DEV Community: Saurabh Bhayana

Claude Models in 2026: Fable 5.1 vs Opus 5.5 vs Sonnet 5.5 vs Haiku 5.5 (Pricing and Which to Use)

Saurabh Bhayana — Fri, 23 Oct 2026 13:00:25 +0000

This guide was first published on saurabhbhayana.com.

The most expensive Claude model costs 100 times more per input token than the cheapest one. Most people who build on Claude never look at that gap. They pick the model they have heard of and pay for it on every request, including the ones a cheaper model would have handled just as well.

This guide covers Anthropic's four current models: Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5. For each one you get what it is for, the full API price list, what real jobs cost in dollars, and working code to call it. The caveat first: AI pricing moves. Every number here comes from Anthropic's own models and pricing pages as of October 2026, and the cost examples leave out thinking tokens, which I explain below. Check the pricing page before you budget a big job.

I build AI automations for clients and run wptaskify, a product that does WordPress work through AI assistants, so I pick between models like these every week. My opinions below come from that work, and I say where they are opinions.

The four Claude models at a glance

Model	API ID	Best for (Anthropic's words)	Speed	Default effort
Claude Fable 5.1	claude-fable-5-1	Demanding reasoning and long-horizon agentic work	Slower	high
Claude Opus 5.5	claude-opus-5-5	Long-running agentic coding and knowledge work	Moderate	medium
Claude Sonnet 5.5	claude-sonnet-5-5	The best combination of speed and intelligence	Fast	high
Claude Haiku 5.5	claude-haiku-5-5	High-volume, latency-sensitive tasks such as classification, extraction and routing	Fastest	medium

All four share more than you might expect. Each has a 1M token context window, which Anthropic puts at roughly 555,000 words, and up to 128K tokens of output. All take text and images, use tools and support vision. All four list June 2026 as their reliable knowledge cutoff, and Anthropic commits to keeping each one available until at least September or October 2027.

Same context window, same output limit, up to 100 times apart on price.

What each model is actually good at

Claude Fable 5.1: the heavy thinker

Fable 5.1 is Anthropic's most capable widely released model. It is built for hard reasoning and for agents that work through long tasks on their own. It always thinks before it answers, single requests on hard problems can run for minutes, and it costs the most. Anthropic's own advice is to try Opus 5.5 first and reach for Fable 5.1 when your tests on Opus at higher effort still fall short. I agree. For most business work, Fable 5.1 is a tool for the hardest 5 percent of jobs, not a default.

Claude Opus 5.5: the default

Anthropic's documentation says to start with Opus 5.5 for most workloads. It handles long coding sessions, research, analysis and writing that needs judgment, and it costs less than the Opus 5 it replaced ($4 and $20 per million tokens, against $5 and $25). Its thinking is always on, and its default effort is medium, one notch lower than the previous Opus, so set the effort yourself if a task needs more care.

Claude Sonnet 5.5: the workhorse

Sonnet 5.5 costs half of Opus 5.5 and runs faster. For high-volume work that still needs real intelligence, such as drafting replies, summarising calls, writing first drafts or running a customer-facing assistant, it is where I would start testing. If Sonnet's output passes your quality check, Opus is money you do not need to spend.

Claude Haiku 5.5: the bulk processor

Haiku 5.5 is the surprise of this lineup. At 10 cents per million input tokens, it is 40 times cheaper than Opus 5.5 on input and 10 times cheaper than the Haiku 4.5 before it. Anthropic positions it for classification, extraction and routing: sorting tickets, pulling fields out of emails, deciding which bigger model a request should go to. Those jobs run thousands of times a day, which is exactly where price per token decides the bill.

Start with the cheapest model that passes your quality check. Move up only when it fails.

Claude API pricing (USD)

All prices are per million tokens, from Anthropic's pricing page. A cache write stores a prompt prefix so later requests can reuse it. A cache read is the price of reusing it.

Model	Input	Output	Cache write (5 min)	Cache read	Batch in / out
Fable 5.1	$10	$50	$12.50	$0.25	$5 / $25
Opus 5.5	$4	$20	$5	$0.20	$2 / $10
Sonnet 5.5	$2	$10	$2.50	$0.10	$1 / $5
Haiku 5.5 (prompts up to 100K tokens)	$0.10	$0.50	$0.125	$0.01	$0.05 / $0.25
Haiku 5.5 (prompts over 100K tokens)	$0.50	$2.50	$0.625	$0.05	$0.25 / $1.25

Four details on that page change real bills:

Batch API: 50 percent off input and output on every model, for jobs that can wait for results.
Prompt caching: reuse the same long instructions or document and pay a fraction for each reuse. A cache read costs 2.5 percent of the input price on Fable 5.1, 5 percent on Opus 5.5 and Sonnet 5.5, and 10 percent on Haiku 5.5.
Long prompts: Fable 5.1, Opus 5.5 and Sonnet 5.5 charge the same rate across the whole 1M context. Haiku 5.5 is the exception, with a higher price once a single prompt passes 100,000 tokens.
US-only processing: asking for inference inside the US applies a 1.1 times multiplier to every price. Fast mode on Opus 5.5 costs $8 and $40 per million tokens for faster output.

What real jobs cost on each model

Token prices mean little until you attach them to work. Here are four jobs, priced on all four models. I use Anthropic's own conversion, about 555,000 words per million tokens, which is roughly 1.8 tokens per English word.

Job	Fable 5.1	Opus 5.5	Sonnet 5.5	Haiku 5.5
Write a 1,500-word draft (1K tokens in, 2.7K out)	$0.145	$0.058	$0.029	$0.0015
Reply to one customer email (1.5K in, 300 out)	$0.030	$0.012	$0.006	$0.0003
Classify 10,000 tickets, Batch API (3M in, 100K out)	$17.50	$7.00	$3.50	$0.18
Summarise a 300-page document (270K in, 2K out)	$2.80	$1.12	$0.56	$0.14

The same 10,000-ticket job, priced from Anthropic's batch rates. The gap is the whole argument for matching model to task.

The math, so you can check it. The draft on Opus 5.5 is 1,000 input tokens at $4 per million ($0.004) plus 2,700 output tokens at $20 per million ($0.054), so $0.058. The 300-page summary on Haiku 5.5 crosses the 100,000-token line, so it uses the higher Haiku rate: 270,000 tokens at $0.50 per million plus 2,000 at $2.50, about 14 cents.

The thinking-token catch

Those numbers leave out thinking. Fable 5.1 and Opus 5.5 always think before answering, and the other two think by default. Thinking tokens are billed as output tokens, so a real request costs more than the table, sometimes much more on a hard question at high effort. You control it with the effort setting, from low to max. Simple jobs at low effort think briefly. Run a sample of your real requests and read the usage numbers in the response before you trust any estimate, mine included.

Tip: Judge cost per finished task, not per request. A cheap model that needs three tries to get a job right can cost more than an expensive one that gets it right once.

Which Claude model should you use?

Job	Start with	Why
Sorting, tagging, extracting fields, routing	Haiku 5.5	Built for it, and volume makes price decisive
Customer replies, summaries, first drafts	Sonnet 5.5	Fast, half the price of Opus, strong writing
Coding, analysis, research, agent work	Opus 5.5	Anthropic's recommended default
Hard reasoning, long autonomous agent runs	Fable 5.1	When Opus at high effort still falls short
Huge one-off jobs that can wait	Any model via Batch API	Half price

A pattern that works well in practice: let a cheap model handle the first pass and hand only the hard cases up. Haiku 5.5 reads every incoming email and decides which ones need a careful answer, and only those go to Sonnet or Opus. Most of the volume never touches the expensive model. The same idea runs through my guide to what AI can automate in a small business, where most tasks are simple, repetitive and perfect for the cheaper tiers.

How do these compare with OpenAI's newest model? I tested GPT-6 Astra separately, and its $10 input and $50 output pricing sits in the same tier as Fable 5.1.

How to write good prompts for Claude

Current Claude models follow instructions closely, so the biggest gains come from saying exactly what you want and why. Three habits matter most.

Give the context and the goal along with the task. Who is the reader, what will they do with the answer, what does good look like.
Say the format you want in plain words. A short list, three sentences, a table with these columns.
Let it think on hard problems by raising effort, rather than writing think step by step into the prompt.

WEAK
Write a reply to this customer.

BETTER
You answer emails for a plumbing company in Riverton.
The customer below is upset about a late technician.
Write a reply under 120 words that apologises once, gives
the new arrival window from the job notes, and offers a
direct number. Plain, warm, no corporate phrases.

<email>...</email>
<job_notes>...</job_notes>

Putting long material inside clear tags like the email and job notes above helps the model keep instructions and data apart. If the reply will be read by a customer, run it past my signs of AI writing checklist before it goes out.

How to use the Claude API, step by step

Create an account in the Claude Console at platform.claude.com and add billing.
Create an API key and save it as an environment variable named ANTHROPIC_API_KEY. Never paste it into your code.
Install the official SDK: pip install anthropic for Python, or npm install @anthropic-ai/sdk for JavaScript and TypeScript.
Send your first message with the model ID you picked, using the code below.
Read the usage numbers in the response to see exactly how many tokens you paid for.

import anthropic

client = anthropic.Anthropic()  # reads ANTHROPIC_API_KEY

response = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=16000,
    output_config={"effort": "medium"},  # low | medium | high | xhigh | max
    messages=[{"role": "user", "content": "Summarise this in 3 bullet points: ..."}],
)

for block in response.content:
    if block.type == "text":
        print(block.text)

print(response.usage)  # input and output tokens you were billed for

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic(); // reads ANTHROPIC_API_KEY

const response = await client.messages.create({
  model: "claude-sonnet-5-5",
  max_tokens: 16000,
  output_config: { effort: "low" },
  messages: [{ role: "user", content: "Classify this ticket as billing, bug or other: ..." }],
});

for (const block of response.content) {
  if (block.type === "text") console.log(block.text);
}
console.log(response.usage);

To switch models, change only the model string. Two things trip people up on the newest models. First, you cannot turn thinking off on Fable 5.1 or Opus 5.5; lower the effort instead. Second, older code that pre-fills the start of Claude's answer or uses a fixed thinking budget returns an error on these models, so update it before you switch.

If you want Claude to work inside your own tools rather than through code you write, such as your website, CRM or files, that is what MCP is for. I explain the difference in MCP vs APIs.

Takeaways

Anthropic's current Claude lineup is Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5, all with a 1M token context window and 128K max output (Anthropic models overview, October 2026).
API prices per million tokens: Fable 5.1 $10 in and $50 out, Opus 5.5 $4 and $20, Sonnet 5.5 $2 and $10, Haiku 5.5 $0.10 and $0.50 for prompts up to 100,000 tokens.
Haiku 5.5 is the only current model with a long-prompt price: prompts over 100,000 tokens cost $0.50 in and $2.50 out per million.
The Batch API halves input and output prices on every model. Cache reads cost 2.5 percent of input on Fable 5.1, 5 percent on Opus 5.5 and Sonnet 5.5, and 10 percent on Haiku 5.5.
Anthropic recommends starting with Opus 5.5 for most workloads and moving to Fable 5.1 for demanding reasoning and long-horizon agent work.
Thinking is always on for Fable 5.1 and Opus 5.5. You control how much they think with the effort setting, and thinking tokens bill as output.
One 1,500-word draft costs about 14.5 cents on Fable 5.1, 5.8 cents on Opus 5.5, 2.9 cents on Sonnet 5.5 and 0.15 cents on Haiku 5.5, before thinking tokens.
Classifying 10,000 short support tickets through the Batch API costs about $17.50 on Fable 5.1 and about 18 cents on Haiku 5.5.

FAQs

What are the current Claude models?

As of October 2026, Anthropic's current models are Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5. All four have a 1M token context window and up to 128K tokens of output. Older models such as Opus 5, Sonnet 5 and Haiku 4.5 are still available as legacy models.

How much does the Claude API cost?

Per million tokens: Fable 5.1 costs $10 input and $50 output, Opus 5.5 $4 and $20, Sonnet 5.5 $2 and $10, and Haiku 5.5 $0.10 and $0.50 for prompts up to 100,000 tokens. The Batch API halves those prices, and cached input costs a small fraction of the normal input price.

Which Claude model should I use?

Anthropic recommends starting with Opus 5.5 for most workloads. Use Haiku 5.5 for high-volume sorting and extraction, Sonnet 5.5 for fast everyday writing and replies, and Fable 5.1 for the hardest reasoning and long agent tasks when Opus falls short.

What is the cheapest Claude model?

Claude Haiku 5.5, at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Through the Batch API that drops to $0.05 and $0.25.

What is the difference between Claude Opus 5.5 and Sonnet 5.5?

Opus 5.5 is the more capable model for long agentic coding and knowledge work and costs $4 and $20 per million tokens. Sonnet 5.5 is faster and costs half as much, $2 and $10, and Anthropic describes it as the best combination of speed and intelligence. Test your task on Sonnet first; if it passes, you save half.

Can I turn off thinking on Claude Opus 5.5?

No. Thinking is always on for Claude Opus 5.5 and Claude Fable 5.1, and requests that try to disable it return an error. Lower the effort setting instead to make the model think less and spend fewer tokens.

How big is Claude's context window?

All four current models have a 1M token context window, which Anthropic estimates at roughly 555,000 words on the current tokenizer. Fable 5.1, Opus 5.5 and Sonnet 5.5 charge the same per-token rate across the full window; Haiku 5.5 charges more for prompts over 100,000 tokens.

How do I get a Claude API key?

Sign up in the Claude Console at platform.claude.com, add billing, and create a key in the API keys section. Store it as the ANTHROPIC_API_KEY environment variable, and the official SDKs will read it automatically.

Originally published at saurabhbhayana.com, where I write about AI automation, AI search and Local SEO for small businesses.

How to Check If ChatGPT Mentions Your Brand (and What to Do If It Doesn't)

Saurabh Bhayana — Thu, 22 Oct 2026 13:00:45 +0000

This guide was first published on saurabhbhayana.com.

Ask ChatGPT for the best plumber in your town. Then open a new chat and ask again. There is a good chance you get two different answers, and a fair chance your business is in neither.

That matters more every month. OpenAI said in February 2026 that ChatGPT had 900 million weekly users. And in BrightLocal's 2026 survey of US consumers, 45 percent said they now use ChatGPT or other AI tools to find local businesses, up from 6 percent a year earlier. A growing share of your customers ask an assistant before they ask Google.

The honest caveat up front: there is no dashboard for this. ChatGPT does not show businesses a report of when it mentions them, and its answers change from one chat to the next. Every method, including the paid tools, is a way of asking it lots of questions and counting. So this guide gives you a test you can run yourself in about an hour, shows how to read it without fooling yourself, and then covers what to fix.

I will use three examples: a plumbing company, a neighbourhood bakery, and a small software company that sells proposal software. Local, walk-in and online. Between them they cover how most businesses get found.

Mention, citation, recommendation: three different wins

Before you test anything, separate the three things an AI answer can do for you. Mixing them up is the most common mistake in AI visibility tracking.

Signal	What it looks like	Why it matters
Mention	Your business name appears anywhere in the answer	The assistant knows you exist
Citation	Your web page is linked as a source	The assistant trusts your page enough to quote it, and it can send a click
Recommendation	Your business is offered as the answer to the question	This is the one that wins customers

Three different wins. A mention is awareness. A recommendation is a customer.

A bakery can be mentioned in a list of ten places and still never be recommended. A software company can be cited for a statistic in its blog post while ChatGPT recommends a competitor's product. Count each one separately.

The 20-prompt test

Write 20 questions your customers actually ask, in their words, not yours. Split them across four types so you see the whole picture.

Type	What it tests	Example (plumbing company)
Brand	Does it know you and get the facts right?	What is Northside Plumbing and what do they do?
Category	Do you show up when people shop?	Best emergency plumber in Riverton
Comparison	How are you positioned against rivals?	Northside Plumbing vs Apex Plumbing, which is better?
Problem	Do you get recommended for the job you do?	My water heater is leaking, who should I call in Riverton?

Five prompts of each type gives you 20. The problem prompts matter most, because that is how real customers ask. Nobody types the name of a business they have never heard of.

PROMPT SET TEMPLATE (swap in your details)

Brand
  What is [business] and what do they do?
  Is [business] in [city] any good?
  What do customers say about [business]?

Category
  Best [service] in [city]
  Who are the top [service] companies near [neighbourhood]?

Comparison
  [business] vs [competitor], which should I choose?
  Alternatives to [competitor] in [city]

Problem
  [the problem in the customer's words], who should I call in [city]?
  How much does [service] cost in [city] and who do you recommend?

How to run it so the results mean something

Use a temporary or logged-out chat. Your own chat history and memory can make ChatGPT more likely to mention you, which flatters the result.
Run each prompt three times, in fresh chats. Answers vary, so one run is a weak signal. Three runs per prompt gives you 60 answers.
Note whether the answer searched the web. When ChatGPT shows sources, it searched. Answers with search and without it can differ a lot.
Log every answer in a spreadsheet: date, prompt, run number, mentioned yes or no, cited yes or no, recommended yes or no, which competitors appeared, and any wrong facts.
Repeat the same prompts in Perplexity, Gemini, Claude and Google's AI Overviews. Same questions, different engines, so you can compare.

Tip: Keep the prompt list fixed and rerun it every month. Changing the questions each time makes the trend meaningless.

How to read the results

Turn the spreadsheet into three rates. Mention rate is the share of answers that named you. Citation rate is the share that linked your page. Recommendation rate is the share that offered you as the answer. Then look at accuracy separately.

An illustrative scorecard. The gap to look for: AI knows the brand when asked by name, but rarely recommends it for the actual problem.

If your results look like this example, they tell you where to work. High on brand prompts and low on problem prompts means the assistant knows you exist but does not connect you with the job. That is a content and reputation problem, not a technical one.

Wrong facts come first, before any rate. If ChatGPT says the bakery closes at 5 when it closes at 8, or that the software company has no free trial when it does, customers act on that. Find where the wrong fact lives on the web, usually an old directory listing or an outdated page, and fix it at the source.

Track the rate across many runs, never a single answer. One answer is luck. Sixty answers is a measurement.

Check that ChatGPT can reach your site

Do this before anything else in the fix list, because it is the only problem that can hide you completely. OpenAI runs separate crawlers for separate jobs, and its own documentation is clear about what each one does.

Crawler	What it does (per OpenAI)	If you block it
OAI-SearchBot	Surfaces websites in ChatGPT's search features	Your site will not be shown in ChatGPT search answers, though it can still appear as a navigational link
GPTBot	Collects content used to train OpenAI's models	Signals your content should not be used for training. Search is a separate control
ChatGPT-User	Visits a page when a user's request needs it	Not an automatic crawler, so robots.txt rules may not apply

So you can keep your content out of training and still appear in ChatGPT search. OpenAI recommends allowing OAI-SearchBot. Here is a robots.txt that does both:

# Appear in ChatGPT search answers
User-agent: OAI-SearchBot
Allow: /

# Optional: opt out of model training
User-agent: GPTBot
Disallow: /

Two things to check beyond robots.txt. First, a blanket rule such as User-agent: * with Disallow: / blocks every crawler, OAI-SearchBot included. Second, some CDN, firewall and hosting settings block AI crawlers on their own, separate from robots.txt, so check those dashboards too. OpenAI says robots.txt changes for search can take about 24 hours to take effect.

Why ChatGPT does not mention you, and the fix for each

Five causes, five fixes. Work top to bottom: access first, then clarity, then reputation.

1. It cannot reach you

Covered above. Fix robots.txt and any CDN or firewall setting that blocks AI crawlers.

2. Your site does not say plainly what you are

Assistants need to connect your name with what you do and where. Many small business sites never state it in one plain sentence. The fix: an About page and homepage that say, in words, Northside Plumbing is a licensed plumbing company serving Riverton and the surrounding towns since 2009. Add Organization or LocalBusiness schema with the same facts, and keep your name, address and phone identical everywhere.

3. Not enough of the web talks about you

Assistants lean on what others say, not only what you say about yourself. Directory listings, genuine reviews, local news, industry roundups and forum threads all build that picture. For local businesses, reviews are the fastest lever, and my guide on how to get more Google reviews covers how to do it inside the rules. For the software company, real comparison articles and listings on software review sites do the same job.

4. You have no pages that answer the questions

Look at the problem prompts you lost. Each one is a page you should have: how much a water heater replacement costs in your area, which cake to order for 30 people, how proposal software compares with templates. Write the answer first, then the detail, with real numbers. My AEO and GEO guide explains the structure assistants find easiest to quote.

5. The web has wrong facts about you

Old hours on a directory, a former address, a product you discontinued. Assistants repeat what they find. Fix the source, not the answer: update the listing, the old page or the profile, then rerun the brand prompts a few weeks later.

Free and paid ways to automate the check

The manual test works, but it takes an hour a month. There are two ways to shorten it.

Free quick check: my team built a free AI Visibility Checker on wptaskify. Type your brand and it shows how an AI model describes you, whether it recognises you, and the biggest gap. It reflects one model's view at one moment, so treat it as a starting point, not a full test.
Paid trackers: a growing group of tools run your prompt set across several assistants on a schedule and chart mention and citation rates over time. They are worth it once you are actively working on AI visibility and need trends, not for a first look.
Do it yourself with a script: if you have a developer, the prompt set and spreadsheet above can be run through the model APIs on a schedule. Note that API answers can differ from the ChatGPT app, because the app can search the web and personalise.

You can try the free checker at wptaskify.com/ai-visibility. Whatever you use, keep the fixed prompt list. It is the only way to see whether your fixes are working.

A monthly 30-minute routine

First of the month: rerun the same 20 prompts, three times each, in a temporary chat. Log the results.
Update the three rates and compare with last month.
Fix any wrong fact the same day, at its source.
Pick the two problem prompts you lost most often and write or improve the page that answers each one.
Check robots.txt and your CDN settings after any site or hosting change.
Every quarter, repeat the test in Perplexity, Gemini, Claude and Google's AI Overviews.

The gap to watch is a good mention rate on brand prompts next to a low recommendation rate on problem prompts. The work is closing that gap, and it is the same work that grows Google traffic, so nothing here is wasted. If you want it done for you, AEO and GEO is part of what my team does. Tell me which questions you want to be the answer to.

Takeaways

OpenAI said on February 27, 2026 that ChatGPT had reached 900 million weekly active users.
In BrightLocal's 2026 survey of 1,002 US consumers, 45 percent said they use ChatGPT or other generative AI for local business recommendations, up from 6 percent the year before.
ChatGPT has no dashboard or report that shows what it says about a brand, so every check is a test you run yourself or a tool that runs prompts for you.
A mention is your name in the answer, a citation is your page linked as a source, and a recommendation is your business offered as the answer. Track all three separately.
AI answers change with wording, time and whether the assistant searches the web, so run each prompt several times and track the rate, not one result.
OpenAI's docs say sites that block OAI-SearchBot in robots.txt will not be shown in ChatGPT search answers. GPTBot is a separate crawler used for model training.
OpenAI says robots.txt changes for search can take about 24 hours to take effect.
Businesses get mentioned when the web says clearly what they do and where: a plain About page, consistent listings, genuine reviews and pages that answer real questions.

FAQs

How do I check if ChatGPT mentions my business?

Write about 20 questions your customers ask, covering your brand name, your category, comparisons with competitors and the problems you solve. Run each one three times in a temporary chat, log whether you were mentioned, cited or recommended, and repeat the same set every month to track the rate.

Is there a tool that shows what ChatGPT says about my brand?

Not from OpenAI. ChatGPT has no dashboard or report of brand mentions. Free checkers give a quick snapshot, and paid trackers run a fixed set of prompts on a schedule across several assistants. All of them work by asking questions and counting the answers.

Why does ChatGPT not recommend my business?

The usual causes are blocked AI crawlers, a website that never says plainly what the business does and where, too few mentions on other sites such as listings and reviews, no pages that answer the questions customers ask, and wrong facts about the business elsewhere online.

Does blocking GPTBot remove my site from ChatGPT?

No. According to OpenAI, GPTBot collects content for model training, and OAI-SearchBot surfaces sites in ChatGPT search. You can block GPTBot and still allow OAI-SearchBot. Sites that block OAI-SearchBot are not shown in ChatGPT search answers.

How long does it take for ChatGPT to pick up changes?

OpenAI says robots.txt changes for search can take about 24 hours to apply. Changes to how assistants describe your business take longer, usually weeks, because they depend on pages being recrawled and on mentions building up across other sites.

Why does ChatGPT give different answers to the same question?

Answers vary with exact wording, timing, the model version, whether the assistant searched the web and the user's own chat history. That is why one answer is a weak signal and you should run each prompt several times and track the rate.

How many people use ChatGPT to find local businesses?

In BrightLocal's 2026 survey of 1,002 US consumers, 45 percent said they use ChatGPT or other generative AI tools for local business recommendations, up from 6 percent the year before. OpenAI reported 900 million weekly ChatGPT users in February 2026.

Originally published at saurabhbhayana.com, where I write about AI automation, AI search and Local SEO for small businesses.

n8n vs Zapier vs Make vs Custom Code: Which One Should Run Your Business Automation?

Saurabh Bhayana — Wed, 21 Oct 2026 13:01:01 +0000

This guide was first published on saurabhbhayana.com.

Zapier, Make and n8n all promise the same thing: connect your apps and let the boring work run itself. They are not the same product, and the difference that matters most is not on their homepages. It is what each one charges you for.

Zapier charges per action. Make charges per module. n8n charges per run. That sounds like fine print until a workflow grows from three steps to twelve, and one tool's bill quadruples while another's stays flat.

This guide compares the three on cost, ease and control, works the same workflow through each one, and then covers the question most comparisons skip: when should you use none of them and write custom code? The honest caveat first: prices change, and these are the prices on Zapier's and n8n's own pricing pages as of October 2026. Make does not publish prices on its public pricing page, so for Make I explain how it bills and give a range rather than pretend to an exact number. Check each page before you commit.

Where I have an opinion, it comes from building automations for clients at Logical Dottech, including one where the client insisted on n8n and we ended up using it for half the job. That story is further down.

The short answer

If you	Use
Are not technical and need a few simple automations working this week	Zapier
Want visual, branching workflows at a lower price per step	Make
Have long workflows, high volume, or want to self-host and own your data	n8n
Need a screen, a login, memory or rules your team can change safely	Custom code (often with n8n for the background jobs)

The rest of this guide explains why, so you can check the answer against your own workflows.

How each tool bills you

This is the part that decides your bill, so it is worth getting exactly right.

Tool	Unit	What counts	What is free
Zapier	Task	Each action that completes successfully	Triggers, polling, Filter, Formatter, Paths, failed actions
Make	Credit	Each module action, most use 1 credit, AI modules can use more	Routers and error handler modules
n8n	Execution	One full run of a workflow, whatever its length	Steps inside the run
Custom code	Server time	Hosting, plus any paid APIs you call	Everything you write yourself

One extra detail from Zapier's page: its tasks are a single pool shared across automations, AI by Zapier and Code by Zapier, and a Zapier MCP tool call uses two tasks. If AI agents are part of your plan, read that pool carefully.

Same workflow, three bills. Zapier and Make count steps. n8n counts runs.

What they cost: the entry plans

Plan	Price (billed yearly)	Included	Billed monthly
Zapier Free	$0	100 tasks, two-step Zaps only, 15-minute polling	$0
Zapier Professional	$19.99 a month	750 tasks a month at the entry tier	$29.99 a month
Zapier Team	$69 a month	2,000 tasks a month at the entry tier	$103.50 a month
Make Free	$0	1,000 credits a month	$0
Make Core	Roughly $9 to $11 a month in 2026 pricing guides	About 10,000 credits a month	Check make.com
n8n Starter (cloud)	20 euros a month	2,500 executions a month	About 20% more (yearly saves 17%)
n8n Pro (cloud)	50 euros a month	10,000 executions a month	About 20% more (yearly saves 17%)
n8n Community Edition	Free, self-hosted	No execution limit from n8n, you pay for the server	Free

Two things to notice. Zapier's yearly price is a third lower than its monthly price, so the plan you see first is the yearly one. And if you go over your task limit with pay-per-task switched on, Zapier's page says overage is billed at 2.5 times the base rate on monthly plans and 1.25 times on yearly plans.

The same workflow on each tool

Here is a workflow almost every service business needs, the same one that tops my list of tasks to automate: a lead fills the website form, the details go into the CRM, the lead gets a reply email, and the team gets a Slack message. One trigger, three actions. Say 500 leads a month.

Tool	Units per lead	Units a month (500 leads)	Fits which plan
Zapier	3 tasks (trigger is free)	1,500 tasks	Above Professional's 750 entry tier, so a higher task tier or Team
Make	About 4 credits	About 2,000 credits	Comfortably inside a 10,000-credit Core plan
n8n cloud	1 execution	500 executions	Inside Starter's 2,500
Custom code	A few milliseconds of server time	Effectively nothing	A small server you may already have

Now add an AI step that writes a personal reply, a lookup to check whether the lead is already a customer, a filter for spam, a second Slack channel for big jobs, and a follow-up two days later. That is a realistic version of the same workflow, and it is ten steps instead of four.

Every step you add raises the bill on Zapier and Make. On n8n, the run is the unit, so long workflows stay flat.

On Zapier the ten-step version uses roughly three times the tasks of the four-step one. On Make, roughly two and a half times the credits. On n8n, it is still one execution per lead. That single difference explains most of the n8n vs Zapier arguments you will read online. For short workflows it barely matters. For long, busy ones it is most of the bill.

Zapier and Make charge for steps. n8n charges for runs. Count your steps before you pick.

Zapier: the easiest to start, the priciest per step

Zapier is the tool to give someone who has never built an automation. You pick a trigger app, pick an action, map a few fields, and it runs. It connects to the widest range of business apps of the three, and for a simple two- or three-step job it is hard to beat.

Best for: non-technical owners, simple linear automations, apps that only Zapier connects to.
Watch out for: cost per step on long or busy workflows, the shared task pool if you also use its AI features, and overage rates if you leave pay-per-task on.
Free plan: real but small. 100 tasks and two-step Zaps only, so it is a trial more than a plan.

Make: visual and cheaper per step

Make shows your workflow as a map of connected circles, with branches, loops and error handling you can see. It takes an afternoon longer to learn than Zapier and gives you far more control for the money. Routers and error handlers are free, which rewards building workflows properly instead of cutting corners.

Best for: branching workflows, people who think visually, moderate volume on a tight budget.
Watch out for: credit use climbing with every module, AI modules using more than one credit, and loops over long lists that burn credits per item.
Free plan: 1,000 credits a month, enough to build and test properly.

n8n: flat pricing, self-hosting and code when you need it

n8n looks like Make, a canvas of connected nodes, but it bills per run and you can host it yourself. It also lets you drop into JavaScript or Python inside a workflow, which removes the wall you eventually hit in pure no-code tools.

Self-hosting is where people get the licence wrong. The Community Edition is free and self-hosted, under n8n's Sustainable Use License. That licence lets you use or modify n8n for your own internal business purposes, or for non-commercial or personal use. You may give it to others only free of charge for non-commercial purposes. In plain terms: running it for your own business is fine. Building a paid product where your customers use n8n as the service is not, without a commercial licence from n8n.

Best for: long or high-volume workflows, teams with someone technical, businesses that want data on their own server.
Watch out for: self-hosting means you own updates, backups and security. Some features are only on paid plans, and Business is self-hosted at 667 euros a month on yearly billing.
Free option: Community Edition, self-hosted. Your cost is the server and your time.

When none of them fit: custom code

All three tools are built for background work. Something happens, steps run, done. None of them gives your customers or staff a login, a screen to use, or memory between runs. The moment your automation needs any of those, you are building an app, and the workflow tool becomes a second system you maintain next to it.

A real example: the client who wanted n8n

A French real estate firm hired us to build a lead system. Every night it watches official public business notices and finds companies their advisors should call. The client asked for n8n, and the reasons were good ones: open source, self-hosted, no vendor lock-in, low running costs, and the ability to change the criteria himself.

For the nightly pipeline, collect the notices, filter them, remove duplicates, push to the CRM, n8n fit well. But the product they actually wanted had more: a chat where advisors type only Bordeaux, only this month and it remembers that, a map, prospect cards, a validation button and an admin console. n8n has no interface for end users, so that app had to be built either way.

Our estimate: about 22 days to build it n8n-first, against about 18 days as a normal web app, because n8n helped with roughly 15 percent of the work and added overhead to the rest. Hosting was a wash, a few euros a month either way. The argument that settled it was simple: using n8n meant building the same app plus maintaining a second system.

What we did: the nightly collection job runs in a workflow tool, which suits it, and the application is ordinary code. The real want, I can change the criteria without calling the developer, became a settings screen where the client ticks regions and notice types. Easier for the client than editing workflow nodes, and nothing can break. The whole system runs for under 20 dollars a month.

Tip: If the honest answer to will a person use this every day? is yes, you are building an app. Price it as one, and use a workflow tool only for the parts nobody sees.

Signs you need custom code

Customers or staff need to log in and use a screen.
The automation needs to remember things between runs, like a conversation or a user's settings.
Volume is high enough that per-step pricing is a real line in the budget.
The logic has many rules and exceptions that turn a canvas into spaghetti.
Facts must come from a strict source and never be invented, such as prices, legal data or medical data. Code makes that rule easy to enforce and test.

Where AI fits in each one

All four can call an AI model for the reading and writing steps: classifying a message, drafting a reply, pulling fields out of an email. The difference is how you pay. Zapier counts AI by Zapier in the same task pool. Make's AI modules can use more than one credit. n8n counts the run once, and you pay the AI provider directly if you bring your own key. Custom code always pays the provider directly, which is usually the cheapest per call.

The rule from my automation guide applies in all four: let AI read and write around facts, never supply them. If you are connecting AI assistants like Claude or ChatGPT to your tools rather than running fixed workflows, that is a different pattern, which I explain in MCP vs APIs.

How to choose in 15 minutes

List the three workflows you want first. For each, write the trigger and count the steps.
Estimate runs a month. Leads, orders, bookings, whatever starts the workflow.
Multiply. Steps times runs gives Zapier tasks and Make credits. Runs alone gives n8n executions.
Check who will maintain it. If nobody technical, lean Zapier or Make. If someone can, n8n opens up.
Ask the app question. If a person will log in and use it, plan custom code for that part.
Start on the free tier of your pick and build the smallest workflow first. Switching early is cheap. Switching after 40 workflows is not.

If you want a second opinion on which way to go, or want the whole thing built and running, that is what my team does. Tell me what you want automated and I will tell you which tool I would use and why.

Takeaways

Zapier bills per task: each successful action is one task, while triggers, filters, formatter steps and Paths are free (Zapier pricing page, October 2026).
Make bills per credit: most module actions use one credit, routers and error handlers use none, and AI modules can use more.
n8n bills per execution: one full run of a workflow counts once, however many steps it has.
On annual billing, Zapier Professional starts at $19.99 a month for 750 tasks and Team at $69 a month for 2,000 tasks. n8n cloud Starter is 20 euros a month for 2,500 executions and Pro is 50 euros for 10,000.
A 10-step workflow costs 10 times as much as a 1-step one on Zapier or Make, but the same on n8n. Long workflows favour n8n, short ones barely matter.
The self-hosted n8n Community Edition is free under the Sustainable Use License, which allows use for your own internal business purposes but not reselling it to others.
None of the three gives your customers or staff a login, a screen or memory between runs. If you need those, you are building an app anyway, and custom code is usually cheaper overall.
The pattern that works most often: a no-code tool for background pipelines, custom code for anything a person has to use.

FAQs

Is n8n better than Zapier?

It depends on the workflow. n8n bills one execution per run however many steps it has, can be self-hosted and lets you write code inside workflows, so it wins on long, high-volume workflows. Zapier is easier for non-technical users and charges per action, so it suits short, simple automations.

How does Zapier count tasks?

A task is counted each time an action completes successfully. Triggers, polling, Filter, Formatter and Paths steps do not count, and failed actions are not counted. Tasks are one pool shared across automations, AI by Zapier and Code by Zapier.

How much does Zapier cost in 2026?

On Zapier's pricing page in October 2026, the Free plan includes 100 tasks a month. Professional starts at $19.99 a month billed yearly or $29.99 monthly for 750 tasks, and Team starts at $69 a month billed yearly or $103.50 monthly for 2,000 tasks.

How much does n8n cost?

n8n cloud Starter is 20 euros a month billed yearly for 2,500 executions, Pro is 50 euros for 10,000, and Business is 667 euros a month, self-hosted, for 40,000. The Community Edition is free to self-host; you pay only for the server.

Is self-hosted n8n free for commercial use?

For your own business, yes. n8n's Sustainable Use License allows use for your own internal business purposes. It does not allow you to provide n8n to others commercially, for example as a paid hosted service, without a separate licence from n8n.

Is Make cheaper than Zapier?

Usually, per step. Make charges about one credit per module action and does not charge for routers or error handlers, and its entry paid plans include far more credits than Zapier's entry plans include tasks. Make does not list prices on its public pricing page, so confirm current prices there.

When should I use custom code instead of a no-code tool?

When people need to log in and use a screen, when the automation must remember things between runs, when volume makes per-step pricing expensive, or when strict rules decide which data is allowed. In those cases you are building an app, and a workflow tool becomes a second system to maintain.

Originally published at saurabhbhayana.com, where I write about AI automation, AI search and Local SEO for small businesses.

What Can AI Automate in a Small Business? 25 Tasks Ranked by Hours Saved

Saurabh Bhayana — Tue, 20 Oct 2026 13:00:41 +0000

This guide was first published on saurabhbhayana.com.

Most small business owners do not need an AI strategy. They need their Tuesday back.

The numbers are not subtle. In a May 2024 survey of 837 US businesses with five or fewer employees, UENI found the median owner works 60 hours a week, and 22.4 percent of that goes to admin, legal and compliance. That is about 13 hours a week of work nobody started a business to do. Add the repeat customer questions, the follow-up emails and the invoices you chase twice, and the real number is higher.

This guide lists 25 of those jobs that AI can take over or cut down today, ranked by how many hours each one gives back. One honest caveat before the list: the hours are estimates. I show the math for every task so you can swap in your own numbers, and your week will not look exactly like mine or anyone else's. Some of these tasks also should not run on full autopilot, and I will say which ones and why.

I run Logical Dottech, a small team that builds these automations for clients, and wptaskify, a product that lets AI do WordPress work on its own. So the advice below comes from things we have built and broken, not from a vendor brochure.

First, the honest numbers on AI in small business

Ask two surveys how many small businesses use AI and you get two very different answers.

Source	When	Who was asked	Share using AI
US Census Bureau, Business Trends and Outlook Survey	Dec 2025 to May 2026	A large sample of US businesses of all sizes	19.8% as of May 3, 2026
Thryv small business AI survey	April 2026	561 US small and mid-sized business decision makers	66%, up from 55% a year earlier

Both can be true. The Census question asks whether a business uses AI in any business function. The vendor survey asks owners, many of whom count opening ChatGPT to write one email as using AI. One measures AI built into how the business runs. The other measures owners who have tried it.

The Census detail that matters most for you: from December 2025 to May 2026, AI use rose at firms with at least 20 employees, but did not change in any meaningful way at firms with fewer than 20. Larger firms are pulling ahead. Firms with 250 or more employees were at 37 percent.

Most of your small competitors have tried AI. Very few have built it into the week. That gap is the opportunity.

The same Thryv survey found 70 percent of owners say they need more training to use AI well. So the problem is rarely the tools. It is knowing which job to hand over first, and how to set it up so it does not embarrass you in front of a customer. That is what the rest of this guide is for.

How to measure the hours a task gives back

Before the list, here is the one formula behind every number in it. Use it on your own business and you will know which task to automate first without guessing.

Hours by hand a week = times it happens a week x minutes each time / 60
Hours back a week    = hours by hand x share the automation handles

The share matters. An automation rarely handles 100 percent of a task. A lead reply bot might answer 85 percent of new enquiries on its own and pass the odd ones to you. A proposal drafter might do half the work, because you still check the price and the scope.

One formula ranks every task: volume times minutes, times the share the automation handles.

To keep the numbers comparable, I use three example businesses through this whole guide. Pick the one closest to yours.

A plumbing company with six people: two in the office, four in vans. Lives on phone calls, quotes and booked jobs.
A bookkeeping firm with four people. Lives on documents, client email and deadlines.
An online store with three people. Lives on product listings, order questions and stock.

Tip: Time one task for a week before you automate it. Write down every time it happens and how long it took. People guess low on small tasks and high on big ones, and the real number almost always surprises them.

The 25 tasks, ranked by hours back

Here is the full list. The volume and minutes are realistic for a business of three to six people, and the last column says whether AI should do the task on its own or draft it for a person to approve. No single business needs all 25. Most will find five to eight that fit.

#	Task	Example week	By hand	Hours back	Mode
1	First reply and follow-up to new leads	25 leads x 10 min	4.2 h	3.5 h	Automate
2	Answering repeat customer questions	50 questions x 5 min	4.2 h	3.0 h	Automate
3	Invoices and payment reminders	30 invoices x 5 min + 10 chases x 10 min	4.2 h	3.0 h	Automate
4	Data entry from forms and emails	60 entries x 3 min	3.0 h	2.7 h	Automate
5	Booking and rescheduling	20 bookings x 8 min	2.7 h	2.5 h	Automate
6	Inbox triage and reply drafts	1 h a day	5.0 h	2.5 h	Draft
7	Product descriptions and listings	20 listings x 10 min	3.3 h	2.5 h	Draft
8	Sorting transactions for the books	150 lines x 1 min	2.5 h	2.0 h	Draft
9	First drafts of quotes and proposals	5 quotes x 45 min	3.8 h	2.0 h	Draft
10	Blog and SEO content drafts	1 post x 4 h	4.0 h	2.0 h	Draft
11	Order status questions	30 questions x 4 min	2.0 h	1.8 h	Automate
12	Screening job applications	40 applications x 4 min	2.7 h	1.8 h	Draft
13	Weekly numbers report	Pulling 6 sources	2.0 h	1.7 h	Automate
14	Call and meeting notes with follow-ups	6 calls x 20 min	2.0 h	1.6 h	Draft
15	Pulling details out of documents	20 PDFs x 6 min	2.0 h	1.6 h	Draft
16	Social media posts	5 posts x 30 min	2.5 h	1.5 h	Draft
17	New client onboarding	3 clients x 40 min	2.0 h	1.5 h	Automate
18	Staff questions about how things work	20 questions x 6 min	2.0 h	1.5 h	Automate
19	Low stock alerts and reorder drafts	Daily stock check	1.5 h	1.2 h	Draft
20	Missed call text back	15 missed calls x 5 min	1.3 h	1.1 h	Automate
21	Scheduling field staff	Daily job board	2.0 h	1.0 h	Draft
22	Small website updates	6 edits x 15 min	1.5 h	1.0 h	Draft
23	Capturing expense receipts	30 receipts x 2 min	1.0 h	0.9 h	Automate
24	Replying to Google reviews	10 reviews x 6 min	1.0 h	0.8 h	Draft
25	Translating customer messages	15 messages x 3 min	0.8 h	0.7 h	Automate

The top 10 alone give back about 26 hours a week in a business that has all of them. Yours will have some, not all.

Notice the pattern at the top. The biggest wins are not clever. They are the jobs that happen dozens of times a week, follow the same shape every time, and sit between your business and your customer. That is where AI is strongest and where a slow human reply costs you money.

Customer-facing tasks: where speed earns money

These tasks rank high for two reasons. They repeat constantly, and every minute of delay costs you. A lead that waits a day for a reply often books someone else.

1. First reply and follow-up to new leads

When a form, email or chat comes in, the automation replies within a minute, asks the two or three questions you always ask, and books a call or sends a quote request to the right person. If the lead goes quiet, it follows up on day two and day five, then stops. For the plumbing company this is the single largest win, because a burst pipe call goes to whoever answers first.

The fix that makes it work: give it your real service area, real price ranges and real availability, and tell it to hand off to a person the moment someone asks for something outside them.

2. Answering repeat customer questions

Do you service my area? What are your hours? How long does shipping take? Most businesses answer the same 20 questions every week. An AI assistant connected to your own FAQ, policies and price list answers them by email, chat or DM, any hour. The trick is that it answers only from your documents, which is a method called RAG. If the answer is not in your material, it says so and passes the question on.

5. Booking and rescheduling

The back and forth of finding a time is pure waste. Connect the assistant to your calendar or booking software and it offers real open slots, confirms, sends reminders and handles the reschedule. For the bookkeeping firm, this one also cuts no-shows, because the reminders actually go out.

11. Order status questions

For the online store, where is my order is the most common message there is. The assistant looks up the order in your store and gives the tracking link and expected date. It needs read access to orders and nothing more. It should never issue refunds on its own.

20. Missed call text back

When a call goes unanswered, the system texts the caller within a minute: sorry we missed you, what do you need, here is a link to book. It is simple, cheap and saves leads that would have called the next company on the list. Field businesses like the plumber miss calls all day while staff are on jobs.

24. Replying to Google reviews

AI drafts a reply to every review that mentions what the customer actually said, and you approve it with one click. Replying to reviews is part of how your profile shows Google you are active, which I cover in my Google Business Profile ranking guide. Keep a person on the negative ones. An auto reply to an angry customer reads exactly like what it is.

25. Translating customer messages

If some customers write in Spanish or another language, the automation translates the message for you and your reply back for them. Modern models handle everyday business language well. For anything legal or medical, use a professional translator.

Back-office tasks: where the 13 hours hide

This is the admin share from the UENI survey. None of it is hard. All of it is slow, and most of it is copying information from one place to another.

About 13 of an owner's 60 weekly hours go to admin. That is the pool most automations draw from.

3. Invoices and payment reminders

When a job is marked done, the invoice goes out with the right line items. If it is unpaid after seven days, a polite reminder goes. After 14 days, a firmer one, and after 21 days, it flags the account for you to call. You stop chasing money by memory. For the plumber and the bookkeeper alike, this one pays for the setup the first month.

4. Data entry from forms and emails

A new customer fills a form, sends an email or a photo of a document, and the details land in your CRM or spreadsheet in the right columns. AI is good at reading messy input, such as an email that says call me after 5, my address is the blue house on Oak, and turning it into clean fields.

8. Sorting transactions for the books

AI suggests a category for each bank line based on the payee and your past choices, and a person confirms the batch. The bookkeeping firm uses this across many clients, so small minutes become large hours. Do not let it file anything with a tax authority on its own.

13. The weekly numbers report

Every Monday the automation pulls sales, leads, jobs booked, unpaid invoices and ad spend from their sources and writes a short plain-English summary: what went up, what went down, what needs a look. It replaces two hours of opening tabs.

15. Pulling details out of documents

Contracts, purchase orders, statements, permits. The AI reads the PDF and pulls out the dates, amounts, names and terms into a table, then a person spot checks it. It is fast, and it is also exactly where AI can misread a number, so the check stays.

23. Capturing expense receipts

Staff snap a photo of a receipt, and the vendor, date, amount and category get filled in and matched to the card transaction. Small task, but it ends the end-of-month shoebox.

Content and marketing tasks: AI drafts, you decide

AI writes fast, and fast is the danger. Readers and search engines can tell when nobody edited the text. Use AI for the first 70 percent and keep a person on the last 30. My guide on signs of AI writing shows exactly what to strip out before anything goes live.

7. Product descriptions and listings

For the online store, AI turns a spec sheet and two photos into a title, description, alt text and SEO fields in your brand voice. A person checks sizes, materials and claims, since a wrong material on a listing becomes a return.

10. Blog and SEO content drafts

AI builds the outline from real search questions and writes a draft. You add what only you know: the job that went wrong, the real price, the local detail. That human layer is what gets content ranked and cited by AI search tools, which I explain in my AEO and GEO guide.

16. Social media posts

Turn one finished job, one review or one blog post into a week of posts, drafted and scheduled at the times your audience is online. You approve the batch once a week instead of thinking of something to post every day.

22. Small website updates

New opening hours, a fresh photo, a price change, a fixed meta description. This is the work I built wptaskify for: you tell an AI assistant what to change on your WordPress site in plain English and it does it, with a backup first. It works because the assistant talks to the site through a proper connection, which I explain in MCP vs APIs.

Team and operations tasks: fewer interruptions

These do not touch customers directly, but they decide how much of the day you spend being interrupted.

6. Inbox triage and reply drafts

The automation sorts incoming email into needs you, can wait and no reply needed, and drafts replies for the first group. You review the drafts in one sitting. This is a draft task for a reason. An AI that sends email as you, unchecked, will eventually say something you would not.

9. First drafts of quotes and proposals

From the call notes, AI fills your quote template with the scope, line items from your price list and the standard terms. You check the price and the scope, then send. It cuts a 45-minute job to about 20, which is why the share is only half.

12. Screening job applications

AI reads applications against the must-haves you set, such as a license, years of experience or the service area, and sorts them into a short list with a reason for each. A person makes every decision. Hiring rules are strict, and an automated rejection you cannot explain is a risk you do not need.

14. Call and meeting notes with follow-ups

Record the call with consent, and you get a summary, the decisions, the action items and a follow-up email draft within a minute of hanging up. It also means nothing promised on a call gets forgotten.

17. New client onboarding

When a client signs, the welcome email, intake form, shared folder, first invoice and kickoff booking all happen in order without anyone remembering to do them. The bookkeeping firm sees this one most, because every new client needs the same ten steps.

18. Staff questions about how things work

Where is the warranty form? What is our policy on after-hours call outs? Put your process documents behind an internal assistant and staff ask it instead of you. Same RAG method as the customer assistant, pointed inward.

19 and 21. Stock alerts and field scheduling

For the store, the automation watches stock, warns before you run out and drafts the reorder for approval. For the plumber, AI proposes tomorrow's job board based on location and skill, and the office manager adjusts it. Both are draft tasks, because the person doing them knows things the system does not, like which van has a flat.

What you should not automate

Every task above sits in one of three modes. Getting the mode right matters more than picking the right tool.

Pick the mode before the tool. Most tasks belong in the middle column.

Automate fully when a mistake is cheap and easy to undo: a booking reminder, a receipt category, a translated message.
Let AI draft and a person approve when a mistake costs money or trust: quotes, review replies, emails sent as you, listings.
Keep it human for decisions about money, people and legal commitments: final prices, refunds above a set amount, hiring, firing, anything filed with a government agency, and the first reply to an angry customer.

The other rule I do not break: never let AI make up data. Here is a real example. We built a lead system for a French real estate firm that watches public business notices every night and finds companies that matter to their advisors. The obvious approach is to feed every record to a model and ask it to pick the good ones. That is slow, expensive, and the model will invent details that are not in the record.

So the data comes straight from the official feed and is filtered with plain code. The model only does the one thing it is good at here: turning an advisor's sentence, such as only Bordeaux, only liquidations this month, into a filter. The whole system runs for under 20 dollars a month, and nothing it shows was made up.

Facts come from your systems. AI reads, sorts and writes around them. The moment it supplies the facts, you have a problem.

What it costs to run

Less than most owners expect, as long as the AI is used only where it is needed. Most of an automation is not AI at all. Moving a form into a CRM, sending a reminder on day seven or pulling a report is ordinary software that costs next to nothing to run.

The AI part, reading a message or drafting a reply, runs on per-use pricing. On the smaller, cheaper models, a short reply draft costs a fraction of a cent. The bill only gets large when someone sends huge amounts of text to the most expensive model for jobs a small one could do. The real cost is the setup: mapping how your business does the task, connecting the tools and testing the odd cases.

Tip: Before paying for anything new, check the tools you already use. Many booking, accounting, email and store platforms now include AI or automation features you are already paying for and have never switched on.

Where to start: a 30-day plan

Do not try to automate 25 things. Automate one, prove it saves time, then do the next. Here is the order that works.

Week 1, measure. Pick three tasks from the table that you know happen a lot. Log every time they happen for five working days, with minutes. Now you have your own numbers, not mine.
Week 1, pick one. Run the formula. Choose the task with the most hours back that sits in the automate or draft mode, not keep human.
Week 2, write the rules. Write down how you do the task today, the facts it needs (price list, service area, policies) and when it should hand off to a person. This page is most of the work.
Week 2, build it small. Set it up for one channel only, such as website form leads, not every lead source at once.
Week 3, run it in draft mode. Let it work but have a person check every output for a week. Fix the rules each time it gets something wrong.
Week 4, measure again. Log the same task for five days. Compare the hours. If it saved time and the outputs were right, switch on full mode where the task allows it, and pick task two.

Three tasks done properly in 90 days will give most small businesses back five to ten hours a week. That is more than an extra working day, every week, without a new hire.

If you would rather have it built for you, this is what my team does every day. Tell me what eats your week and I will tell you honestly which tasks are worth automating and which are not. If you want to turn these skills into a business of your own instead, read my guide to AI businesses you can start with zero employees.

Takeaways

Owners of US businesses with five or fewer employees work a median 60 hours a week and spend 22.4 percent of it on admin, legal and compliance, about 13 hours (UENI survey of 837 owners, May 2024).
Only 19.8 percent of US businesses reported using AI in any business function as of May 3, 2026, and use at firms under 20 employees did not grow from December 2025 to May 2026 (US Census Bureau, Business Trends and Outlook Survey).
A vendor survey found 66 percent of US small businesses use AI (Thryv, 561 owners, April 2026). The gap with Census comes from who is asked and what counts as using AI.
Hours back per week = times the task happens a week, times minutes each time, divided by 60, times the share the automation handles.
The tasks that give the most time back are the repetitive, customer-facing ones: first reply to new leads, repeat questions, invoices and payment reminders, data entry and booking.
Most tasks work best as AI drafts and a person approves. Full autopilot suits only tasks where a mistake is cheap and easy to undo.
Never let AI make up data. Pull facts from a real source such as your CRM, price list or a government feed, and let the model only read, sort and write around them.
Start with one task. Time it for a week, automate it, measure the hours, then move to the next. Three tasks done well beat 25 half-built.

FAQs

What tasks can AI automate in a small business?

The ones that repeat often and follow the same shape: replying to new leads, answering repeat questions, booking, invoices and payment reminders, data entry, order status, receipts, reports and meeting notes. AI can also draft quotes, listings, social posts, review replies and blog content for a person to approve.

How many hours can AI save a small business?

It depends on how often your tasks happen. Use the formula: times a week, times minutes each, divided by 60, times the share the automation handles. In the examples in this guide, the top 10 tasks add up to about 26 hours a week for a business that has all of them. Automating three tasks well typically gives back five to ten hours a week.

How many small businesses actually use AI?

It depends on the survey. The US Census Bureau found 19.8 percent of US businesses used AI in any business function as of May 3, 2026, with no meaningful growth at firms under 20 employees from December 2025 to May 2026. A Thryv survey of 561 small business owners in April 2026 put it at 66 percent, because it counts owners who use AI tools in any way.

What should a small business not automate with AI?

Decisions about money, people and legal commitments: final prices, larger refunds, hiring and firing, anything filed with a government agency, and the first reply to an angry customer. Also never let AI supply facts. Prices, dates and customer details should come from your own systems, with AI only reading and writing around them.

Is AI automation expensive for a small business?

Running costs are usually low, because most of an automation is ordinary software and only the reading and writing steps use AI, which costs a fraction of a cent per short message on smaller models. The main cost is the setup: mapping the task, connecting your tools and testing the unusual cases.

Do I need to know how to code to automate my business with AI?

Not for simple automations. Many booking, email, accounting and store tools have AI and automation built in, and no-code tools connect them. Custom code starts to make sense when you need your own interface, a login for customers or staff, or rules the no-code tools cannot express.

Which task should I automate first?

The one with the most hours back that does not need a human decision. For most service businesses that is the first reply to new leads or repeat customer questions. For online stores it is usually order status questions or product listings. Measure for a week first so you choose on real numbers.

Originally published at saurabhbhayana.com, where I write about AI automation, AI search and Local SEO for small businesses.