Off-the-shelf chatbot platforms (Intercom Fin, Zendesk AI, Ada) win when you need a support bot live in days and your use case is answering questions from existing documentation. A custom AI agent wins once the workflow requires multi-step actions across your own systems, non-generic logic, or becomes something customers actually pay for rather than a cost center you're trying to shrink.
The decision isn't "AI vs no AI" — it's who owns the orchestration layer. Platforms own it and rent it to you. A custom build means your team owns it, for better and worse.
What you're actually buying with a platform
Chatbot platforms sell you three things: a pre-built UI, a support-specific reasoning layer tuned on their customers' data, and integrations to the tools you probably already use (Zendesk, Salesforce, Shopify). That's a real head start. For a team with no ML or backend capacity, standing up Intercom Fin against your help center can go live in a week.
The tradeoff shows up in three places:
- Pricing that scales against you. Most platforms charge per resolution or per seat. At low volume this is cheap. At the volume where the bot is actually saving you headcount, the bill starts looking like the salary it replaced.
- A ceiling on what the agent can do. Platform automation builders are designed for FAQ answering and simple actions (check order status, issue a standard refund). The moment you need the agent to reason over your pricing logic, call an internal API with business rules, or hand off differently depending on account tier, you're fighting the platform's workflow builder instead of writing code.
- Your conversation data lives in their system. Exporting it for analysis, fine-tuning, or switching vendors later is rarely as clean as the sales deck implies.
None of this makes platforms wrong. It makes them a bet that your use case stays inside their guardrails.
What a custom agent actually costs
A custom agent means you (or an agency) design the reasoning loop, tool calls, and guardrails yourself, on top of a model API (Claude, GPT, or similar). Anthropic's own guidance on building effective agents is a useful baseline: the pattern is a model that plans, calls tools, and checks its own output against a task-specific standard — which is exactly the part platforms abstract away and charge you not to think about.
Building this yourself costs more upfront: engineering time for the orchestration layer, tool integrations, evaluation harness, and guardrails against hallucinated actions. Where it pays off:
- Marginal cost is usage, not seats. Once built, your ongoing cost is largely LLM API spend, which you can tune directly — see our take on LLM cost optimization for the levers that actually move the number (prompt caching, model routing, batching).
- No ceiling on workflow complexity. If the agent needs to check inventory, apply a discount rule, and message a customer on WhatsApp in one flow, that's normal agent design, not a platform feature request. We've written about what this looks like end-to-end in AI WhatsApp sales agents.
- You own the data and the evals. You decide what "correct" looks like and test against it, rather than trusting a vendor's black-box resolution rate. If you're negotiating with any AI vendor, our guide to evals in AI vendor contracts applies whether you build or buy.
The real cost gate isn't "can we afford a custom build" — it's "does this workflow rely on retrieving and reasoning over our own private data." If yes, you'll want a proper retrieval layer regardless of platform; see RAG explained for founders for what that actually requires.
The decision, without the hedging
Ask three questions:
- Is the workflow generic (support FAQ, order status) or specific to your business logic? Generic → platform. Specific → custom.
- Does the agent need to take actions across more than 2-3 systems, or just answer questions? More systems and more branching logic pushes you toward custom, because platform automation builders get unwieldy past a handful of conditional steps.
- Is this a cost-saving support function, or a product feature customers pay for? If it's the latter, a platform badge in your product ("Powered by Intercom") is a real signal problem — customers can tell the difference between a rented bot and something you built.
If you answered "generic / simple / support" across the board, start with a platform and revisit in 6-12 months once you have real usage data. If you answered "specific / multi-step / product feature" on even one, the platform will feel like a cage within a quarter, and the migration cost later will be higher than building it right the first time.
A note on the false middle ground: some platforms now let you "extend" their bot with custom code or webhooks. This is worth evaluating case by case, but in practice it inherits the worst of both — you're paying platform fees and doing custom engineering. Before agreeing to that path, get a straight answer on what the extension layer can and can't touch, ideally in writing.
We've built both — support bots on top of platforms for teams that needed speed, and fully custom agents for teams where the agent is the product. The difference in your workflow's complexity is usually obvious once you list out the actions the agent needs to take, not just the questions it needs to answer. If you're not sure which side of that line you're on, let's talk.
Originally published on the Pykero blog.
Top comments (0)