When a frontier AI lab ships a new model, the usual pitch is conversation: open a chat, ask anything. Anthropic's latest cheap model, Claude Haiku 5.5, shipped with a different pitch — delegation. Anthropic president Daniela Amodei called it "a great subagent" for "straightforward tasks that come up again and again," and the model's own docs position it for high-volume work: classification, extraction, routing.
At roughly $0.10 per million input tokens — with cached prompt reads around a penny per million — it's a model positioned less as a chatbot and more as a subagent: the junior hire you give routine work to while you (or a smarter model) keep the judgment calls.
Most small business owners treat AI as one assistant. They open a chat window and hand the AI everything: routine work, judgment calls, sensitive replies. That's like running a company where nobody gets hired below senior level — it works until you see the bill.
Here's the advice that's missing from all the benchmark coverage: the cheap model is not the trick. The delegation plan is the trick.
What a Ten-Cent Junior Is Actually Good At
Cheap, fast models are reliable at work with a knowable shape:
- Classification: is this urgent, routine, or a complaint?
- Extraction: pull action items and client names out of meeting notes
- Summarization: condense a thread into a three-line digest
- First-pass drafting: produce a rough reply that someone else edits
- Skeleton-building: assemble the shell of a report so a human fills in the judgment
What they're bad at: judgment, nuance, novel situations, and anything where a confident-sounding wrong answer costs you a customer.
The Delegation Playbook
Here are four concrete delegations to start with, each with its own escalation rule.
1. Inbox triage drafts
Every morning, a junior agent reads new emails and produces a triage table: sender, topic, urgency, and a draft reply. The draft never sends automatically — it sits in a queue for review.
Escalation rule: If the email is a complaint, mentions legal or refund language, or the agent's confidence on the topic is below roughly 80%, mark it "needs human" and skip the draft entirely.
2. Meeting notes → CRM update drafts
Paste your meeting notes in; the junior drafts the CRM updates — new contact info, next steps, deal notes — as structured records ready to import.
Escalation rule: Any note that mentions a new deal, a risk, or a promise you made gets flagged for you to write yourself. Data-entry is delegable; commitments are not.
3. Weekly report skeletons
Monday morning, the junior assembles last week's numbers into your standard report format, with placeholders for the commentary only you can write.
Escalation rule: If any metric lands outside its normal range, the junior adds a "check this" flag instead of quietly filling in context it doesn't have.
4. First-pass customer replies
The junior drafts responses to routine questions — order status, scheduling, documentation requests — based on your FAQ and macros.
Escalation rule: If the recipient is a high-value client, the thread runs past three exchanges, or the question isn't clearly covered by the FAQ, it goes to you untouched.
Write the Escalation Contract Like a Job Description
Notice the pattern in those four rules. Every delegation comes with three sentences:
- What the junior produces — a draft, a table, a skeleton. Never a final output.
- What triggers escalation — a condition you can actually test: refund language, confidence below a threshold, an out-of-range number.
- What "escalation" means — stop, attach context, hand upward. Not "try harder."
If you can't write those three sentences for a task, you're not ready to delegate that task yet. Start with one where you can.
Why Routing Is the Whole Trick
The benchmark coverage argues about whether the cheap model is "good enough." That's the wrong question. In a small business, no model is good enough unsupervised — the value is that the routing lets you buy 90% of your AI work at ten cents and reserve the expensive tier for the 10% that's actually judgment.
It's also why you don't need to re-architect anything tonight. Pick one recurring task from last week. Write the three-sentence escalation contract. Hand the drafts to the cheap tier for a week. You'll either have a new junior hire, or you'll have learned exactly which part of the task needed judgment — and both answers are worth the experiment.
If you want the full walkthrough with more delegation templates — including the escalation-contract format above — The AI Agent Owner's Playbook covers it end to end, $49 CAD, instant download.
One honest note: we're a $0-revenue small team running a public 14-day experiment on whether this playbook advice actually works — the receipts are here.
SMB Scale Up helps small businesses build AI agent systems that actually work — without the enterprise price tag.
Top comments (0)