The problem
Too many tools, too little time
Ask a small business owner where their week goes and you'll hear the same thing: not the work they're good at, but the work around it. Chasing invoices. Confirming appointments. Copying details from email into a spreadsheet. Following up with a supplier.
Each job is small. Together, they quietly take hours. And each one crosses several tools — inbox, accounting, calendar, WhatsApp — so simple automation rules often can't cope with the exceptions.
Why the current approach fails
Owners either do it all themselves, hire help they can't always afford, or try rigid automations that break the first time an email doesn't match the template. Meanwhile the word "agent" is everywhere, usually attached to demos that are far too ambitious to trust with a real business.
The AI thought
Small, bounded, supervised
Here's what's genuinely new. Language models can now plan a sequence of steps, use tools — read a spreadsheet, search an inbox, draft an email — and handle variation that would break a fixed rule. That's what makes something an agent rather than an automation.
But capability is not the same as trust. The useful framing for a small business is:
- One job with a clear start and finish.
- Few tools, with the minimum access needed.
- A human approval step before anything external, financial or irreversible.
What should NOT be delegated
Moving money. Signing anything. Messages that represent the business in a sensitive situation. Deleting records. An agent can prepare all of these; a person should do them.
An agent isn't valuable because it's autonomous. It's valuable when its autonomy has edges.
The solution
An agent with edges
Pick the job
Good candidates recur weekly, follow rules you could write on one page, and have a visible output: a set of drafts, an updated sheet, a summary.
Draw the boundary
Write down exactly what the agent can read, what it can write, and what it must ask about. Grant access accordingly — not more.
Run in shadow mode
For the first weeks the agent only prepares; you act. Every change you make to its output is a lesson about the rules.
Earn autonomy
When a step's output goes unchanged week after week, that step can run on its own. The approval gate for money and external messages stays.
How it works
From problem to outcome
Problem
The owner loses hours each week to back-office chores that span several tools: email, spreadsheets, invoices, calendars.
Input
A clear goal ('chase unpaid invoices older than 30 days'), access to the few tools needed, and the business's rules.
AI / Agent
An agent plans the steps, reads the relevant records, drafts messages and prepares updates across tools.
Decision
Low-risk steps proceed; anything external, financial or irreversible is queued for the owner's approval.
Action
Approved messages are sent and records updated. Every action is logged.
Outcome
A recurring chore that now takes minutes of review instead of hours of doing.
Level 1
Manual
The owner does the chore by hand across several tools.
Level 2
AI-assisted
AI drafts messages and summaries; the owner does the rest.
Level 3
AI-automated
A fixed workflow runs on a schedule with set rules.
Level 4
AI-agentic
An agent plans and executes the job across tools, adapting to each case, with approvals for anything that matters.
Implement this
0/8 done
Identify the workflow
List back-office jobs that recur weekly and touch more than one tool. Pick one with clear rules and a clear finish line.
Identify repetitive tasks
Break the job into steps. Mark which are lookups, which are drafts, and which are decisions.
Define required inputs
What the agent must read (invoices, a calendar, an inbox label) and what it may write. Grant the minimum access.
Select the appropriate AI capability
A language model with tool use is enough. Prefer an established automation platform over custom code for a first agent.
Build the workflow
Run it in 'draft only' mode for two weeks. The agent prepares everything; you do the sending.
Add human approval where required
Payments, external messages, deletions and anything legal require approval. Keep that rule even when it works well.
Measure the result
Time spent on the job before and after, how often you change the agent's drafts, and any errors caught.
Iterate
Loosen approval only for steps where you've stopped making changes. Then — and only then — consider a second agent.
Business impact
What it's worth
- Time. The owner reviews instead of doing.
- Consistency. The job happens every week, not only in quiet weeks.
- Scalability. The pattern — job, boundary, shadow mode, earned autonomy — is reusable for the next agent.
The takeaway
Start with one agent, one job and one approval step. Earn autonomy; don't assume it.
Share this thinking
The newsletter
Think better about AI.
One practical AI idea. One real-world problem. Delivered periodically.