By Troels Johannesen
Small firms are quietly putting AI agents on quotes, bookkeeping prep, customer replies and scheduling. The tools are useful. The problem is simpler: most owners still cannot say what the agents did last week, how long it took, or who checked the work before it left the building.
Below are five rules that work in a typical small service business. They are about ownership and records, not about buying software.
1. Give every agent a named owner
An agent without an owner becomes nobody’s problem. Put a person next to each one on the rota, the same way you name who opens the shop or who owns payroll.
That person decides what the agent is allowed to touch, when it runs, and what “good enough” looks like. If the agent drafts client emails, the owner is the person whose name would have been on those emails anyway.
Example: in a five person accountancy practice, the office manager owns the agent that chases missing receipts, and the senior bookkeeper owns the one that prepares VAT workings.
2. Log agent work against real tasks
Do not let agent activity float as a mystery line in a chat history. Attach it to the same jobs you already track: the quote for Client A, the month end pack, the support ticket.
When the agent finishes a pass, the record should show the task, a short note on what it produced, and a time estimate. You do not need perfect precision. You need a trail you can read on Monday morning.
3. Record who reviewed it
Unchecked agent output is how small firms create quiet risk. Before anything goes to a client, a regulator, or a supplier, note who reviewed it and when.
The review can be light. Skim the draft, fix the tone, confirm the numbers. What matters is that someone signed for it. If nobody reviewed it, treat it as unfinished work, not as done.
Example: a small building firm lets an agent draft quotes, but nothing goes to the customer until the estimator has checked the materials list and added their initials to the job.
4. Compare agent time with human time
Owners often feel agents are “free” because there is no wage line. That hides the real cost: prompting, fixing, and redoing work when the agent misses the brief.
Once a week, pick a few repeated tasks and compare. How long did the agent plus review take versus a person doing the whole job? Keep the agents where the combined time is clearly better. Drop or redesign the ones that only create cleanup.
Example: if an agent drafts a customer reply in one minute but the reply needs ten minutes of rewriting, while a person writes it from scratch in eight, the agent is costing you time.
5. Stop any agent nobody is checking
If an agent has no owner, no task log, and no review trail for two weeks, turn it off. A quiet agent that nobody watches is not saving time. It is building a backlog of unknown quality.
Stopping it is not failure. It is the same hygiene you already use for unused subscriptions and half finished automations.
A simple weekly check
Spend fifteen minutes each week on three questions: Which agents ran? Who owned and reviewed them? Where did agent plus review beat a person on the same task?
That is enough to keep AI on the rota without turning the business into a surveillance project. You stay in charge of the work, including the work the agents do.
About the author
Troels Johannesen is the founder of Hourtick, a time tracking and task workspace for people and AI agents, and the owner of WeCode, a digital agency in Copenhagen, Denmark.