The seat-count framing is misleading
AI employees are sold like headcount, which invites the question “how many”. The better question is how many distinct jobs you have that are high-volume, verifiable, and documented — and for most small businesses the honest answer is one or two.
The framing matters because seats imply the constraint is budget. It isn’t. The constraint is how many jobs you have of the right shape, and how much reviewer attention exists to keep them honest.
Count jobs, not departments
Write down every recurring task in the business, then keep only those that are done at least weekly, have a checkable right answer, and already have a written rule or could have one. Most small businesses producing this list honestly end up with three or four candidates, of which one is clearly the best starting point.
Each one carries fixed overhead
Every role needs its own scope, knowledge, evaluation, and reviewer. That overhead does not shrink with the second one, which is why scaling from one workflow to five is not five times the first.
Some of it does become shared — voice, prohibitions, escalation policy and disclosure are defined once. The per-role costs are scope, knowledge, evaluation and review, and those are precisely the expensive ones.
One AI employee doing a job properly beats four doing jobs approximately.
Start with the job that generates the most volume
Whatever your team does most often, most repetitively, with the clearest right answer. For most small businesses that is inbound enquiries or operational chasing rather than anything strategic.
The limiting factor is reviewer time
In a small business the same person owns the knowledge, reviews the output, and does the actual job. Two AI roles pointing at one reviewer means neither gets reviewed, and an unreviewed system decays quietly.
Before adding a second role, ask the intended reviewer what they would stop doing to make room. If there is no answer, you are not adding a role — you are adding an unmonitored system with the same name as one that works.
Expand on evidence, not on ambition
Add a second role when the first has a correction rate that has stopped falling, a measurable return, and a review routine that has survived a busy month. All three are observable facts, unlike the projections in a proposal.
That last condition is the one teams skip. A review habit that holds during a quiet quarter and collapses at year-end has not been tested, and year-end is exactly when an unreviewed system does its damage.
The one-role variant nobody sells
A single AI teammate covering several small tasks within one workflow — drafting, summarising and chasing inside the same process — is often the right answer for a very small business, because it shares one scope, one knowledge base and one reviewer.
It is sold less often because it counts as one seat. It is frequently the version that works.
The honest ceiling
Small businesses tend to plateau at two or three well-run roles, because that saturates the work with the right shape. That is not a disappointing result — it is a bounded, maintainable system that keeps paying back, and it beats five roles nobody has looked at since launch.