Tesseract SA

What a small business should not automate

Not everything is worth automating. Here are the five families of tasks to leave alone in a small business, a three-question test to decide, and what to do instead.

Matte 3D illustration: white cubes travel along a grey belt while a single red cube sits apart on a plinth, the one task kept in human hands.

What should a small business not automate?

Short answer: five families of tasks. Those that come up too rarely, those whose rules do not fit on one page, those where a mistake reaches a client or commits money without review, those built on a human relationship, and those nobody has described yet. In these cases you either do not automate, or you automate only the preparation and a person decides.

A studio that automates processes has an obvious incentive to automate everything. We prefer to say the opposite: part of what small businesses ask us for is not worth doing, and another part should never be done. This article gives the five refusals we apply, a three-question test to decide, and what to do instead.

Refusal 1: the task that comes up too rarely

Automation has a fixed cost: design, testing, monitoring, maintenance. If the task seldom occurs, that cost is never paid back. The French outlet Journal du Net puts it this way about AI agents: a task that comes back twice a month almost never justifies an agent, because the time spent building, testing and maintaining it exceeds what it saves.

A simple illustrative calculation. A 20-minute task done twice a month is 40 minutes a month, roughly 8 hours a year. If setting it up and keeping it running takes more than that over two years, automation is a net loss. The same goes for volumes too small to measure quality: Tensoria, an automation provider, cites roughly thirty cases a month as too few to calibrate an AI system, and recommends human review in that case.

To put numbers on your own case, our calculation of the real cost of a repetitive task gives the method in hours and francs.

Refusal 2: the process whose rules do not fit on one page

Automation executes rules. If the person doing the task cannot explain their decisions on one page, there is nothing to execute. Journal du Net makes this its second criterion: when decisions depend on many exceptions or case-by-case negotiation, the rule has to be defined first. The Intelligence Academy says the same: a manual process that is not documented cannot be automated cleanly.

A typical example: “we sometimes give good clients a goodwill gesture”. As long as nobody can say who a good client is, which gesture, and from what amount, a tool will only freeze the arbitrariness. The right first project is then a writing exercise, not a development one: describe the rule, get it approved, and only then hand it to a machine.

For the same reason, rules that change every month are poor ground. Tensoria notes that maintaining a precise prompt costs as much as maintaining classic rules: you move the upkeep around without removing it.

Refusal 3: what goes to a client or commits money without review

Journal du Net’s third criterion is the cost of error. A mistake caught before it leaves the company is cheap. A mistake that reaches a client, a supplier or an authority is not. That covers a bank transfer, a final accounting entry, a contract, a reminder sent with the wrong amount.

Tensoria lists critical processes among the cases to rule out, where a residual error would have consequences too heavy, and recommends human approval or a deterministic rule. Our position is slightly more nuanced: we do not forbid automating these processes, we automate everything that comes before the decision and we keep the decision. The exact placement of that checkpoint has its own article: where to place human approval.

Sensitive data falls under the same refusal. Journal du Net rules out cases where the data cannot be entrusted to a third-party provider on the terms offered. In Switzerland, that starts with the revised data protection act (nLPD): check before sending anything to a model, and have a lawyer review your case if doubt remains.

Refusal 4: the human relationship, when it is the service

Le Blog du Dirigeant, in an article on automation in very small and small businesses, draws the line like this: whatever involves emotion, nuance or a sensitive decision stays human. It names complaints, negotiation and advice, the need for reassurance, and decisions that commit the client. The signs of overuse are recognisable: automatic messages in series, replies that ignore the context.

In practice, an unhappy client should not receive a reply drafted alone and sent alone. An insurance broker explaining why a premium is rising does not delegate that moment. Everything around it can be automated, though: pulling the history, preparing a draft, listing the points to cover. That is the logic of our pages on chasing unpaid invoices or summarising customer service calls: the machine prepares, the person speaks.

A practical test: if the client is buying your availability and your judgment, do not automate them. If the client is buying a result (a document, a confirmation, a report on time), automate the path.

Refusal 5: the process nobody has described yet

This is the most common refusal and the least visible. A task that “has always been done this way” often exists in three variants, depending on who does it. Automating the first variant you meet gives a tool that fails in the other two cases and is then abandoned. The useful work is to watch the task being done, note the differences and decide which variant becomes the rule.

The result shows quickly: once the task is described, some steps disappear (they had no reason to exist), others get simpler, and automating the rest takes less effort. Sometimes the description is enough and the tool becomes unnecessary.

The three-question test

Journal du Net sums up its sorting in three criteria: frequency, stable rules, cost of error. We turn them into questions, with the rule it draws: one criterion missing, keep human approval; two missing, do not automate for now.

QuestionGreen lightAmber or red
Does the task come up often?Several times a day or a weekA few times a month: setup cost exceeds the gain
Do the rules fit on one page?Yes, and they rarely changeNo: write the rule first
Can a mistake be caught before it leaves the company?Yes, it stays internal and reversibleNo: the AI prepares, a person approves

We add two questions that are not in that grid: is the human relationship the service being sold, and has the task been described by someone other than its author? A “yes” to the first or a “no” to the second is enough to slow down.

What to do instead

  • Document the task. One page is enough. It is useful even without any tool.
  • Keep a fixed rule rather than an AI when the rule is simple. Tensoria reminds that AI should not be used where an exact rule is enough: AI then adds cost and complexity.
  • Automate the preparation, not the decision. Gather the documents, draft a message, compute a balance, and let the person decide.
  • Do nothing. A rare, low-stakes task that takes ten minutes can stay manual. That is a decision like any other.

If you hesitate between several tasks, our method for choosing your first process ranks them by real gain. The express diagnostic in five questions gives a first opinion, including a negative one. And you can browse the processes the studio automates: those are the ones that pass the test above.

Frequently asked questions

Which tasks should a small business not automate?

Tasks that come up rarely, whose rules do not fit on one page, where a mistake reaches a client or commits money, that rest on a human relationship, and that nobody has described.

How often must a task occur to be worth automating?

There is no universal threshold. A task that comes back twice a month almost never repays its setup cost. Compare hours saved per year with design and upkeep time.

Can a high-stakes task be automated with human approval?

Yes. The AI prepares the file, a person approves before the irreversible action. It is often the right option when only one of the three criteria is missing.

Should you always use AI to automate?

No. When an exact rule is enough, classic automation costs less and behaves predictably. AI is justified when cases vary too much to be written down in advance.

What should you do with a poorly defined process?

Describe it first: watch the task being done, note the variants, choose the rule. Once described, it often gets simpler, and automating it takes less effort.

Sources
Read next