Seshat AI
Request an assessmentAssessmentFR

All articles

For SMEs

Where to start with AI in a small business: five questions

Five simple questions to ask before choosing a tool, a provider or a budget: the problem, the data, the validation, the measure, the pilot.

By Stéphane Kpoviessi · · 5 min read

In a small or mid-sized company, AI rarely arrives through a precise need. It arrives through pressure: a competitor talking about it, a tool an employee already tried on their own, a salesperson promising gains. The temptation is to start with the tool. That is almost always the wrong door.

Here are five questions to ask before you sign anything, or before you contact us. They are for whoever has to decide: owner, team lead, operations manager. None of them requires understanding the technology.

1. What is the problem, in one sentence, without the word “AI”?

“We want to use AI” is not a problem. “Our salespeople spend hours finding the right version of a contract” is one. “Every tender response starts from scratch” is another.

If your sentence works without the word “AI”, that is a good sign: you have a problem you can measure, and AI may be only one of the ways to treat it. A filing tool, a file naming rule or a better written procedure sometimes settles the matter without a language model.

The situations we see most often come down to three:

  • Knowledge is everywhere and nowhere: procedures, contracts, notes. The information exists, but nobody knows where.
  • The same tasks, done by hand: data entry, reminders, reports.
  • AI yes, but in-house: the documents must not leave for a third party.

2. Who lives with this problem every week, and how much time does it take?

The gain of a project is measured on real people. Name them. Ask them, without imposing a number, how much time they spend on it and what annoys them most.

Write the answer down before launching anything: it is your starting point. Without it, you will never know whether the project was worth it, and you risk judging the result by impression.

3. What data is needed, and can it leave the company?

A useful assistant works on your documents: contracts, quotes, customer files, procedures. Before choosing a tool, sort these documents into three groups:

  1. Open: no risk if they get out (catalogues, public documentation).
  2. Sensitive: a leak would hurt (offers, price lists, customer exchanges).
  3. Restricted: they must not leave (personal data, industrial secrets, files under a confidentiality obligation).

The answer changes the choice of solution. For groups 2 and 3, a self-hosted solution, running on your infrastructure or on hosting you control, avoids sending your documents to a third-party service. That is what we call sovereign AI here: not a label, a decision about where your data lives and who can reach it.

It is not always necessary. For group 1 data, an online service may be enough, and an honest assessment will tell you so.

4. Who validates when the AI is wrong?

It will be wrong. The question is not “how often” but “what happens next”. For each use, decide in advance:

  • what the tool may do on its own (search, summarise, suggest);
  • what a human must always validate (emailing a customer, changing a reference document, committing spend);
  • how an error gets reported.
A document with a passage highlighted in orange, linked to an answer shown in a window, and a hand ticking a box to validate.
A sourced answer can be checked in seconds: the original passage is visible, and a person validates (AI-generated illustration).

A sourced answer, one that shows the passage of the document it came from, changes the game: the team checks in ten seconds instead of trusting blindly. It is the most useful criterion for judging a demo.

Be wary of promises of “fully autonomous” agents. In a small business, what creates value today is a tool that prepares and a human who decides.

5. How will you know it worked, within 30 to 90 days?

A useful pilot is short and covers one use case only. Before starting, define three things:

  • the starting indicator (question 2);
  • the target indicator: time, errors, response delay;
  • the human validation rule (question 4).

At the end of the period there are three honest outcomes: extend, fix, or stop. Saying “we stop” is a good result: you spent a pilot, not a year-long project.

What an assessment should give you

If you bring someone in to answer these questions, you should leave with material you can read, discuss and refuse:

  • a map of the processes concerned;
  • a list of use cases ranked by value, speed, risk and feasibility;
  • a tool recommendation, including when the best answer is not ours (an open source solution, an off-the-shelf tool, custom development);
  • a 30 to 90 day pilot plan, with its indicator.

Where we stand, honestly

Seshat AI offers this path: a costed assessment, a pilot, then operation if the pilot justifies it. The Seshat runtime, the technical foundation, is open source (Apache-2.0) and its code is public. SeshatOS and SeshatCloud are under construction: we do not present them as finished products, and a document pilot is built with you, to measure. Maturity labels are on the Products page.

Keep reading