Automating regulatory monitoring without losing the decision

Regulatory monitoring lends itself well to automation: high volume, known sources, and decisions that repeat. Here it serves as the example for a wider question, that of AI in a small business.

For in many companies, AI arrived through a subscription. One employee tried a chatbot, then a second, then the business bought licences. The uses settled in, often helpful, almost always individual and invisible: nobody really knows who uses it, for what, or what difference it makes. There is another way to start: with a real process, broken down step by step to find where a model belongs, and where it does not.

Why the chatbot is rarely the right starting point for AI in a small business

A general-purpose chatbot improves one person’s work. It drafts faster, rewords, summarises. But it does not improve a process: the quote still goes through the same hands, in the same order, with the same waits. The gain stays in the pocket of whoever uses it.

It also raises a question discovered late: that of data. Whatever is pasted into a chatbot, from a personal account or a poorly configured plan, leaves the business without anyone having decided it. We devoted an article on ChatGPT and the GDPR to it.

Nor does it leave a usable trail. Nobody knows which decisions were made with its help, with what degree of certainty, or who checked them. There is no threshold and no clearly assigned responsibility. For an activity subject to quality or compliance requirements, that is a blind spot.

Generative AI in business is not the problem. For writing, it remains an excellent tool, and it would be absurd to go without it. But a writing tool is not a strategy. Strategy begins when you look at an entire process and ask, step by step, what it is made of.

The case: a regulatory monitoring process to automate

The case we take is that of a scientific think tank, bringing together researchers and industry around a heavily regulated field. It has to follow regulatory developments in that field for two reasons: to prepare a professional event that takes stock of them, and to keep its members informed throughout the year.

This case is nothing exceptional. Any small business following a flow of outside information, whether calls for tender, standards, supplier prices or competitors’ publications, faces the same equation: a lot of volume, little time, and a final decision that must remain human.

The constraints are clear. Sources are many and varied: official texts, agency opinions, scientific publications. Reliability is not negotiable: a missed or misclassified development shows, and it puts the organisation’s credibility at stake. The final scientific decision belongs to an expert, not to software. Finally, the data processed must stay within the European Union.

The division of roles was set from the outset. An expert in the field owns the scientific decision: what is relevant, what is important, what deserves to be shared. Maeliom owns the technical implementation: collection, the switches and the assisted writing chain.

Breaking down the process: decisions on one side, writing on the other

The work starts with a list. Not a list of tools: a list of steps, as a human watcher would carry them out. For each, a single question: what kind of step is it?

Step	Nature
Collect the publications	Mechanical, no AI
Is it relevant to our field?	Reflex decision
Which topic, which category?	Reflex decision
Is it urgent?	Reflex decision
Summarise the text	Writing, by an LLM
Validate and share	Human decision

Of six steps, three are switches: closed answers, which fall under what we have called elsewhere reflex rather than deliberation. Only one genuinely calls for writing, and it is the only one where a large language model belongs. Collection needs no AI at all: it is mechanical work that a program does better, and reproducibly. The last step stays human, as a matter of principle.

The right-hand column is not a matter of vocabulary. It says, for each step, what is expected of the tool and how it will be judged. A mechanical step is judged on completeness: was everything collected? A reflex decision, on its accuracy and its ability to flag its doubts. A piece of writing, on its faithfulness to the original text. A human decision is not judged the same way: it is owned.

This breakdown changes how the system is designed. The three switches can be handed to a fast, frugal model, possibly run locally, that returns a level of confidence for each answer. Writing, which is costlier, is only called on the publications that have passed all three filters. The large model works less, and on what matters.

The choices that make the difference

A threshold rather than blind trust

Each switch returns an answer and a probability. Above a threshold, the publication continues on its way. Below it, it goes up to the expert, flagged as uncertain. This mechanism does not try to eliminate error, which would be illusory, but to make it visible. A publication confidently misclassified goes unnoticed; one flagged as doubtful is settled quickly by someone who knows the subject. The threshold itself is not fixed: it is tuned through use, by looking at what the expert corrects.

Hosting within the European Union

The publications collected are largely public. But what the organisation does with them (its priorities, its analyses, the topics it deems urgent) is not, and constitutes its expert work. Processing therefore had to stay within the European Union, with operators whose applicable law is known. We have explored this question, and what it implies for the choice of models, in an article on private LLMs and the GDPR.

The human at the end of the chain, not in the middle

In many AI projects, the human rereads everything: every summary, every classification. The gain then disappears, replaced by tedious proofreading. Here, the expert steps in at two points only: on the cases the system flags as uncertain, and at final validation before sharing. They validate a result; they do not redo the work. That is what makes their involvement sustainable over time, and what guarantees the scientific decision remains theirs.

Applying it to your business: the method in four stages

This breakdown is in no way specific to regulatory watch. It follows the method we apply to any transformation, in four stages.

Understand. Map a real process, as it actually unfolds, not “AI in general”. Who receives what, who decides what, where the work waits.

Position. Sort each step into one of four kinds: mechanical, reflex, writing, human judgement. This sorting is what says where a model belongs, and which one.

Build. Start with a pilot, on a single process, with explicit thresholds and a baseline measurement. A successful pilot on one process is worth more than a vague rollout across ten.

Stay. Track the errors caught, adjust the thresholds, and decide on the basis of that tracking whether to extend, correct or stop. This is the stage at which the gain is measured, by comparing time spent and errors caught against the baseline.

Three processes lend themselves well to a first trial in a small business. Sorting incoming requests, whether they arrive by email or through a form: what kind, for whom, how urgent? Lead qualification: does this contact fit the target, what stage is it at, should it be called back this week? And checking supplier documents (delivery notes, invoices, certificates), whose compliance is often still checked by hand: is the document complete, does it match the order? In all three cases, most steps are switches, and writing only comes at the end, when a reply is needed.

AI in a small business is not a technology project. It is a rereading of its processes, step by step, with one simple question each time: decision, writing or judgement? Once that rereading is done, choosing the tools becomes almost secondary, and the gains become measurable, because you know where to look for them.

This article closes a series of three. The first explains why not every decision deserves an LLM; the second, who sees these decisions go by when they are handed to a model.

Common questions