Deterministic AI: does it even exist?

A language model computes probabilities, a template follows rules. Why the two can still be combined, and what that produces for document creation.

The term contradicts itself

Determinism means same input, same result. Every time. That is exactly what rule-based systems have delivered for decades. Select “fixed term: yes” in a template and you get the fixed-term clause that is stored there. Word for word, in the same place, in the right format.

Large language models work differently. They compute probabilities for the next word. Even with an identical prompt the wording can vary, and that variance is not a defect but the principle. It is the reason AI is so strong at open tasks: summarising, translating, rephrasing, pulling what matters out of unstructured material.

“Deterministic AI” therefore sounds like a contradiction, and strictly speaking it is one. Something else is meant: a system whose result is predictable even though AI was involved in producing it. That is achievable as soon as you stop making the AI responsible for everything.

Where AI is good and where rules are better

The dividing line becomes clear once you look at the individual steps of document creation.

AI is strong at reading a file, an email thread or an informal description and working out what it is about. It recognises that “delivery from Q2, offtake 12,000 units, penalty for delay” describes a framework supply agreement with a liquidated damages provision, and maps the details to the right fields. It can suggest a suitable template when someone describes in their own words what they need.

It is weak where binding commitments are required. Which limitation of liability clause is permitted? Above which order value does the second signature apply? Which style sheet matches the firm or company layout? Questions like these should not be answered afresh each time, and certainly not on the basis of probabilities.

That split is the whole trick. The questionnaire of a template can be filled in with AI. Generating the document then follows the stored rules exclusively — the core of rule-based document automation. The result always looks the way you expect it to, and the route there was fast all the same.

”AI writes well, but these are not my clauses”

We hear this sentence in almost every conversation with law firms. It captures the problem more precisely than most commentary.

Lawyers rarely dispute that a language model produces a decent contract text. It reads cleanly, and the structure is usually sound. It simply does not come from their own house. The liability clause is not the one they have used for ten years and that has held up twice. The definitions differ, the choice of law is phrased differently.

That starts review work which eats the time saved. If every generated clause has to be checked against your own standard, little has been gained. If the comparison is skipped under time pressure, a risk arises that nobody would take knowingly.

This is exactly where the combination comes in. The clauses come from your own template and are fixed. The AI does not decide how the clause reads, only which answer in the questionnaire applies. That answer can be checked in seconds, because it comes from a manageable set of options.

What this looks like in practice: the Word add-in

At Lawlift this division of labour is implemented in the Word add-in, which is to say where legal work already happens.

The process begins with choosing a template. Instead of clicking through a folder structure, you describe in your own words which document you need. The system proposes matching templates with a degree of fit.

The template’s questionnaire then opens: predefined answers that determine which clauses go into the document in which form.

And then the AI reads an unstructured set of material — an email thread, a PDF or Word file, even images — and answers the questionnaire from it. The user checks the results and, at the end, presses “Generate document”. The document is then produced, with all approved clauses and the right layout.

The same process in chat: the route through MCP

The second point of access is the Model Context Protocol. Lawlift can be connected through it to any MCP-capable chat.

In the chat you search for the template, fill in the questionnaire and generate the document. There are three ways to fill it in, and they can be mixed: answer the questions yourself, give additional context in the chat, or point to a file and ask the chat to fill the questionnaire from it.

The last route uses the strength of AI exactly where it lies. This step still has to be checked, because a model can miss a fixed term or map a date incorrectly. Sign-off stays with a person.

What happens after that is fully determined again. The document is assembled according to the stored rules and comes out of the chat fully formatted. No reformatting, no surprises in the numbering.

The side effect: almost no AI costs

If you have a language model write a twenty-page document end to end, you pay for every word it produces. At high document volumes that adds up, and the cost rises linearly with the quantity. If the same document is assembled by rules, the document text consumes practically no tokens. The template supplies it and the software puts it together.

Tokens are only consumed by the part where the AI is genuinely needed: understanding the inputs, mapping them to the questionnaire, individual adjusted clauses. That is a fraction of the total text.

The cost advantage is not the actual point, but a consequence of the same decision. Use AI only where it contributes, and you get more reliable documents and lower running costs at the same time.

Where to start

Take a document you produce often and whose clauses are uncontroversial in-house. Sales and procurement contracts work well, because the volume is high and the standard is clear. Build a template with a questionnaire from it. AI-supported templating shortens this step considerably, because variable passages and conditions are proposed automatically from an existing document.

Then decide where the AI should help. Filling in from the file? Adjusting individual clauses? You make that decision per use case, and it is reversible.

That answers the question in the headline. Deterministic AI does not exist. A deterministic result despite the use of AI very much does, and in practice that is what counts. What it looks like when this foundation is missing is described in Shadow AI in contract drafting. Our whitepaper AI vs. rule-based document automation goes deeper.