The Assistant That Shows Its Work

You’re right to be suspicious

Start with the objection, because it’s correct: a language model, on its own, is a machine for producing plausible text. Ask it a question and it will answer – fluently, confidently, and with no internal difference between remembering and inventing. In most industries that’s an inconvenience. In work that ends with a professional seal, it’s a hazard: a confidently invented coordinate, a plausible-sounding prior job, a bearing that was never computed – any of them, acted on, becomes someone’s liability with a license attached.

So the question we built around was never “what can AI do?” It was: what would it take for someone who stamps their work to act on an answer? Not enjoy the demo – act on it. That question has a specific engineering answer, and it comes in three rules.

Rule one: computed, never generated

The most important decision in Theo and Heath is a division of labor that never gets crossed: the engine calculates; the model narrates.

When Heath runs a bearing and distance, the trigonometry happens in deterministic code – the same arithmetic every time, to the same decimal. When Theo reports what jobs sit within a half mile of a parcel, a spatial query found them – not a model’s impression of what might be nearby. When Heath summarizes which agreements have gone quiet, the workflow engine flagged them against real dates and thresholds before the model said a word.

The language model’s job is the last step only: turning computed results into a clear answer. It is never asked to produce a number, a coordinate, a date, or a status from its own imagination – because narrating a fact and inventing one are, to a model, the same act. The architecture’s job is to make sure it only ever has facts to narrate.

Rule two: every claim carries its receipt

Open any answer from Theo and the sources are attached: the job number, the field book, the page, the scan date. Ask Heath about a parcel and the answer points at the agreement, the payment, the permit it came from. Not “sourced from your records” as a vague comfort – this record, right here, click it.

That does two things. The obvious one: verification takes ten seconds. A professional can check the answer against the instrument it cites and then act, which is the entire trust transaction in a liability profession – the same standard the profession already applies to its own work.

The less obvious one: citation is a leash on the whole system. An assistant that must attach a source to every claim cannot drift into fiction, because fiction has no source to attach. The requirement doesn’t just help you catch errors – it structurally prevents the class of answer that sounds right and traces to nothing.

Rule three: no answer beats a plausible one

The hardest behavior to build into an assistant is the willingness to disappoint. Models are trained to be helpful; the failure mode of helpful is filling silence with something that sounds right.

So the third rule is explicit: when the record doesn’t contain the answer, say so. Ask Theo about prior work on a site the firm has never touched and the honest response is “nothing found” – not a maybe, not a hedge dressed as an answer. Ask Heath about a clause no agreement contains and you get the absence, stated plainly. In this work, “we don’t have that” is a useful fact – it tells you what diligence is still owed. A fabricated answer is worse than no assistant at all.

Why this survives a technical audit

Here’s the practical difference this architecture makes, beyond the daily workflow. “We use AI” is a claim anyone can print on a slide. An answer that traces – this number came from this computation, over these records, cited here – is a receipt. When a utility’s IT team, a client’s counsel, or your own E&O carrier asks how the assistant works, the answer isn’t a shrug about model behavior. It’s an audit trail: deterministic computation you can re-run, sources you can open, refusals you can see.

That’s the bar we think every AI tool in this industry should be held to – ours included. When a vendor says “AI-assisted,” ask the three questions: Where do the numbers come from? Where’s the citation? What does it do when it doesn’t know? The answers separate the assistants from the party tricks – and your profession, of all professions, already knows how to demand shown work.

Trust in AI isn’t a feeling. It’s an architecture. Ours shows its work.