The demo is optimized for the wrong hour
Sit through enough AI pitches and you notice the shape of them. There's a moment engineered for the room, where the tool does something nobody expected and everyone leans in. That moment is real, and it's the least useful part of the evaluation, because it tells you what the software does on its best day with a document somebody chose for it.
What a firm actually needs to know is duller. On a Thursday in March, with a badly scanned medical record and a paralegal who has done this eleven times already, does the tool do the same thing it did in January? Can the person using it predict the answer's shape before they click? When it doesn't know, does it say so, or does it produce something plausible and let a human find out later?
Predictable is a specific thing, and it isn't a promise of identical text
Let's be honest about the technology. Language models are probabilistic. Run the same request twice and the wording can shift. Anyone telling you their AI product is fully deterministic is either using the word loosely or selling you something.
Predictability that a practice can actually use lives somewhere else. It means the scope is fixed: the system sees the documents in this matter and nothing else, every time, whether or not anyone remembers to set it. It means the output arrives in the same form, with the same fields, so a reviewer builds a habit instead of reading each result fresh. It means every factual claim carries a pointer back to a page, so checking is following a link. It means a version that behaves a certain way today behaves that way next quarter, because nobody upgraded it while you slept. Those properties are design choices. They survive the underlying model being probabilistic.
Where the surprises get expensive
Variability doesn't cost you at the moment it happens. It costs you downstream, in three places a small practice can least afford.
- Review time. If the output looks different every run, nobody can skim it. Every item gets read from scratch, and the tool stops saving the hours it was bought to save.
- Training. You can teach a staff of six to use a tool that behaves consistently. You can't really teach them to use one that changes, so knowledge stays in the head of whoever uses it most, and leaves when that person does.
- Supervision. A supervising attorney or agency principal is answering for work done with software. That's much harder when the software's behavior last month is not a reliable guide to its behavior this month.
This is the same ledger we described in the verification tax. Every unpredictable output moves cost from the machine back onto a person, quietly, without appearing on any invoice.
What boring looks like when you build for it
Boring is not an absence of effort. It's what you get when a handful of unglamorous decisions are made on purpose.
- A fixed evidence boundary. Work on a matter draws from that matter's documents. Not the firm's whole library, not the open web, not last year's similar case.
- Citations as a requirement, not a feature. A claim without a source doesn't ship as a finding.
- An honest refusal. When a scan is unreadable or the record doesn't answer the question, saying so is the correct output. A tool that always produces something is a tool that will eventually produce something wrong with total confidence.
- Pinned versions. The system that ran your work in January is still available in June, and updates happen when the firm schedules them.
- One shape for the output, so the approval queue reads the same way in week one and week fifty.
- A record of what ran, including which version produced which result, which is the practical half of a tamper-evident audit trail.
Software that changes under you
Here's the part that doesn't come up in the sales conversation. When a tool runs on somebody else's servers, its behavior is theirs to change. A model gets swapped for a newer one. A safety filter gets tightened and starts declining a category of document your practice handles all day. You find out the way everyone finds out, which is that something that worked on Friday works differently on Monday, and nobody can tell you exactly what moved.
For most software that's a minor annoyance. For software sitting in the path of client work, with a licensed professional's name on the result, it's a real supervision problem. The Tiber River Legal Workbench runs on hardware the firm owns, which makes the answer structural instead of contractual. Updates are installed, not received. The firm picks the week. If a new version behaves differently on your documents, you find that out in a test you chose to run, not in a client deliverable.
What you give up, honestly
A system with a fixed boundary won't surprise you with a connection drawn from outside the matter, and one that refuses when the record is silent will sometimes refuse where a chattier tool would have guessed correctly. Those are real tradeoffs. Most firms and agencies I talk to don't want the guess. They want to hand a client a chronology and know where every line came from.
A test for your next demo
Ask to run the same request twice, on your documents, and look at whether the two results point at the same pages. Ask what happens when you feed it something it can't read. Ask who decides when the version you're using changes, and how you'd learn that it did. None of those questions make for a thrilling meeting. They're the ones you'll wish you had asked.
Tell us what your practice needs to be able to count on
The Tiber River Legal Workbench is a matter intelligence appliance for litigation and personal injury practices, running on hardware your firm owns. Predictable behavior is a design goal, and the people who know what has to stay steady are the ones doing the work. We're inviting a small number of Maryland firms to shape it as design partners.
Start the conversationOr write to josh@tiberriverconsulting.com