AI skills & training for legal teams · Salt Lake City

Some of the work has to be done right every time. None of it is why anyone hired you.

The check nobody sees. The fourth read of the same agreement. The citation you meant to verify. We find the work in your practice that fits that description — then teach your people to handle it with AI, and build the skills that run it the same way every time.

Reserve a workflow review See a skill running

No charge · one hour · reviews are scheduled at least a week out

The work in question

You already know which tasks these are.

  • 01

    The twenty-minute check before you draft — standing orders, general orders, whether the rule moved since you last looked. Nothing to bill. Nobody thanks you.

  • 02

    The fourth read of an agreement, because the first three were interrupted and you no longer trust your own margin notes.

  • 03

    The template you pull from a prior matter, then spend an hour scrubbing for someone else's facts.

  • 04

    The citation you are almost sure about, that you will check on Monday, that you will not check on Monday.

  • 05

    The summary you write twice — once for the file, once for whoever asks about this matter in four months, which will be you.

None of it is hard. All of it is exacting. It is the part of the work that stays invisible until the one time it doesn't.

Why it keeps happening

It isn't a discipline problem. It's arithmetic.

You know the check. You have done it a thousand times. But it costs twenty minutes, produces nothing anyone will ever read, and stays invisible right up until it isn't. Against a deadline, invisible loses. It will lose again next month.

General-purpose AI was supposed to help here. Mostly it moved the work. Ask a chatbot for the controlling rule and you get a fluent answer with a citation that may not exist — so now you verify the machine as well as the file. That is a task added, not removed.

17–33%

hallucination rate of leading legal-research AI tools

[Stanford RegLab 2024]
87 1,433

court cases citing AI hallucinations, May 2025 to May 2026

[Charlotin database, May 2026]
>$145k

U.S. sanctions for AI hallucinations in Q1 2026 alone; one case reached $110,000

[Charlotin database, May 2026]

The gap isn't raw capability. It's that nothing in a general-purpose tool knows your checklist, your courts, or what you refuse to sign.

What we do

A skill is your own checklist, written down and made to run.

We do two things, in that order. First we teach your people to use AI where the work already lives — including the connectors, which is the part almost everyone skips. Then we build the skills that handle what training alone can't.

A model that can't see the matter is a model you feed by hand, and hand-feeding is its own unrewarded task. Connecting AI to where your documents, matters, and mail already sit is most of the value — and it is the step almost nobody takes.

01

Map the workflow

We follow a real matter end to end and write down every step, handoff, and wait. Not the process as documented — the process as performed.

02

Locate the friction

Where does time pool? What gets redone? What gets skipped under pressure? We separate what AI genuinely helps with from what it should never touch.

03

Teach, then build

Training first, so your team can use what already exists. Then we build the skills that encode your discipline and run it every time.

We build alongside a Utah litigation practice, so what we ship is tested against real filings and real deadlines — not imagined ones.

What one actually looks like

The pre-draft procedural check

[built for a litigation practice]

The friction. Courts change how they want things done — standing orders, general orders, administrative directives — and those changes don't arrive as amendments to the published rules. A motion drafted against last month's procedure can be defective today.

Every careful litigator knows to re-check. It takes twenty minutes, produces nothing billable, and is the first thing to go at 11pm the night before a filing deadline.

The skill. A gate that runs before drafting, every time. It checks the assigned judge's standing orders, the district's general orders, and whether any rule or statute the draft relies on has been amended — then applies the right one. A person skips this under pressure. A skill has no pressure.

Pre-draft gate · procedural currency
matter  civil motion · state district court
✓ judge standing orders — 1 formatting requirement applies
✓ district general orders — no change
‼ rule amended since last known version
↳ current text retrieved and applied
✓ statutory currency — no amendment
conflict rule — the standing order controls within its scope; the general-rule argument is preserved as a fallback.

One module of a larger system we built and use. Yours would encode your practice, your jurisdiction, your checklist. [illustrative run]

How the engagement works

One hour a day. One or two weeks. A flat fee.

You bill by the hour, or you answer to people who do. You cannot disappear for a week to be transformed, and nobody should ask you to.

So the engagement is built the other way around. One hour a day is your time. The building happens between sessions, on ours. You are never asked to sit and watch someone write software.

Days 1–2

We watch the work

A real matter, end to end, as actually performed. We ask a lot of questions and write everything down.

Days 3–4

We train your team

Using AI where the work already lives, connectors included — plus an explicit list of what AI should not touch in your practice.

The rest

We hand over what we built

Built between sessions, then run against your own matters with you, adjusted, and handed over working.

From $4,500 for a one-week engagement. Scope and fee agreed in writing before day one. No hourly billing, no change orders, no meter running on our side either.

After we leave

You keep what we build.

It's yours

The skills we build belong to your practice. Not licensed back to you, not hosted at our discretion, not switched off if you stop paying us.

It runs where you say

On your machine or your own infrastructure. You decide what stays local and what, if anything, is allowed to leave.

Things change — we say so up front

Courts reformat pages, models update, systems move. We tell you what will need maintenance and what it costs. It's documented, so anyone can pick it up. Including someone who isn't us.

The obvious objection

We just told you AI fabricates citations. Here's why that isn't an argument against this.

The failure mode isn't mysterious. A general-purpose model answers from pattern rather than from your sources, and nothing checks it before it reaches you. So the checking is what we build in.

Grounded

Every claim is tied to a source it can show you. If it can't point at one, it doesn't get to assert it.

Reviewed

A second model, from a different vendor, re-checks the first — so one system's blind spot isn't the whole system's blind spot.

Blocked

Anything unverified is blocked rather than flagged. A warning you can click past is a warning you will click past at 11pm.

Recorded

What ships carries a record of what was checked and against what — the thing you'd want if anyone ever asks.

Where your files actually go

The sorting decision runs locally, on hardware you control. Privileged and personal material stays there. Only low-risk work is ever sent to a cloud model — and you set where that line sits. When it isn't sure, it keeps the work local.

We won't tell you nothing ever leaves. That would be false for any useful system. What's true is that you decide what leaves, and you can see the decision.

hares

[working prototype]

We proved this architecture in a contract reviewer that refuses to release unverified claims, keeps privileged material local, catches fabricated citations, and seals an audit record.

It runs on a synthetic corpus. It is evidence that the architecture holds — not a product we're selling you. [synthetic data · U.S. case law only]

Watch it work →
[grounded]

We don't ship claims we can't source — in our software or our advice. If a sentence sounds good but lacks evidence, we cut it.

[private by default]

The routing decision runs locally; privileged work stays on infrastructure you control. When sensitivity is uncertain, we keep it local.

[human-in-the-loop]

AI augments judgment; it never replaces it. Nothing here decides anything. The lawyer signs.

The next step

Reserve a workflow review.

One hour, no charge. We look at how a matter actually moves through your practice and tell you where AI would help — and where it wouldn't. If the answer is that you don't need us, we'll say so.

Reserve a workflow review

hello@denovodev.ai

On confidentiality. A review looks at how you work, not at privileged content. We don't need your client files to map a workflow. Happy to sign an NDA before the first conversation if you'd prefer.

Reviews are scheduled at least a week out.

Tell us where it hurts

Opens your mail client to hello@denovodev.ai. Nothing is stored on this site.