Skip to content
COR Intelligence
Back to latest

Why we built a FLIT drafter using local AI, and saved hours of partner time per will

Pete Gypps11 May 2026 · updated 27 August 2026
8 min read1,821 words

Lawyers are right to be scared of cloud AI. So we built a will-drafting tool that does not use it, the AI runs on a computer in the firm’s office, with no internet connection and no client data ever leaving the building. Along the way, it cut a one-to-three-hour partner task down to about thirty minutes, and brought the four-to-eight-hour complex matters under control too.

A solicitor’s hands writing on legal documents at a desk, with the headline "Local AI · No cloud, will-drafting that stays in your office" overlaid in white type on a deep navy tint.
Local AI, running in the firm’s office. No client data crossing the internet.

What if a tool that saved your senior team hours on every will managed it without sending a single piece of client information out of your office? That was the question we asked when a STEP-qualified private-client firm came to us looking for a faster way to draft FLIT wills, without going anywhere near cloud AI.

The firm had two requirements they would not negotiate on. First, no client data could leave the building. Second, the senior solicitor had to be the last word on every draft. Everything else was on the table.

We built them a tool that runs on a computer in their office. The partner pastes their meeting notes in, the AI fills in the firm’s FLIT template, the partner reviews the judgement-critical clauses and signs off. On a straightforward matter it brings a one-to-three-hour task down to about thirty minutes. On the complex ones (unusual property, tax positions, family structures) it brings what used to be four-to-eight hours under control too.

What is unusual about this (and what almost nobody in legal-tech is talking about) is the AI itself. It runs on a machine in the firm’s office. No cloud connection. No call to OpenAI, no call to Anthropic. The technical term is "local AI models", small AI systems that run entirely on hardware you own. The legal term is: confidentiality, untouched.

The drafter is the visible bit. What I want to talk about is what is underneath it, because the underneath is what makes it defensible, and the underneath is exactly what most legal-tech AI offerings are skipping.

What it does, in one paragraph

The solicitor opens the application on their own workstation. No website, no login, nothing leaving the office. They paste their meeting notes into a single box, exactly the kind of notes they would jot during a client meeting, unstructured, often messy. The system reads them, fills in every field a FLIT will requires, and surfaces anything ambiguous as a question for the solicitor before any draft is produced. The solicitor reviews and edits everything the system extracted (full control, with the firm’s template strictly followed) then clicks produce. A Word document in the firm’s house style appears on disk. Everything from "I paste" to "I have a draft" takes well under a minute on machines most firms already own.

That is the visible part. Here is the part most "AI for law" offerings skip.

Every draft leaves a complete record

Before the draft is shown to the solicitor, the system saves an audit record on the firm’s own server. The record holds: who pressed the button, when, on which machine, against which version of the firm’s template, using which AI, producing what document, down to an exact fingerprint of the file itself. A second copy is saved straight away to a separate place the firm controls (a second machine in the office, or an encrypted cloud copy) so that even if someone got into the main record and tampered with it, the second copy catches the change.

If the firm chooses the cloud copy, the encryption happens on the firm’s own machine before anything ever leaves the office. The key stays in the office. What the cloud provider stores is encrypted data they cannot read, not their staff, not their administrators, not a future buyer of their business. It is the only kind of cloud backup that is compatible with a regulated firm’s duty of confidentiality: the cloud is paying for storage, not access. If the firm later wants to walk away, they take their key and the cloud copy becomes useless to anyone else.

The records are linked in a chain. Every new record carries a fingerprint of the one before, and is stamped by the machine itself, not by a file someone could copy off and reuse from another computer. An independent check walks the chain from top to bottom and reports any break in under a second. We ran a deliberate test where someone with full administrator access quietly changed one field of one record, the check flagged it immediately, with the exact record and the exact point the chain broke.

No draft comes out without a paper trail your regulator would accept. No exceptions, no fallbacks, no "we’ll fix it next render."

The principle the system is built around

If for any reason the record cannot be saved, the draft is deleted and the solicitor sees an error. That is how the principle holds in practice.

Why that matters for a private-client team

A solicitor’s working file has to survive the testator. Often by decades. Will files in particular live for the testator’s lifetime plus six years post-death, practically speaking, fifty to a hundred years. That horizon makes most AI offerings nervous, because most of them weren’t built for it.

The questions a regulator, an opposing solicitor in a contested probate, or an ICO inspector will eventually ask are entirely predictable:

  • Who drafted this clause, on which machine, on which day?
  • Has anyone changed the firm’s template between then and now? If yes, when, and who?
  • Was the same model used, or a different one? Can you prove it?
  • If any client data left the firm’s premises, where, and on what basis?

"Trust me, it is in the case management system" does not answer any of those. A complete record, linked in a chain, that re-creates itself end-to-end from a check the firm itself controls, does. The whole point of a system that does the work is that it leaves a record you can actually defend.

The other answer the audit trail unlocks: the firm’s data never goes anywhere. The system runs on a computer in the firm’s own office. Nothing crosses the internet, no third party is involved, no processor agreement to maintain on top of everything else, no Article 28 mapping exercise. The AI lives on a machine the firm owns, and that machine lives in the firm’s office.

Why we said no to cloud AI

Most of the AI offerings landing in solicitors’ inboxes right now send your data to Anthropic, OpenAI, Google, or one of the other cloud providers. The reasons are obvious: it is the fastest way to build something, the AI is very good, and most of the engineering work is taken off the table. The whole industry (including most of our own clients in other sectors) is running to those services.

For a private-client legal firm drafting wills, we think that’s the wrong call. Not a slightly wrong call. A wrong call you’ll be answering questions about for the rest of your professional career.

When a solicitor pastes a client’s notes into a tool that sends them to OpenAI or Anthropic, what actually happens is: the client’s name, the testator’s address, the spouse’s name, the children’s names and dates of birth, any health conditions mentioned, any reason for disinheriting someone, and the entire free-form context the solicitor wrote into their notes, leaves the firm’s premises, crosses the public internet, lands on a third-party server somewhere in the US or Ireland or wherever the provider routes it that hour, gets processed by software the firm has never read and never will, and may (depending on the provider’s current policy and the option the firm has signed up to) be stored, logged, used to train future AI systems, or all three. Most providers offer "zero retention" options; most of those cost more, and the firm has no independent way to verify the policy is being honoured.

There are firms for which that risk is acceptable. A solicitor drafting a will is not one of them. Solicitor-client privilege is a structural duty, not a preference. It survives the client’s death. The Solicitors Regulation Authority’s outcomes-focused regulation puts the burden on the firm to demonstrate adequate protection of client confidentiality, not on the regulator to demonstrate inadequate protection. A subpoena issued to the AI provider (and there will be such subpoenas) is not a hypothetical risk: it’s a question of when, not if, the first one lands.

The AI we use in the office is smaller than the giant cloud ones. Of course it is. But for a narrow task like "turn these notes into the fields of a FLIT will," the modern smaller AI systems we use are entirely capable, and improving fast. We have measured the trade-off carefully. The firm gets a draft that needs a partner’s review (which it would need anyway) in exchange for never sending a single piece of privileged client information outside the office.

If the wider industry’s run to cloud AI is the bet, ours is the counter-bet: that the firms still standing in five years will be the ones who never had to explain a data-exposure incident to the SRA, the ICO, or the Law Society Gazette.

This is one piece of work

FLIT drafting is one piece of work that fits the pattern. The same shape, paste in the kind of notes the firm already has, surface the matter-specific questions to the right person, produce the firm’s house-style document, leave a complete record, works equally well for conveyancing packs, suitability letters, fee-note generation, year-end accounts triage, AML triage, conflicts checks, file-review summaries, and several others.

We do not think every firm should start with FLIT. We think every firm should start with whichever piece of work their senior fee-earner spends the most time on that is not actually their senior judgement. For most owner-operator legal firms, that is not a mystery. Most partners can name the document they keep complaining about drafting before you have finished asking.

What this looks like for your firm

We build one focused piece of work at a time, on machines the firm owns, in 90 days. Before the build, a two-week paid review of your last 30 days of senior fee-earner time names the single piece of work with the biggest recoverable hours. The review credits 100% against the build cost if you proceed; refundable if no credible plan comes out of it.

The build comes with a written guarantee: the review names a specific recovered-hours target for the named senior fee-earner. If they are not hitting at least 80% of that target ninety days after the system is in place, we continue at no further cost until they are.

We are a small firm and we do one engagement at a time, properly. If you want to walk through your last 30 days and find out where your senior hours are actually going, an initial call is sixty minutes. Reply to enquiries@corsolutions.co.uk to book it.

Useful? See more of us on Google.

One click adds COR Intelligence to your preferred sources in Search and its AI features.

Read next

More from COR Intelligence