For law, accounting & consulting firms

Your firm has answered this before. Stop paying an associate to answer it again.

Precedent, prior advice, matter files and working papers — searchable by meaning, answered with the source paragraph attached. Built for firms where a partner's recollection is currently the knowledge management system.

Book a knowledge assessment Read the results

Fixed fee. Runs inside your infrastructure. Nothing leaves your perimeter without your written sign-off.

Our associates were spending the first two hours of every new matter finding out whether we'd done something like it before. Now that's a search. The partners stopped being a lookup service.

HM Helena Marsh Managing Partner · Marsh & Coyle LLP

Firms live, from 9 to 240 fee earners

4.1 hrs

Recovered per fee earner per week, median across deployments

93%

Median accuracy on firm-written evaluation sets at handover

Client data incidents. Every deployment runs inside the firm's own tenancy

Measured against each firm's own baseline before deployment, not self-reported afterwards. Two engagements are excluded from the median: one stopped at assessment on my recommendation, one where the firm's document estate wasn't ready.

Firms currently running Billable

Marsh & CoyleLLP · 84 fee earners Underhill TaxAccountancy · 41 Bramley AdvisoryConsulting · 27 Kestrel & FinchLLP · 240 Orsett PartnersAccountancy · 63 Ravenscroft LegalLLP · 19

The economics

In a firm, an unrecovered hour isn't waste. It's revenue you didn't invoice.

That's what makes professional services different from every other business buying this. The time doesn't just cost you — it was sellable, and now it isn't.

Re-doing work the firm already did

An associate spends three hours researching a question a partner answered in 2023, because nobody could find the memo. The firm bought that answer once and paid for it twice.

Partners as a lookup service

Your most expensive people spend a meaningful part of the week answering questions that are written down somewhere. That's the highest-cost search function ever built.

Time that can't be written off gracefully

Research time either gets billed — and a client queries it — or gets written down. Both are bad, and the second one is invisible until you look at realisation rates.

Knowledge leaves when people do

When a senior associate goes, so does their map of what the firm has done. The documents remain; the ability to find the relevant one doesn't.

Results

Three firms, measured against their own baseline

Each includes what didn't work. A case study without a disappointing number in it has been edited.

Marsh & Coyle LLP Commercial litigation · 84 fee earners · London Live 14 months

Precedent and prior-advice search across 19 years of matter files

Their document management system had everything and surfaced nothing — search worked on filenames and metadata, so finding a prior position on an issue meant asking whoever had been there longest.

We indexed matter files, advice memos and counsel opinions, with access permissions mirroring their existing matter-level restrictions exactly. First measured pass was 74%; most failures were documents with no useful text layer. Running OCR across the pre-2015 archive took it to 94%.

The unexpected benefit was for the junior end. A first-year can now find how we argued something in 2019 without needing to know that we did. — Helena Marsh, Managing Partner

4.6 hrsRecovered per fee earner per week

94%Accuracy on their 120-question set

£312kAnnualised recoverable time, at their blended rate

What didn't work: we tried extending it to draft first-pass advice notes. Partners rewrote all of them and we withdrew it after five weeks. Retrieval yes, drafting no.

Underhill Tax Tax accountancy · 41 staff · Manchester Live 9 months

Technical answers from HMRC guidance, internal policy and prior filings

Tax is unusual: the source material changes constantly, so an answer that was right in March can be wrong in April. Their problem wasn't finding guidance, it was knowing which version applied to the year in question.

We built date-aware retrieval — every answer states which tax year and guidance version it draws on, and refuses rather than guessing when a query straddles a rule change. It declines about 9% of questions, which the partners consider correct behaviour.

I was braced for it to be confidently wrong about something. Instead it keeps telling us the answer depends on a year we haven't specified, which is exactly what a good junior would say. — Priya Raval, Partner

3.2 hrsRecovered per person per week

9%Of questions deliberately refused

6 wksFrom kickoff to firm-wide rollout

What didn't work: the first version ignored effective dates entirely and produced plausible answers against superseded guidance. That was caught in evaluation, before anyone relied on it — which is the entire argument for evaluating first.

Bramley Advisory Management consulting · 27 consultants · Remote Live 5 months

Reusing prior engagements without re-reading them

A consultancy's real asset is what it learned on past engagements, and it's usually locked inside 200-slide decks nobody opens again. New proposals were being written from scratch against problems the firm had already solved twice.

We indexed deliverables, research and interview notes across six years, with client-confidentiality boundaries enforced at retrieval so nothing from one client's engagement can surface inside another's work.

Proposal writing went from a week to two days, and the proposals got better, because we're now citing work we actually did instead of describing capability in the abstract. — Tom Bramley, Founding Partner

-58%Time to produce a proposal

91%Accuracy on their evaluation set

0Cross-client retrieval leaks in 5 months of audit

What didn't work: the confidentiality boundaries were initially too strict and blocked legitimate internal research. Took three iterations with their risk lead to land it correctly.

In their words

What partners say six months in

He told us in week one that a third of our archive wasn't machine-readable and that we should fix it before spending anything else. That cost him money and bought all of our trust.

JK James KestrelSenior Partner · Kestrel & Finch LLP

Our COLP asked six questions about data handling. Bahman answered all of them in writing within a day, with the architecture diagram. That's not typical of technology suppliers.

SO Sarah OrsettManaging Partner · Orsett Partners

We'd already spent forty thousand with a larger firm on something that never worked. This cost less, arrived faster, and we own all of it. I'm still slightly annoyed about the forty thousand.

DR Daniel RavenscroftPartner · Ravenscroft Legal

How it works

Eight weeks, and the firm owns everything at the end

You don't reorganise the document management system first. Sorting through it is a large part of what you're paying for.

Weeks 1–2

Assessment

What's in the archive, what's machine-readable, where the permissions live, and which practice area to start with. Sometimes this ends with "fix the archive first" — and I'll say so.

Weeks 2–3

Define correct

A partner and a senior associate write 100–120 real questions with the answers they'd expect. This is the accuracy bar, and the firm sets it rather than me.

Weeks 3–6

Build and measure

Retrieval, matter-level permissions, citation enforcement, refusal behaviour. Measured against the question set weekly, so progress is a number rather than an update.

Weeks 7–8

Roll out and hand over

Practice area first, then firm-wide. Runbook, evaluation harness and a session with your IT function, so the firm can run it without me.

Confidentiality

Built for firms with professional privilege obligations

The reason most firms haven't done this yet is not cost. It's that nobody wanted to explain to a client that their file was uploaded somewhere.

Runs in your tenancy

Your cloud accounts, your storage, your keys. I don't host anything and there's no vendor to trust.

Permissions mirror the DMS

Matter-level and client-level restrictions are enforced at retrieval. Nobody sees through a wall they couldn't already see through.

No training on your data

Contractual no-training terms with the model provider, or fully local models where the classification requires it.

Full retrieval log

Who asked what, which documents were returned, what was shown. Retained to your policy, available to your COLP.

Answers cite the source

Every response links to the paragraph it came from, so a fee earner verifies rather than trusts.

Refuses rather than guesses

Outside the corpus, it says so. In this sector a confident wrong answer is the only unacceptable failure.

Investment

Published, so you can take a number to your partners' meeting

Fixed fees. Nothing is priced per seat, because charging you more for growing your own firm is a strange way to do business.

Start here Knowledge assessment

Is your archive ready, and which practice area should go first?

$6,500

Two weeks · credited against a build

Audit of your document estate
What's machine-readable and what isn't
Recommended starting practice area
Written report for your partners' meeting
An honest "not yet" where that's the answer

Book an assessment

Practice area build

One practice area, live, measured against your own question set.

$34k – $58k

6–8 weeks · assessment credited

Full retrieval over that area's archive
Matter-level permissions enforced
Evaluation harness your IT team can re-run
Retrieval logging for your COLP
Accuracy bar agreed before work starts

Discuss a build

Firm-wide & ongoing

Every practice area, plus someone keeping it accurate as the firm changes.

from $4,500/mo

After the first build · 60 days notice

Additional practice areas rolled out
Quarterly accuracy review with a written report
Model upgrades tested before they ship
Your IT team trained to take it over
Explicitly designed to become unnecessary

Talk about firm-wide

The arithmetic your finance partner will do anyway

A 40-fee-earner firm at a $220 blended rate, recovering three hours per person per week across 46 weeks, is roughly $1.2m of recoverable time a year. Even at a 25% realisation rate on that, a $45,000 build pays back inside a quarter.

I'd rather you ran that calculation with your own numbers before we speak — and if it doesn't clear comfortably, say so and I'll tell you honestly whether a smaller starting point makes more sense.

Model and infrastructure costs sit in your own accounts, typically $200–$900 a month depending on firm size. Not marked up.

Questions

What managing partners ask first

How is this different from what our DMS provider is offering?

Your DMS vendor's AI works on the documents inside the DMS, which is a real advantage if everything lives there. Ask them two questions: what accuracy it achieves on your own material, and whether you can measure that yourself. Most can't answer either.

The other difference is scope. Firms rarely keep everything in the DMS — there's advice in email, working papers in shared drives, counsel opinions in someone's folder. I build across all of it, and I'll tell you during the assessment if your DMS vendor is genuinely the better option for you.

What about privilege and client confidentiality?

Everything runs inside your own tenancy, with permissions mirroring your existing matter and client restrictions enforced at the retrieval layer — so nobody can reach a document they couldn't already open. Every retrieval is logged and available to your COLP.

Where the classification demands it, fully local models mean nothing crosses your perimeter at all. That's a capability trade-off we'd discuss in week one, not a surprise later.

Will it hallucinate a case or a rule?

It answers only from documents it retrieved from your archive, with the citation generated from the retrieval rather than from the model — so an answer with no source cannot reach a fee earner. Below a confidence threshold it declines rather than answers.

Then it's measured: your own 100+ question set, run before rollout and quarterly afterwards, so accuracy is something you watch rather than assume. That measurement is contractual, not a promise.

Half our archive is scanned PDFs. Is that a problem?

It's the single most common finding in the assessment, and it's solvable — OCR across a legacy archive is a known quantity, and it's usually the difference between 70% and 90%+ accuracy.

What matters is knowing before you commit to a build rather than discovering it in week five. That's most of what the assessment is for.

Our fee earners won't use new software.

Correct, and the firms where this works are the ones that put it where people already are — inside the DMS interface, or as a search box on the intranet homepage, not a new application with a new login.

The other thing that drives adoption is citations. Fee earners trust it once they've clicked through to the source a few times and found it was right. That's why the citation isn't a feature, it's the whole design.

You're one person. What if you're unavailable?

Everything runs in your infrastructure, in your repositories, with a runbook and an evaluation harness your IT function can operate. No proprietary layer and nothing that stops working if I do.

I take two clients at a time and will tell you a start date rather than overcommit. For firms that require a corporate counterparty on the panel, I'd say so early rather than waste your procurement process.

Who does the work

One engineer, and the same person from first call to handover

I'm Bahman Shadmehr. Before this I spent eight years building systems where quiet failure was expensive: automated trading platforms placing orders on the Texas power market, a national payment gateway I helped decompose while it was live, and the infrastructure underneath both.

That's the relevant experience for a firm, though it isn't obvious why. Those systems ran unattended and had to be trusted, which means the engineering was mostly about knowing when something had gone wrong — measurement, logging, and refusing to act when confidence was low. A knowledge system inside a firm has exactly the same requirements, for exactly the same reason.

You'll deal with me directly throughout. There's no account manager, no offshore team, and no handoff between the person who sold it and the person who builds it.

2024 — 2025Automated energy trading, ERCOT — order pipeline, monitoring, alerting

2023 — 2024US equities trading platform — AWS services on the trading path

2020 — 2021Zibal payment gateway — led monolith-to-microservices migration, live

2022 — 2023Vgang — Kubernetes platform, integration microservices

2017 — 2020Rahpa — async backends, real-time systems

Python, Go, Kubernetes, AWS, GCP, Postgres, Redis. Remote, UTC+3 — European hours and US mornings. 76 technical posts · LinkedIn

Next step

Ask your associates what they looked up twice this month

Then email me the list. It's a better brief than any discovery call, and I'll tell you within a day whether your archive can support answering it.

info@bshadmehr.me

Two firms at a time Reply within one business day NDA before anything substantive

Useful in a first message

Firm size and practice areas.
What your document management system is.
Roughly how far back the archive goes.
Whether the older material is scanned or native.
Who'd need to approve this — and what they'd worry about.