The real cost of AI for a small law firm, worked out properly

Ask what AI costs a small firm and most answers stop at the subscription. The subscription is real, but it is only one part of the total. This article sets out a five-part cost model a 2-20 fee-earner firm can use before signing anything, with a worked example whose assumptions are all stated, and the two mistakes that make AI business cases collapse.

A five-part cost model

1. Licences and usage fees. The visible line. What matters beyond the headline price is the pricing unit (per seat, per matter, per document, per query) and the tier structure: check the price at the usage level you expect in six months, and read the minimum-seat, lock-in and overage terms before the trial ends.

2. Setup and learning time. Someone configures templates, sets permissions, trains the team and answers questions in the first months. In a small firm that someone is usually a fee earner or the managing partner. Cost it as hours multiplied by that person's internal cost rate - not their charge-out rate, and not zero.

3. Verification time. AI output on legal work has to be checked before it is relied on. The SRA's warning notice on AI misuse records its concern about court and other documents containing false or incorrect information, including citations, as a result of AI misuse - and notes reports of potential breaches received from senior members of the judiciary. Checking is therefore part of the cost of using the tool. Measure it: if a tool saves a fee earner 40 minutes and creates 25 minutes of checking, the net saving is 15 minutes, and 15 minutes is what the business case may claim.

4. Risk work. Where client data goes into a tool, the SRA's position is that firms should understand the contractual and technical safeguards that apply and ensure they are appropriate to the nature and sensitivity of the information being processed - and that paid and free tools alike may not provide the safeguards needed. Closing that gap takes work: reading vendor terms, mapping data flows, tightening permissions. That work is part of the price of using the tool, whether it is done in-house or bought in.

5. Exit cost. Export rights, deletion terms, data portability. A tool that cannot return your data cleanly has a hidden exit fee discovered at the worst time. Check it before signing, not after.

A worked example (illustrative assumptions, not market data)

A 6-fee-earner firm trials a drafting assistant. Stated assumptions:

• Licence: GBP 45 per seat per month x 6 seats = GBP 270 per month
• Setup and training in month one: 20 hours across the team at an internal cost rate of GBP 60 per hour = GBP 1,200
• Vendor terms review and data-flow mapping: half a day, defined here as 4 hours at GBP 60 = GBP 240, once
• Measured gross saving: 10 hours per month across the team
• Verification time: assumed at 30% of the gross saving = 3 hours per month = GBP 180 per month

Month-one total cost: 270 + 1,200 + 240 + 180 = GBP 1,890, against GBP 600 of gross time value - firmly under water, as month one usually is for any tool with real setup.

Steady state, from month two: costs are 270 licence + 180 verification = GBP 450 per month, against GBP 600 of gross time value = a net benefit of GBP 150 per month. (Equivalently: 7 net saved hours are worth GBP 420, against the GBP 270 licence - the same GBP 150.) The GBP 1,440 one-off setup and review cost takes 9.6 months of GBP 150 steady-state monthly benefit to recover. Every one of those numbers moves with the measured saving - which is the point of the model: nothing here can be known from a demo.

The two mistakes that make the numbers lie

Counting gross savings. A demo shows a task done in minutes. It does not show the checking, the failed outputs, or the matters where the tool was the wrong instrument. Only a baseline - volume, hands-on time, turnaround, corrections, measured before and after on one real workflow - survives a partnership meeting. Our evaluation guide (/insights/how-to-evaluate-a-legal-ai-tool/) covers how to run that trial properly.

Skipping the terms. The safeguard and pricing questions that change the real cost - retention, training use, subprocessors, deletion, seat minimums, overage - are all in the vendor's terms, not its website. Our standing question set (/insights/legal-ai-vendor-data-security-questions/) is the checklist version.

Before you buy

Cost all five parts, measure net savings on one real workflow, and read the terms before the trial ends. The free twelve-question readiness score at /readiness/ gives a quick baseline. For a paid engagement: Margo Legal's readiness-audit scope and price are not yet published; the page currently describes the Workflow Value Workshop at GBP 2,500 fixed - one workflow, one half-day, ending in a written recommendation: fix, buy, pilot or stop (/consultancy/). Margo, the product, is in controlled development; nothing above depends on it.

Next step: score one contract workflow in twelve practical questions.