Machine Learning for Legal Practices

Most legal AI talk is about drafting, but the numbers are elsewhere

Much of the conversation about AI in law firms concerns generative tools that draft, summarise and research. Those are useful, and we cover them in our post on AI for the legal industry. The quieter opportunity is predictive: using a firm's own matter history to price work, plan resources and manage cash.

A 50-lawyer regional firm moving more work onto fixed fees faces a hard question every day: what will this matter actually cost to run? Partners estimate from experience, and experience is good at typical matters and poor at spotting the ones that will double in time. The firm's time recording system already holds thousands of completed matters that could answer the question more consistently.

Where machine learning helps a law firm

• Matter cost and duration estimates. Predicting hours and elapsed time from matter type, client, opposing party, value and early events.
• Technology assisted review in disclosure. Ranking documents by likely relevance so reviewers see the important ones first and review can stop sensibly.
• Document classification. Sorting incoming post, emails and client documents into matters and types automatically.
• Lock-up and cash forecasting. Predicting when work in progress will be billed and paid, by matter and client.
• Matter risk flags. Spotting matters whose time recording is running ahead of the usual pattern for their type.
• Conflict and intake triage. Suggesting potential name matches and routing enquiries to the right team.

Estimating fixed fees from matter history

For volume work such as conveyancing, employment claims, probate or debt recovery, matter histories are plentiful and fairly consistent. A model can learn how many hours a given type of matter really takes and which early signals mean trouble: a particular opposing firm, a leasehold title, a contested estate, a late response to the first letter.

1. Export completed matters with type, value, fee arrangement, time recorded and key dates
2. Clean matter types, which are rarely coded consistently across departments
3. Model hours and duration, producing a range rather than a single figure
4. Compare predictions with partners' own estimates on recent matters
5. Use the range to set fee bands and identify matters that need a scope conversation

The comparison step is where buy-in comes from. When the model and the partner disagree, looking at why is often more valuable than either estimate on its own.

A fixed fee is a forecast with a price attached. Most firms set it with one data point: the partner's memory.

Technology assisted review is well established

Predictive coding for disclosure is not new. English courts have accepted technology assisted review since 2016, and disclosure rules in the Business and Property Courts expect parties to consider it on larger document sets. The method is to have experienced reviewers code a sample, train a model on their decisions, rank the rest and keep reviewing until relevant documents stop appearing at a meaningful rate.

What a firm needs is not a bespoke model but a sound process: agreeing the approach with the other side, validating recall on a sample, and documenting decisions so they can be defended. Established review platforms provide the tools. Custom work is more useful around the edges, such as classifying a firm's own document archive.

Professional duties shape every design choice

Solicitors remain responsible for their work and for supervising the tools and people they use. The regulator expects firms to understand the technology, protect client confidentiality and act in clients' best interests. Legal professional privilege adds another layer: client documents should not end up anywhere the firm cannot account for.

Lock-up forecasting and cash

Many firms carry months of work in progress and unpaid bills. Predicting which matters will bill late and which clients will pay slowly lets finance teams and partners chase earlier and plan borrowing sensibly. The data sits in the practice management and accounts systems, and the model is similar to any accounts receivable forecast, with the twist that billing points in legal work depend on matter milestones.

Even a modest improvement in lock-up can release a meaningful amount of cash for a mid-sized firm, which is often a more persuasive business case to a managing partner than any claim about efficiency.

When a firm is not ready

Small firms and firms with inconsistent time recording should not start with predictive models. If fee earners record time in weekly blocks, or matter types are chosen from a list of 300 options used differently by every team, the history cannot support a reliable estimate. Tidying matter types and time recording first is the real project. Our post on software for legal firms covers the systems side.

How we would start

At SpiderHunts we would begin with an export of three or more years of closed matters from one practice area with plenty of volume. Within a few weeks that shows whether hours are predictable enough to price from and what the early warning signs of an overrunning matter look like. Only then would we build anything fee earners see.

That modelling sits within our machine learning development work, and it is always designed to sit inside the practice management system the firm already uses.

Frequently asked questions