AI for Law Firms: The Architecture vs Software Problem

Every law firm SaaS vendor currently has "AI-powered" somewhere on their homepage.

Clio. Smokeball. MyCase. Harvey. Spellbook. Legora. Thomson Reuters. LexisNexis. Dozens of smaller specialists. All promising to transform legal work with AI drafting, AI review, AI research, AI intake, AI summarisation, AI billing.

A managing partner at a fourteen-person commercial firm told me in March that she had twenty-three vendor decks sitting in her inbox from the first quarter alone. Each one promising different AI. Each one priced per seat per month. Each one "integrates with your existing stack."

She has not bought any of them. She asked me a question that cuts to the bottom of this entire market: "How am I supposed to know which of these is real and which is vapour, when I cannot even describe my own intake process in one page."

That is the whole problem with AI for law firms in 2026.

It is being sold as a software problem. It is not a software problem. It is an architecture problem wearing a software costume.

Why AI for Law Firms Is a Software Buying Problem (For Most Firms)

The market for legal AI has matured in one specific direction over the last two years. The tools have got meaningfully better, and the category has got meaningfully more crowded. Partners at small firms now face a decision they are badly positioned to make: which tool to buy, from which vendor, in which combination, at what price, for which of the ten or so things AI can now theoretically do.

Most partners solve this by doing what partners have always done when there is a decision they do not have time to think through properly. They delegate it. Sometimes to an operations manager, sometimes to an IT consultant, sometimes to the most tech-enthusiastic associate.

That person chooses based on the best demo, or the best price, or the best sales person, or the tool the firm down the road just bought. A twelve-month contract gets signed. The tool gets deployed. Nobody has asked the question that would have made the choice obvious, which is: what process, specifically, is this supposed to improve, and how will we know if it worked.

Six months later the tool is used by two fee-earners and ignored by everyone else. The partner renews the contract because cancelling is embarrassing. The firm has paid for AI. The firm has not implemented AI.

This is not a failure of the tool. The tools are fine. This is a failure of purchase logic.

The Actual Problem: Nobody Designed the Workflow Before Choosing the Tool

Every failed legal AI implementation I have looked at has the same root cause. Nobody sat down and designed the process the AI was meant to support.

Take a common example. A firm buys an AI intake tool. The pitch was that it would respond to enquiries in under a minute, qualify them, book a call. Real value, measurable outcome.

The tool gets installed. Enquiries come in. The AI responds. Nothing else happens, because nobody decided who picks up from the AI. The AI books a call into a calendar nobody is monitoring. Or qualifies a lead that never gets routed to a fee-earner. Or sends a follow-up email that contradicts the voicemail the receptionist left yesterday.

The AI did its job. The workflow around the AI was never designed. The result is worse than before, because now the client has had three touch-points that do not talk to each other.

This is true for every AI category in legal. AI drafting only works if there is a defined review and approval flow. AI research only works if there is a knowledge base the firm actually trusts. AI intake only works if there is a routing layer downstream of the intake. AI document review only works if there is a defined acceptance standard.

In every case, the AI is the final layer. The workflow is the foundation. Most firms are trying to buy the roof without laying the floor.

3 Places AI Is Ready for Law Firms

There are three workflows in legal where AI is mature enough, cheap enough, and reliable enough to ship genuine value today. In all three, the workflow has to be designed properly first. If you design the workflow, the AI layer is obvious.

1. Structured Document Review

For high-volume, repetitive document review (NDA review, standard contract review, disclosure review, lease review, policy review), AI is ready. Spellbook, Harvey, Legora and others do this well. A trained junior and a well-configured AI can review a stack of contracts in half a day that would have taken a week two years ago.

The workflow that has to exist first: a written acceptance standard per document type. What must every NDA contain. What are the red flags. What is the fallback position on each clause. Without that, the AI produces output nobody trusts and every review gets manually re-done.

2. First-Draft Generation for Standard Matters

For standard correspondence, first-draft pleadings, client-facing summaries and standard advice letters, AI now drafts at the level of a competent second-year. Fee-earners edit rather than write from scratch. Time saved per matter ranges from thirty minutes to three hours depending on matter type.

The workflow that has to exist first: a standard clause and template library per matter type, tied to the firm's actual style, with the review loop clearly defined. If you have that, AI drafting is a straight multiplier. If you do not, every draft starts a new argument about what the firm actually does.

3. Internal Knowledge Search and Matter Research

A properly-configured AI layer on top of your past matters, precedents, and internal advice notes will answer eighty per cent of internal questions in seconds. Associates stop interrupting partners. Partners get an hour a day back.

The workflow that has to exist first: your knowledge base has to be clean. One version of the precedent. One location for the advice note. A shared vocabulary for how matters are tagged. Without that, the AI returns three contradictory answers and nobody uses it after week two.

3 Places AI Is Not Ready for Law Firms

There are also places where the vendor marketing is ahead of the actual tool, and where deploying AI today will cost you more in cleanup than it saves in hours.

1. End-to-End Client Intake

The full flow from "enquiry arrives" to "engagement letter signed" is a sequence of decisions that still need human judgement. Conflict checks, commercial triage, pricing decisions, scope decisions. Vendors will sell you an "AI intake assistant" that promises to handle all of this. What it actually handles is the acknowledgement message and the calendar booking. The rest is still your partners.

Buy the acknowledgement and the calendar. Design the rest of the flow yourself. Do not buy the full pipeline, because the full pipeline does not exist yet.

2. Autonomous Advice

Any tool that claims to give AI advice directly to clients on legal matters is not ready. The liability model is not there, the quality is not consistent, and the reputational risk is real. AI is ready to help a fee-earner prepare advice three times faster. It is not ready to replace the fee-earner.

3. Full Matter Management

The pitch is an AI that runs your matters end-to-end, triggers the next step, sends the client updates, manages the deadlines. In practice, every firm's matters have enough variance that this breaks the moment you hit an unusual matter. You end up with a system that handles the easy matters and creates noise on the hard ones, which is the opposite of what you wanted.

What you can build today: matter management with automated status updates, triggered milestones, and client-facing visibility. Not AI-driven. Rules-driven. AI helps at the edges (summarising a matter, drafting the next email), not at the core.

The Architecture Shift

The question a managing partner should be asking right now is not "what AI should we buy." It is "what workflows do we run, and which of them are costed out properly, documented properly, and measured properly."

For most firms at five to fifty people, the honest answer is: none of them. The intake process is what the intake person does on a Tuesday. The matter management is what the partner remembers to chase. The document review is whatever the associate thinks looks right.

Once you do the architecture work, the AI question becomes trivial. You know which workflow is ready for an AI layer because you know what the workflow is. You know which tool to buy because you know what that tool needs to do. You know how to measure it, because you designed the workflow with a measurement in mind.

Without that architecture work, buying AI is buying a lottery ticket. Sometimes it pays. Most times it does not. You cannot tell which is which in advance.

This is not a criticism of the tools. Harvey is a genuinely impressive piece of software. So are Spellbook and Legora. The problem is that great tools in undesigned processes produce no results, and nobody in the vendor market is incentivised to tell you that.

How to Decide What to Do Next

If you run a five to fifty person law firm and you are trying to work out your AI strategy, here is a simple sequence worth running before you sign another vendor contract.

1. Name your five highest-volume workflows. Intake, matter setup, document collection, drafting, billing, status updates, close-out.
2. For each one, can you describe the flow on one page. Inputs, owners, decisions, outputs, escalations.
3. For each one, how many hours per fee-earner per week does it consume.
4. Which one is the biggest time sink.
5. Before you AI-automate it, design it.

Most firms never get past step two. The process lives in one partner's head. That is the thing that needs to change, not the tool stack.

Once the flow is designed, the AI becomes a layer choice, not a strategy. You pick the tool that fits the flow, not the other way around.

Start With the Diagnostic

If you run a law firm between five and fifty people and at least two of the workflows in this piece are undesigned, the next step is a two-week operational diagnostic. £5K. You get a written audit of the five highest-leverage workflows, a ranked leak list, and a scoped recommendation of the single highest-value fix.

Around sixty per cent of diagnostics turn into a build engagement. The rest walk away with a pricing-grade audit they can use to drive internal change. Either way you stop making tool decisions in the dark.

The 10 Operational Leaks guide is worth reading before any vendor call.