The Stack You Didn’t Choose.

Here is how AI shows up in an in-house legal team in 2026.

IT rolls out Microsoft Copilot across the business, often thinly deployed, and legal gets it whether it wanted it or not. The contract platform announces AI features at renewal, priced into the uplift, with no procurement decision for legal. Three panel firms arrive with three different AI portals, each hoping its platform becomes your platform, or at least makes you symbiotic with it. Some may even white-label you an AI product of their own.

Other departments run their own AI projects, some already making calls with legal consequences. Two of your lawyers are quietly using personal ChatGPT accounts; in Thomson Reuters’ 2026 Future of Professionals survey, roughly a third of professionals admitted using AI tools their organisation hasn’t approved.¹ And somewhere in the budget cycle, someone asks whether you should buy Harvey or Legora, which you cannot afford on your existing budget.

That is six doors, five pricing models, and one deliberate decision. The stack assembles itself. The question is whether you govern it or discover it.

The strange part: in-house counsel are the most powerful buyers in this market and the least powerful buyers in their own building. In Litera’s Spring 2026 survey, 51% of law firm respondents said a client had directly influenced an AI investment decision in the past year; only 15% said their AI investment was entirely internally driven.² Meanwhile, your own AI stack may have been chosen by an IT committee you have never met. You dictate to the firms you buy from, and you get dictated to by the business.

There is a reason for the squeeze. In-house legal is a service provider at the heart of the business, which is exactly why so many of these decisions are not legal’s. Law firms sit above you, keen to show off their AI. Business units press in from every side, expecting legal to keep up. In the middle sits a legal team whose own stack may be the least deliberately assembled of anyone’s. And trust inside legal is thin: in Factor’s 2026 survey, 83% of lawyers had access to AI tools but only 22% reported high trust in them.²⁷

This is how I would try and untie the knot: understand the products, understand your constraints, and then, most importantly, understand your needs. Skip to the third part if you only read one thing.

Start by watching the big firms. They publish their moves, the press covers them, and they can afford to try everything. Their AI strategies have split three ways.

Sit with that. The best-resourced legal buyers in the world looked at the same market and split three ways, with many running more than one strand at once. There is no consensus play; the closest thing is hedging. Treat anyone selling you certainty accordingly.

The same three strategies exist in-house, at smaller scale. But in-house is the harder problem. A firm decides once, for itself. You have a stake in your firms’ decisions as well as your own, and their choices land on your desk as portals, workspaces and AI-drafted work. And you face doors no law firm has to worry about, so your map needs more rows.

The market looks chaotic because vendors describe themselves by what they’d like to be. Ignore the marketing and every AI product or initiative that touches an in-house team comes through one of six doors.

When a vendor demo refuses to fit the map, ask these four questions. The answers place it.

Two complications. First, the types are collapsing into each other: the labs are moving up into legal product (Anthropic now sells Door 1 and, since May 2026, something close to Door 2⁴), the copilots are absorbing legal agents, and the incumbents are bolting AI on. That is what the four questions are for. Second, the industry’s own leaders say the categories are unstable. Harvey’s CEO, 12 May 2026: “long term we would end up competing with the model companies.”¹⁴ Legora’s CEO: models are “necessary … but insufficient on their own.”¹⁵ OpenAI’s president, in a post reported on 22 May 2026: “the model alone is no longer the product.”¹⁶ When both sides of the table agree the ground is moving, get ready for earthquakes.

Product knowledge is useless until you face three constraints: who holds the authority (and the budget), how pricing is shifting, and the fact that products die.

Start with an uncomfortable question: who actually controls your team’s AI decisions? There are three real answers, each with a different menu.

Row C looks like defeat and isn’t. A mandated Copilot is a constraint, not a verdict: the licensed tools are a free way to learn, and the models inside them are real. Used bare, though, they disappoint. In the first published blind-judged benchmark of its kind (April 2026, run by the vendor Ivo, so read it with that in mind), a frontier model out of the box scored 3.50 out of 10 on contract review. The purpose-built tool scored 4.52; a senior lawyer 4.56. Nobody aces this test, but the ordering is the point.¹⁷ The setup wasn’t even-handed: the purpose-built tool came configured for the task; the frontier model came bare. But the asymmetry is the finding. The gap was what the model was given to work with, not the model. Raw model access is not capability. The playbook you feed it is the capability.

One constraint sits across every row: the buffet. Instruct several firms and several platforms will come at you through door 5, each convenient, each moving a little more of your knowledge into someone else’s product. You cannot adopt them all, and deciding matter-by-matter under deadline pressure means the firms decide for you. Set the rule once, for every firm: which portals you will work in, whether your playbooks may go into a firm’s platform, and what firms must tell you about AI-produced work.

For two years legal AI was priced like software: per seat, per year, predictable. That era is ending. On 23 June 2026 Legora moved its flagship tier to consumption pricing, in its own words “from hours billed and seats licensed, to outcomes delivered.”¹⁸ Expect others to follow, because the vendors’ own costs run on consumption. Harvey’s founders said in May 2026 that their token throughput (tokens are the units AI work is metered in) had grown twelvefold in five months, and that the company is not currently profitable.¹⁹

Consumption pricing cuts both ways. In your favour: cost becomes legible per matter, which a seat licence never showed you. Against you: volatility. Agentic workflows, where the AI takes many steps on its own, multiply consumption in ways nobody can forecast, including the vendor.

Remember margin stacking too. The further from the source you buy, the more layers of markup the same computation passes through: model, then platform, then embedded product, each taking a cut. That is not an argument for always buying at the source; the markup buys real things (workflow, security, someone to sue). It is an argument for knowing how many layers you are paying for, and asking each what it adds.

Two moves follow. First, settle AI cost allocation in your engagement letters and panel terms now, while it is cheap to raise: is AI the firm’s overhead or a disbursement, and if a disbursement, at cost or marked up? One sentence today avoids an argument in 2027. Second, ask every vendor Rubric 2’s fourth question and get per-matter cost visibility written into the contract.

Legal tech products fail, get acquired, or get abandoned. ROSS Intelligence, a respected legal research startup, shut down in 2020 while defending copyright litigation that is still running; the appeal was argued on 11 June 2026.²⁰ Its customers were pointed at three competitors and left to manage the move. Even the safest names get repriced overnight: on 3 February 2026, after Anthropic shipped a legal plug-in, Thomson Reuters shares fell 16% in a day.²¹ If the market cannot price these products with confidence, neither can you.

You cannot pick survivors, but the risk isn’t uniform. Thin products that wrap a frontier model in a chat window are the most exposed. Deep workflow platforms are defensible but expensive to run.¹⁹ Products anchored to systems of record (your DMS, your CLM, the law reports) are the most insulated, because the data holds even where the AI is mediocre. Rate every product in your stack on that spectrum. Then do the one thing that makes tool death survivable anywhere on it: keep the knowledge outside the tool.

Before the needs, name the failure mode, because it’s the default. Nobody decides it; then it accretes.

The contracts team pilots the CLM vendor’s AI module. Disputes standardises on Copilot, because IT switched it on. One lawyer builds a brilliant personal workflow in a paid ChatGPT account. Two firms’ portal seats arrive with the next two big matters. Procurement renews the DMS with the AI uplift because the discount expired Friday, and procurement’s own contract tool starts suggesting fallback positions nobody in legal has seen.

Eighteen months later: six tools, four pricing models, no shared playbooks. Each product has been learning your team’s precedents and positions inside its own walls, in formats none of them will hand back. The CLM knows your contract positions, the portal knows your instructions, the chat histories know everything else. Three lawyers do the same task three different ways at three different costs. Everyone has become the champion of a different product, which is another way of saying the AI deployment is not managed at all. And when the board asks what legal’s AI position is, the honest answer is: we have six, and they’ve never met.

The cost is bigger than duplicated spend. Whatever makes your team valuable (the positions you take, the way you paper a deal, what you learned the hard way on the last dispute) is exactly what these products absorb as your lawyers use them. If you don’t own your knowledge management, you are giving it away: chopped into a thousand pieces and handed to every vendor, firm and business unit with a prompt box. Nobody would photocopy the precedent bank and post it to six suppliers. An unmanaged AI deployment does the equivalent every day, without anyone deciding to.

None of the individual decisions was wrong. The absence of a decision was wrong. That is what a team AI strategy is for, and it is why the strategy has to start from needs, not products.

Run the logic from the work, not the demo. Most in-house AI needs reduce to seven. Six map onto the doors cleanly. One maps to nothing on the market, and that one is the point. (The types below are the six doors from Part 1: source, legal platform, enterprise copilot, legaltech bolt-on, your firms’ AI, the business’s AI.)

Now read the fourth column. The tools supply almost none of it. The rows keep asking for what only you can bring: your positions, your precedents, your matter histories, a named owner. No product ships with any of it; the benchmark result in Part 2 is that gap, measured. The model is the engine. Your knowledge is the fuel. Everyone is selling engines.

The strategic core is unglamorous: get your filing system in order, and make it readable by the various tools. Keep the master copy of your team’s knowledge in plain, portable formats you own. Ordinary files, ordinary folders, readable by anything; in practice, Word documents and plain-text files in storage the team controls. Not inside Harvey’s vault, Legora’s playbooks, the CLM’s clause library, or Copilot’s memory. Those are working copies, rebuilt from the master whenever you like.

Three reasons, all downstream of Parts 1 and 2:

This is also the answer to the authority problem in Part 2. A row-C team, no budget and a mandated Copilot, can build the knowledge layer with nothing but discipline, and it converts into leverage the moment budget arrives. It is the only AI investment available at every authority level, including none.

If the chaos scenario is the default, the alternative must be light enough to use.

That’s the strategy. It contains no product name. The products are Part 1; they slot into a strategy shaped by Parts 2 and 3, never the reverse.

The teams that come out of this era ahead won’t be the ones that picked the winning vendor. Nobody reliably can: the vendors disagree about which layer survives¹⁴¹⁵¹⁶, and so do the big firms²⁴²⁵²⁶. The winners will be the ones who understood the products well enough to label them, faced their constraints, and knew their needs well enough to notice that the most important asset was never on any vendor’s price list.

You have more leverage than you think. The firms are already taking direction from clients like you²²²²³. The vendors want your logo. And the knowledge layer, the thing that makes every product on the map work at all, is already sitting in your team’s documents and heads, waiting to be structured by the one party with no incentive to lock you in: you.