Second in a series on funding AI work inside law firms. Part one argued that the process has to come before the fund.
If an innovation fund releases money against milestones, someone has to decide what counts as an AI innovation milestone.
This is where funnels usually fall apart. It isn't because firms can't measure things; firms measure everything. It's because the measurements that matter early are unrecognizable as legal work or deemed irrelevant. These signals often look like marketing. They look small. And a partnership trained to evaluate opportunities against the revenue of a real matter will throw them away on sight, if not careful.
For example, a client agreeing to pay $15,000 toward developing something is a rounding error next to a single decent engagement. It is also one of the strongest signals you can get that a new service offering has real potential.
Start with what you can afford to lose
Before the first milestone, you must establish the business case and the affordable loss.
The business case names the problem. That means where work is slowing down, where margin is leaking, or where the client feels friction the firm hasn't addressed. The affordable loss answers a different question: what is the firm willing to spend to find out whether this is solvable, assuming it learns nothing else?
That number sets the size of everything downstream. It's also the honest version of a conversation firms usually have dishonestly, by building an ROI model that projects returns nobody believes. You don't know the return yet. You do know what you can afford to put at risk. Start there.
This is not a soft way to decide. Deciding on the basis of affordable loss rather than expected return is characteristic of expert entrepreneurs. Novices build the forecast. Experts cap the downside and buy the option.
The ladder of AI innovation milestones
What follows is roughly ordered by cost and by strength of signal. Each rung is cheap relative to the one above it, and each one can kill the project before you spend the money on the next.
Do clients recognize the problem at all?
The cheapest tests have nothing to do with technology. Conversations with clients about the problem itself. Traffic to a page that describes it. Attendance at a webinar on the theme. Sign-ups to a mailing list about it.
These feel soft, and expect partners to say so. But they're testing something specific and important: whether the problem you've identified is a problem your clients would name themselves. If forty people register for a webinar on a friction point you thought was invisible, you've learned something. If six do, and three of them work at your firm, you've learned something too, for the cost of a webinar rather than the cost of a build.
Does a partial version change anything?
Then comes a minimum viable version. Deliberately not the finished thing.
This is where the law profession's instincts work against it hardest. Lawyers are trained on completeness; a partial answer is a wrong answer. But an AI tool that handles sixty percent of a task and hands the rest back to a human is not a failed tool. It's a first result. The question at this rung isn't "is this good enough to ship?" It's "does a rough version of this change how the work actually goes?"
Plenty of initiatives die here, and that's the rung doing its job. The data turns out to be worse than expected. The workflow isn't ready. The tool that demos beautifully performs poorly on real files.
Will a client use it on real work?
A pilot with a client, on live matters, with the partial solution. Not a demo. Not a sandbox.
This is the first rung that costs real partner attention, which is why the rungs below it exist.
Will a client pay anything for it?
Any amount. This is the signal firms are most likely to discard and least likely to be able to replace.
A client who pays $15,000 has done something categorically different from a client who says the idea sounds interesting. Enthusiasm is free. Budget is not. Someone inside that organization had to justify the line item to someone else, which means the problem was real enough to defend in a meeting you weren't in.
Compare it to a matter and it looks like nothing. Compare it to every other form of evidence available at that stage and it's probably the best you'll get.
Will a client put their own people on it?
Stronger still: a client willing to co-develop. Their time, their subject-matter experts, their process knowledge.
For most clients, people's time is scarcer than money. A client who commits it has decided the outcome matters to them, not just that your firm is pleasant to work with. That's also the point at which the thing you're building stops being a guess about what clients want.
Define what the client is getting
The last two rungs raise an important question: what does the client actually get?
Most of what firms build here is a better way of delivering a service, not a product, so this is rarely a question of ownership. It's a question of terms. The client might get preferential access or pricing once the service is live, early influence over how it works, recognition as a co-developer, or simply a better service sooner.
Be clear about what you are offering. First, define what the client is buying, so their $15,000 doesn't quietly turn into an expectation of exclusivity. Second, define what the firm may do with what it learns on their matters, including how their data is handled and whether the approach can be offered to other clients.
Why early evidence gets dismissed
There are two reasons early evidence often isn't taken as seriously as it should be.
The first is the comparison problem. Every proof point on the ladder gets measured against billable revenue, because that's the yardstick the firm has. Against that yardstick everything early looks trivial, so the honest response is to change what's being measured. These milestones aren't measuring profit. They're measuring how much uncertainty has been removed and whether the next tranche is justified.
The second is that the lower rungs look like marketing rather than practicing law. Whether clients care about a problem is a market question, and it has a market answer. Getting that answer for the price of a landing page is a bargain compared to finding out after the build.
Decide the threshold before you start
Write down what result would justify the next tranche before you run the test. Not "we'll see how the webinar goes," but "if fewer than twenty-five people from target clients register, we stop." The number can be a judgment call. What matters is that it exists before anyone is emotionally invested in the answer.
Without that, every result becomes an argument. Eighteen registrations is disappointing if you said twenty-five, and encouraging if you said nothing. The people who ran the pilot will always be able to construct a reading in which things are going well, and they won't be lying. They'll just be human.
One clarification for anyone whose firm has lived through a classic stage-gate process. The threshold is not a standard the project must meet to prove its worth. It's a device to stop biased reasoning after the fact. If thresholds start being used to screen out anything that looks unfamiliar, you've rebuilt the gate that kills novel ideas.
What you're actually buying
None of this proves an experiment worked. That isn't the point. The point is that at each gate the firm knows something it didn't know before, and that knowledge changes where the next dollar and the next round of partner attention go.
A pilot that ends with "the data was too messy" is not wasted money. It becomes wasted money only if the firm starts the next initiative without fixing the data first.
Start small enough to learn. Use real work. Involve the people who do it. Decide the threshold before enthusiasm takes over. Then read the result honestly. If the evidence is weak, stop. If the technology is promising but the workflow is broken, fix the workflow. If a client is willing to pay something, take that seriously, however small the number.
Next in the series: Where will the money come from?
If you're already thinking about an innovation fund, I work with firms on the pieces that make one fundable: designing the process that produces evidence, setting milestones a partnership will accept, and getting the incentives right so partners will back work outside their own practice.
That includes the unglamorous half — knowing when a project should stop, and stopping it early enough that the money still bought something.
Book a Discoveries partner call →
Rise Above AI Chaos: A Business Fable About Leading Organizations Through the AI Revolution