Why Law Firm AI Pilots Stall Before Production, and How to Clear the Path
Key Takeaways
- AI has cut the time it takes to build a working prototype, so development is rarely the slow part of a law firm AI project any more.
- The time now sits in 3 queues: infosec review, architecture review and deployment.
- Each queue gets shorter when the firm agrees a reusable pattern once, instead of reviewing every project from scratch.
- Firms that agree the review path before the pilot starts get from AI pilot to production sooner.

A CTO at one of the firms we work with asked us a fair question recently. Everyone is using AI to build now, so why does it still take about the same time to get anything into production?
Our answer was short. AI made the build faster. It didn't touch the parts that were already slow.
A law firm AI pilot to production journey has 4 stages: build, infosec review, architecture review and deployment. A couple of years ago the build was often the longest of them. Now a working prototype of a redaction workflow or a matter chronology tool can be in front of a test group within weeks.
The other 3 stages haven't moved. So the gap between "it works in the demo" and "fee earners use it every day" is where the time goes.
Why the AI Pilot to Production Gap is Widening at Law Firms
Faster builds mean more pilots. More pilots mean longer queues for the same small group of reviewers.
At most firms, infosec, architecture and change are handled by a handful of people who also run everything else: the DMS upgrade, the cyber insurance renewal, the next audit. Every AI pilot lands on their desks as a new request.
And each one looks new, because it usually is. A different model, a different vendor, a different data source. So each one gets a fresh questionnaire, a fresh design review and a fresh slot in the change calendar.
The result is a pile of finished pilots waiting for sign-off. The business sees a tool that works and asks why it isn't live. IT sees a risk it hasn't assessed yet. Both are right.
The SRA's Risk Outlook report on AI is clear that firms stay accountable for the services they deliver, whether or not AI is involved, and that they need to supervise the systems they use. So the reviews stay. What a firm can change is how long they take, and how often they start from zero.
Infosec Review: The First Queue
In our experience, infosec review is where an AI pilot waits longest. The questions are reasonable ones. Where does the data go? Who can see it? Does the vendor train models on it? What gets logged, and for how long? How do we switch it off?
The delay comes from answering them again for every tool, usually through a long vendor questionnaire that goes back and forth for weeks.
What helps:
- Fix where the data lives. If AI tooling runs inside the firm's own tenant, the hardest questions get answered once. Client data stays in the firm's environment, existing access controls apply, and logs sit where the security team already looks. It's why we build AI legal tooling inside the firm's tenant.
- Write a standard AI security baseline. One document that sets out what any AI tool has to meet: data residency, no training on firm data, retention and deletion, permissions inherited from the firm's identity systems, and audit logging. Tools that meet it get a short review. Tools that don't go through the full one.
- Map it to a recognised framework. The NCSC's guidelines for secure AI system development split the work into 4 areas: secure design, secure development, secure deployment, and secure operation and maintenance. A baseline mapped to those areas gives the security team a structure it can stand behind with the board and with insurers.
Architecture Review: When Every Project is Treated as New
Architecture review asks whether the tool fits the firm's estate. Does it connect to the DMS, the practice management system and the identity provider? Does it create another copy of client data that someone will have to govern?
It slows down when there's no agreed reference pattern for AI. Without one, the architect designs from first principles each time, and the safest answer is often "not yet".
What helps:
- Agree one reference architecture for AI tools. Where models are hosted, how they reach firm data, how permissions are applied at the point of retrieval, and where prompts and outputs are stored. A pilot that follows the pattern needs a short conformance check.
- Keep the knowledge layer separate from the tool. If structured data, permissions and prompt libraries live in a layer the firm owns, each new tool is another interface on the same foundations. We covered this in our piece on legal AI interoperability.
- Name the exceptions upfront. Some pilots will need something the pattern doesn't cover, like a new data source or an externally hosted model. Flag those on day one so the full review runs in parallel with the build, instead of starting after it.
Deployment and Change: Getting it in Front of Fee Earners
A tool that clears both reviews still has to reach fee earners. That means packaging, testing, a slot in the change calendar, training, support and, at many firms, a partner willing to go first.
This is where a lot of pilots quietly stall. The pilot group liked the tool, but a firm-wide rollout needs a sponsor, a support model and a clear answer to "who do I call when it gets something wrong?"
Fee earners also need to know when to trust the output and when to check it. A 20-minute session built around their own matters will do more than a general AI awareness course. And supervising partners need to know what the tool does, so they can sign off work that used it.
What helps:
- Decide the production owners before the pilot. One person in the business who owns adoption, and one in IT who owns support. Our piece on who should own AI at a law firm sets out the options.
- Agree what good looks like in advance. Pick 2 or 3 measures the pilot has to hit, such as time saved per matter or accuracy against a test set your lawyers have signed off. If it hits them, it goes live. If it doesn't, it stops, and everyone knows why.
- Build the pilot where production will run. Use the firm's tenant and real permissions from day one. Then deployment means widening access to something that already runs in the right place.
Agreeing the Path Before the Pilot Starts
Most of the time in a law firm AI pilot to production journey is spent waiting for decisions that could have been made before anyone wrote a line of code.
So make them first. Before the next pilot starts, get 4 things agreed in writing:
- The security baseline the tool has to meet, and who signs it off.
- The reference architecture it has to follow, and what counts as an exception.
- The success measures that decide whether it goes live.
- The production owners in the business and in IT.
With those in place, infosec and architecture review become checks against something everyone has already agreed. Deployment goes into the change calendar on a date set at the start.
Keep the documents short and keep them live. A baseline that runs to 40 pages won't get read by the people building pilots, and one that hasn't been touched since last year won't reflect the models or vendors the firm uses now. A quarterly review by the same small group (security, architecture, a business sponsor) is usually enough. Each pilot that goes through the path also tells you where it's still slow, so the second and third pilots should move faster than the first.
If those 4 things aren't in place yet, that's what our AI Readiness in Five Days assessment covers. It looks at infrastructure, process and governance, and ends with a costed 90-day plan. We explain what those 5 days should tell you in our AI readiness assessment guide. For firms that already know what they want to build, our AI services team builds it inside your tenant, against the pattern you've agreed.
AI made building faster. The firms that reach production first will be the ones that sorted out the review path while their pilots were still being built.