.png)
In our latest AI research report, a survey of 528 in-house legal leaders across six countries, every team already using AI plans to increase its AI budget next year. Not one plans to cut. At the same time, 83% say they cannot measure whether that spend is working. Only 7% have actually been able to scale AI across their legal organization.
That’s a gap is not a tools problem, and it isn't a money problem. It's a discipline problem.
I recently hosted a webinar with three practitioners: Kaj Rozga, Senior Counsel at a Fortune 500 and host of the Version Up podcast, Richard Strachan, Assistant General Counsel - Global M&A, Supply Chain, AI & Data Privacy at Weatherford and David O'Hara, Corporate Counsel at Axiom. Three very different operating models, but they kept arriving at the same conclusions.
The trigger was two open roles, not a strategy deck
At Weatherford, two roles sat on the commercial legal team and in supply chain. Instead of backfilling both, Richard's team asked whether the right tool in the hands of the existing lawyers could absorb the full workload with one hire instead of two.
They chose a purpose-built contract review tool, and the honest version of what happened next matters more than the happy ending. The team's initial reaction was mixed. Tools look impressive in a demo, then you discover how long it takes to build playbooks your lawyers actually trust. Weatherford's contract work required bespoke playbooks, and the first two months were bumpy.
Six months in, the team is handling with one hire the volume that once required two. Richard says his US lawyers would not go back to the old way of working, but the headcount number is the one that settles the business case.
The lesson I keep coming back to is that legal teams routinely underestimate the configuration work these tools require. Enter a pilot without playbooks your team understands and trusts, and you will struggle to prove value and may abandon a tool that would have worked. We see this constantly with clients early in adoption.
AI delivers more value when the technology, legal expertise, and workflow are designed together.
Your operating model decides whether AI arrives top-down or bottom-up
Kaj offered a useful counterweight. He works in a decentralized operating model as a niche subject matter expert, and his organization's best AI use cases have bubbled up from the teams doing the work rather than arriving as a central mandate. Lawyers practice in particular ways, and their workflows are often bespoke. Local experimentation lets small teams think critically about their own needs.
His caveat matters, though. Bottom-up ideation still requires top-down support. You need budget, you need technologists, and you cannot hand a lawyer an agent builder and expect an agent. The enterprise funds the experimentation, then identifies which six-person team wins can scale.
David also sees the same pattern across Axiom's client base. The teams in that 7% treat AI as a business transformation rather than a series of tool purchases. They have executive sponsorship and designated champions across practice groups, and they've abandoned the idea that one tool will be everything for everybody. A tool that transforms one lawyer's practice may barely touch another's.
What you can measure, and what you probably can't
So how do the successful teams handle that 83% measurement problem? The honest answer from our panel is imperfectly and deliberately.
Richard's team tracks time saved on contract review. If a contract that took three hours now takes one, that's a two-hour savings. The harder question is what those two hours become: more proactive time with the business, faster cycle times, or more training. Weatherford is still working through how to translate time savings into dollars, and Richard said so plainly.
Kaj pushed further. He calls himself as a “Moneyball guy” who likes data, yet he pointed out that nobody demanded an ROI analysis when enterprises rolled out Word or Outlook. He's also skeptical of self-reported time savings: lawyers undercount how long work took before AI, overlook the time spent fixing AI output, and inflate their estimated savings. High volume, commoditized work can be measured with hard data. Higher-value advisory work is far more subjective, and there you are better off watching trend lines like outside counsel spend than mapping ROI to specific tools.
David added the category that CFOs struggle with most: intangibles that turn out to be real. Consistency of work product is one. A review performed fresh on Monday morning and one performed at 4 PM on Friday should catch the same high-risk issues, and with a well-configured tool, they do. For global teams operating across jurisdictions, templates, and languages, centralized playbooks create a level of quality control that is genuinely hard to price. So is retention. Lawyers freed from repetitive work do not want to give the tool back, and that shows up in who stays.
General-purpose tools set the baseline, then stop
Our research found that two-thirds of legal teams are running general-purpose tools in its default configuration. The three most widely used AI tools in legal—ChatGPT, Microsoft Copilot, and Google Gemini—were not built for legal work, and most teams have not configured them to behave as if they were.
The panel converged on a layered answer. Kaj is more bullish on general-purpose tools than most legal tech watchers: an assistant wired into an enterprise suite is convenient, connected to your systems and data, and a pretty good baseline for day-to-day work. His favorite use case is preparing for meetings with the business. Legal vertical tools write like a senior law firm associate, he noted, which isn't much help when you need a quick, bottom-line one-pager for senior leadership.
Richard agreed on the baseline and then explained where it breaks down. Everyone in his department uses a general-purpose assistant for general advisory work. But contract redlining, legal research grounded in a strong case dataset, and large-scale due diligence review each demanded purpose-built tools, because a general-purpose tool simply can't deliver the same accuracy or consistency for those workflows.
David has actually run this comparison for a client that was initially limited to a general-purpose assistant. The team found open text chat daunting and didn't know where to start, and months of Copilot access produced very little day-to-day use. When they piloted DraftPilot, adoption took off, because the tool arrived with playbooks, prompts, and workflows designed around how attorneys actually review documents. Somebody has to do that engineering. Purpose-built tools have done it on the front end.
As for how many tools is too many, Richard's test is simple. Each tool has to prove its value, and it has to integrate into existing workflows. A Word plugin your lawyers never have to leave is sustainable. Fifteen separate logins is sprawl.
What this means for outside counsel
One more finding from our research: 92% of in-house teams now expect AI-driven savings to show up in their law firm bills, and 58% say those savings have not arrived. Richard described how this reshapes the work. His team takes research and diligence to a strong internal draft, then pays outside counsel a premium for judgment rather than billing thirty associate hours from scratch. Kaj's view is that firms that passing savings through will keep the work and that firms that move slowly will watch clients build the capability in-house.
Literacy before tools, use cases before ROI
Asked what they would do differently if they could start over, 37% of leaders we surveyed said they would establish better measurement practices from the start, and 36% said they would build or bring in AI legal expertise before deploying. That is the playbook, as stated by the people who wish they had followed it.
Don't hand teams a tool with nothing but an expectation of efficiency. Build baseline AI literacy first, define structured use cases, and set realistic expectations. This is not a magic button. Think about it like the shift from library research to online research: it’s second nature now, but it took time to get there.
The teams in the 7% aren't smarter or better funded. They picked high-impact use cases, invested in configuration and playbooks before demanding ROI, matched tools to practices instead of forcing one tool on everyone, and measured what could honestly be measured. None of that ships with the software. Somebody has to map the work, rebuild the workflows around what the tools are actually good at, and carry the volume while the change lands. That is the work we do at Axiom. We are not selling you the license, so we have no reason to tell you your pilot went well.