Legal departments rushing to implement generative AI are hitting a wall. Most business AI pilots fail to deliver measurable business value. That is not our opinion. That is according to an increasing number of studies that report dizzying AI pilot and project failure rates across industries. MIT cites a 95% failure rate, IDC says 88%, the RAND Corporation says 80%, to name a few. The numbers are not pretty and are not getting better.
But what about the 5% to 20% of pilots and projects that succeed? In legal, some organizations are achieving remarkable results, transforming how legal work gets done and defying these statistics.
While there is no shortage of advice on how to succeed with AI, in our experience the difference lies not with the technology but with the approach teams take to onboarding it. That should be obvious, as success always depends on a proven process. But it is not obvious, based on the reported failure rates.
We now have data to go with the experience. Axiom’s 2026 In-House Legal AI Report surveyed 528 in-house legal leaders across six countries. What the leaders getting AI right share is not budget. It is not headcount. It is discipline. Five practices. All reproducible.
Principle 1: Invest in Your People
Generative AI operates differently from traditional legal software. It is probabilistic rather than deterministic, creative rather than purely logical, capable of understanding natural language but also capable of hallucinations. This fundamental shift requires more than training lawyers how to use features. It demands broader education, including how to manage risk.
Attorneys need to understand what hallucinations are, why they occur, and why some AI tools are more prone to them than others. Without this foundational understanding, the first unexpected output can end their trust in legal AI permanently.
What the data shows: 40% of legal teams name the skills gap as a top barrier to AI value. 59% rate their teams as intermediate skill level. Only 10% rate their teams as advanced. And fewer than one in five team members use company-licensed AI tools daily. Nearly one in three rarely or never use them. Teams have the licenses. They do not have the adoption. — 2026 Axiom In-House Legal AI Report, N=431
Successful implementations combine broad AI education with role-specific training tied directly to real-world daily work. Generic training gets forgotten. Training connected to specific, regularly exercised workflows strengthens trust and creates a working AI culture that lasts.
The debate about structured training versus self-discovery misses the point. Both are necessary. Some attorneys thrive on experimentation, others need intensive support, others need a mix. The key is understanding the team’s culture and providing the appropriate resources aligned to each group.
Change management is not a one-time communications task. The teams that succeed treat it as an ongoing investment. They do not hand people a tool and a one-hour training session and expect them to fundamentally change the way they work. That is not how change happens.
The teams advancing fastest are investing in people as intentionally as they invest in technology.
Principle 2: Start Small
Start with a Strategic Use Case
Success begins with choosing the right problem to solve. Too many organizations put the cart before the horse. They start with technology and search for applications. The most successful teams identify their pain points first, then evaluate whether AI can help. We recommend viewing potential use cases through three lenses:
1. User desirability: What daily pain points do your attorneys face most frequently? What value would solving these problems deliver? The answers must come from the attorneys themselves, not from assumptions about their needs.
2. Commercial viability: Where does your organization see the largest volume of work? For many firms, commercial contracts represent the bulk of activity, making AI contract review and redlining natural starting points.
3. Technical feasibility: What can today’s technology reliably accomplish? What is the implementation complexity? Where can you achieve quick wins while building toward a larger goal?
This three-lens framework consistently points organizations toward high-volume, repetitive tasks as initial targets. Pre-execution document review, particularly redlining on counterparty paper, emerged as ideal at Axiom because it is performed daily, represents significant volume, and the technology has matured enough to deliver immediate and measurable value.
What the data shows: Perceived ROI concentrates in three document-heavy workflows: legal research (57%), contract drafting and review (19%), and document summarization (16%). The other use cases surveyed are cited by fewer than 4% of teams as top ROI drivers. Starting small is not a lack of ambition. It is the strategy. — 2026 Axiom In-House Legal AI Report, N=431
Structure Your Pilot for Success
Many legal AI implementations are doomed from day one because the pilot is inadvertently designed to fail. Most pilots in the 2026 Axiom survey ran three to six months. That is too long. A well-structured legal AI pilot does not need six months. It needs six to eight weeks, a defined use case, measurement baked in from day one, and a clear decision at the end.
Longer than that and you are not running a pilot. You are running an experiment with no endpoint. Attorneys stop engaging. Urgency fades. And when the pilot finally ends, nobody can tell you whether it worked.
What the data shows: Fewer than one in four legal teams run structured pilots with defined success criteria, timelines, and measurement. Six in ten set general goals with a flexible approach. One in six run informal trials with interested users. Without a definition of success, a pilot produces activity, not insights. —2026 Axiom In-House Legal AI Report, N=431
Four structural elements distinguish successful pilots from the rest:
1. Integration with real work. The most common failures occur when pilots use dummy or synthetic data. Lawyers who are already pressed for time will not engage in additional work that does not deliver real benefits. Recently completed deals offer ideal test cases because they carry real-world complexity with known outcomes, which lets teams evaluate AI output against what actually happened.
2. Intensive support infrastructure. Weekly office hours with product and process specialists are essential. Different attorneys have different technical aptitude and need varying levels of support. Problems must be identified early and solved quickly. Without early intervention, attorneys lose trust in AI. Once trust is lost, it rarely returns.
3. Structured feedback mechanisms. Twenty-minute interviews at pilot midpoint and endpoint, coupled with quantitative and qualitative surveys, reveal whether the AI technology works and how it fits into actual workflows. This provides the data needed for informed decisions about whether to proceed, pivot, or abandon.
4. Vendor partnership assessment. The pilot period reveals how legal AI vendors respond to feedback and adjust to needs. These are the factors that are often more important than comparative feature sets for long-term success.
Focus, Focus, Focus
When attorneys receive tools with a wide range of capabilities, they might use different features each week but never develop proficiency in any of them. When teams focus on one specific capability for four weeks, attorneys build repetitions, develop confidence, and become advocates. It takes 10 to 14 days just to get comfortable with the concept of using legal AI, and another three to six weeks for habit formation.
Starting with a narrow focus enables personal discovery. Once attorneys master core use cases and understand the technology’s parameters, they start to identify adjacent applications themselves. This serendipitous personal expansion often leads to discovery of the most valuable use cases.
Principle 3: Set the Rules First
Most legal teams are running AI the same way a marketing department would. ChatGPT, Copilot, and Gemini are the three most widely used AI tools in legal. None were built for legal work. And 68% of teams running those platforms use them in general-purpose configuration, not legal-specific. The legal-grade behavior that would justify a legal workflow—configuration, prompt engineering, workflow scaffolding, review protocols—lives outside the tool. None of it ships with the software.
What the data shows: 28% of legal teams say their chosen AI tools covered narrower use cases than expected. 35% would have done more rigorous evaluation before committing. Only 11% conducted rigorous evaluation with defined criteria and benchmarking before selecting their tools. General-purpose AI tools are growing more than twice as fast as legal-specific tools. — 2026 Axiom In-House Legal AI Report, N=431
Setting the rules first means several things. Before you select a tool, define what legal-grade performance looks like for your specific use case. Before you run a pilot, document your AI governance policies: data handling, confidentiality protocols, acceptable use, review requirements. And before you scale, audit what your AI tools are actually doing, not just which ones you bought. For help developing an AI governance policy, it is often a good idea to consult with an artificial intelligence lawyer.
Lastly, setting the rules first means knowing the rules that govern you. Before launching any AI pilot, review the bar rules, ethics opinions, and applicable regulations in your jurisdiction, including competence, confidentiality, supervision of non-lawyers, and fee-reasonableness obligations. Requirements vary by state and country and are evolving rapidly. Confirm your AI use program is compliant before you go live.
Most teams skip at least one of these steps. The data on AI disappointments reflects the consequence. Teams report that AI disappoints more on adoption than on performance. The technology is not the problem. The governance layer around it is.
The good news: most teams already believe in governance. Across 11 governance practices surveyed, fewer than one in 10 teams have no plans for peer review of AI work product, AI use policy documentation, or workflow integration. The intent is nearly universal. The gap is between planning and doing.
Principle 4: Know What Problem You Are Solving
83% of legal teams cannot reliably measure whether their AI spending is working. 17% have established metrics. The rest are spending on faith. “Establish better measurement practices from the start” was the most common answer to what teams would do differently if starting over, cited by 37% of respondents.
The teams that do measure split into two distinct groups. Those tracking formally count hard outcomes: labor cost reductions, time savings, reduction in outside counsel spend. Those tracking informally count squishy outcomes: user adoption rates and satisfaction scores. The measurement approach determines what teams believe AI is doing for them. ROI built on anecdotes is not ROI.
What the data shows: Among teams with formal metrics, the top reported outcomes are reducing error rates and reducing outside counsel spend. Among teams measuring informally, the top outcomes are user adoption and satisfaction scores. Both groups agree AI is helping. They disagree on how. — 2026 Axiom In-House Legal AI Report, N=431
Measure, measure, measure. Facts, facts, facts. Track efficiency gains and new capabilities. What questions can you now answer that you could not before? What risks can you now manage better? What projects can you now take on that you would have passed on? Document the expected benefits and the unexpected discoveries.
You cannot know what problem you are solving if you have not defined what solving it looks like. A pilot with no measurement baseline cannot produce a decision. It can only produce activity.
Principle 5: Leverage Your Power Users and Partners
Nobody is figuring this out alone. 98% of current AI users say outside guidance on AI tool selection would help. 77% of non-adopters say the same. The market is not short on tools or ambition. It is short on trusted advice about which tools are right for which work.
When we asked teams where they turn for guidance, alternative legal service providers (ALSPs) ranked first, ahead of management consulting firms (39%), industry associations (35%), law firms (33%), and AI vendors (31%). The teams doing this at scale have already figured out who actually knows how to do it.
What the data shows: 96% of current AI users say they would use an ALSP specifically for AI adoption and implementation. Legal teams prefer ALSPs over law firms for AI-enabled legal work by more than two to one: 52% to 24%. The work being outsourced most maps directly to where AI ROI is highest: high-volume contract review, legal research, and document summarization at scale. — 2026 Axiom In-House Legal AI Report, N=431
ALSPs do this work every day. The configuration, the workflow design, the training … none of it ships with the software. ALSPs have already built it. Law firms have not. 92% of respondents expect AI-driven savings to show up in their law firm bills. 58% say they do not.
Given where the industry is in its evolution, legal AI implementations often feel exploratory despite the best efforts to structure them. With hundreds of new vendors entering the market, making sense of all the options is challenging, even for those focused on it full time. Organizations that succeed leverage external expertise—peers who have already crossed the chasm, consultants who specialize in legal AI implementation, ALSPs that have embraced an AI-first mentality.
The teams getting AI right are not trying to solve this in a vacuum. They leverage their power users. They leverage their partners. They phone a friend. That is not a weakness. That is the strategy.
A Fundamental Rule for All Five Principles: Keep Humans in the Loop
Face AI’s Existential Threat Head On
With the five principles in place, one best practice applies to all of them: Keep people at the center. That begins with addressing a fear your team already feels. It is no secret that the buzz around AI is that it will put lawyers out of work. Your lawyers are reading and hearing these stories. Do not dance around the subject. Provide a thoughtful vision of the future for your team.
Yes, the efficiency gains from AI raise uncomfortable questions about billable hours, junior attorney development, and revenue models. But AI’s impact is more nuanced than simple displacement. General counsels report that AI enables them to answer business questions previously too expensive to investigate. Contract reviews skipped during due diligence for time constraints become feasible. Projects requiring months of manual effort get done in weeks.
Rather than simply doing the same work faster, AI-accelerated legal departments are expanding their capabilities. Risk assessments become more comprehensive. Deal analysis goes deeper. Back-office contract remediation that would have taken months now fits within deal timelines. These facts represent an accretive evolution, not a destructive one.
Some attorneys even report relief that AI handles the mundane work they dislike, freeing them up for more intellectually engaging tasks. The attorney concerned about losing billable hours may instead find they are handling more transactions, providing richer insights, and tackling previously impossible projects.
Use AI to Augment, Not Replace, Judgment
This is not a principle among the five. It is the rule that runs underneath all of them. Whatever process a team takes, however they invest in people, structure their pilots, set governance, measure outcomes, or find partners; attorney judgment stays in the loop. That does not change.
Attorney judgment—the ability to assess risk, navigate ambiguity, advise under uncertainty—is not replaceable by a language model. The configuration, the workflow design, the review protocols: all of it is human work that makes AI legal grade. The teams getting AI right use it to augment what their people do, not substitute for it.
Organizations frequently discover capabilities accidentally and come upon new applications while solving initial problems. This is precisely why human oversight matters at every stage. AI surfaces options. Attorneys decide. That division of labor is not a limitation of the technology. It is the right way to use it.
The strongest legal outcomes combine AI capabilities with experienced attorney judgment.
The Layer That Matters
What ties all five principles together is a single pattern. The top disappointments teams report are not about AI performance. They are about the work of adoption: implementation time, training burden, and ongoing maintenance. Tool selection ranks ninth among the factors successful adopters credit for their results. What ranks first is strong integration with existing workflows. What follows is well-defined use cases, investment in training and change management, and internal expertise. None of those is a tool decision. All of them sit around the tool.
That is the services layer: the configuration, the workflow design, the people scaffolding, the measurement infrastructure that makes the difference between a licensed tool and a realized outcome. The teams getting AI right have built it or found partners who already have it. The teams that have not built it are the ones spending on faith.
Here are three common pitfalls that emerge consistently:
1. The fourteen-day trial trap. Traditional software trial periods do not work for transformational technology. Attorneys will take 10 to 14 days just to become comfortable with the concept, then need another three to six weeks to build habits. Do not let your organization feel pressured to commit to an AI tool after only two weeks.
2. The build-versus-buy dilemma. Internal builds succeed about 33% of the time while purchased solutions succeed roughly 67% of the time. Most organizations underestimate the complexity of building an AI and the ongoing maintenance and support costs required.
3. The pilot philosophy misconception. Pilots should begin with high conviction that the technology will likely work, based on thorough, upfront due diligence. They are for confirmation and learning, not exploration. Post-pilot does not mean immediate enterprise-wide deployment, but rather a graduated rollout at a pace the legal team can absorb.
Disciplined Action Beats Perfect Planning
Putting all of this into practice comes down to a few disciplined moves. Start with the methodology, not the technology. Apply the three-lens framework to identify where AI can deliver immediate, measurable value in your specific context. Look at your highest-volume work and biggest pain points.
Give it time. Allocate six to eight weeks minimum for meaningful pilots. This provides enough time for attorneys to develop habits and for you to gather meaningful data about adoption and value.
Give it budget. If you are serious about reaping the value of legal AI, budget appropriately for training, support, and dedicated implementation resources. The technology is often a fraction of the investment. How you resource the pilot will determine success or failure.
Choose partners over products. High-quality, long-term relationships matter more than showy technology features. Assess vendors’ responsiveness, understanding of legal workflows, and commitment to your team’s success during the pilot period.
The teams getting AI right are not the fastest movers. They are the most disciplined ones. They started with a problem worth solving. They picked a use case small enough to prove. They configured their tools for the actual work. They measured from day one. They brought their people along. And when that worked, they did it again.
The issue is not whether to adopt legal AI. That is settled. The question is how to do it well. That lies in focused execution, structured support, and gradual expansion from proven foundations. That is how your legal department can join the minority achieving real measurable returns rather than the majority struggling with pilots that go nowhere.