By Rupali Patel Shah, Head of Legal Solutions, DiliTrust
The organizations getting real value from AI right now share one thing: they treated adoption as a transformation problem. The ones staring at bills they can’t explain and outputs they can’t fully trust treated it as a procurement event. That distinction — more than which platform you chose, more than how fast you moved — is where the AI story in legal actually lives.
I’ve been writing about this pattern for months, across information governance, regulatory uncertainty, and technology implementation. The formula is remarkably consistent: overinvest in the tool, underinvest in people and process, expect transformation, get disappointment. AI is not exempt from this formula. If anything, it is the highest-stakes version of it we have seen.
The bill is real. The value is harder to find.
AI budgets inside legal departments grew an average of 67% year on year, per Deloitte’s June 2026 report, The AI Imperative: Reshaping of the Legal Industry — a survey of more than 100 General Counsels and senior legal leaders across nine sectors. [1] Two years ago, 76% of legal departments reported no AI adoption at all. Today, 61% are in active deployment. [1] The spend is accelerating fast.
The returns are not keeping pace. An MIT study found that 95% of enterprise AI pilots delivered no measurable P&L impact. [2] Bain & Company tracked 200 companies over five months on GenAI specifically in legal — and lawyers showed the greatest dissatisfaction of any corporate work group. Only 53% said tools met or exceeded expectations, and that number fell 18% over the study period. [3] Deloitte is careful to note that the most significant value from AI in legal is “still to come.” [1]
That phrase is doing a lot of work. Because the bills are not waiting.
The cost problem has a name: Token maxxing
Here is what happens when you deploy AI tools broadly without a clear outcome in mind. Adoption metrics get built around usage — who is using the tools, how often. Reasonable intent. Unintended consequence: employees quickly learn that high AI usage signals relevance in an environment already anxious about headcount. Consuming more tokens becomes the goal rather than the outcome. As Cornell University’s AI Innovation Hub has observed, “employees learn to maximize usage rather than value.” [4]
One large technology company burned its entire 2026 AI budget in four months after deploying a tool to roughly 5,000 engineers. Its CTO spent $1,200 in a single two-hour demo session. [5] Per-token AI prices have fallen 80–90% over the past two years. Enterprise AI bills went up 320% in the same period [6] — because agentic workflows can generate 10 to 20 separate LLM calls per task [7], and only 15% of companies can forecast their AI costs within 10% accuracy. [8]
The current pricing is also partially subsidized by venture capital. When that changes, the bills become significantly harder to absorb. Goldman Sachs projects AI agents will drive a 24-fold increase in token consumption by 2030. [7] We genuinely do not know how expensive this gets.
You paid a premium. You may not own what you built.
A particular category of AI tool has compounded this in legal: products built as workflow harnesses sitting on top of foundation models — adding document management, matter history, sector-specific prompting — sold at a meaningful markup. Firms signed multi-year agreements before understanding the pricing architecture. Many discovered they were paying a premium subscription plus consumption-based API fees plus integration costs plus training, all on top of a foundation model they could have accessed directly at a fraction of the price.
Stanford Law’s analysis of AI vendor contracts found that 92% claim data usage rights beyond what is necessary for service delivery. Only 17% explicitly commit to complying with applicable laws — compared to 36% in standard SaaS agreements. [9] And for legal specifically, tools that integrate with matter management accrue institutional context — precedent libraries, lawyer corrections, risk thresholds built over months of use. When a firm leaves the platform, that learning evaporates. We still have no clear picture of the long-term privilege implications.
This is the lock-in that doesn’t appear in the pitch deck.
Some of this is about the users, not the tools
Here is the uncomfortable part: not all of this is the technology’s fault.
The most rigorous study to date on AI and experienced developer productivity — a randomized controlled trial by METR, published in 2025 — found that AI tools made developers 19% slower on real production tasks. The developers predicted they would be 24% faster. After completing the study, they still believed they had been 20% faster. [10] The gap between how productive people feel when using AI and how productive they actually are is one of the more consequential findings in the current research.
What I called “vibe coding” in my April piece applies well beyond software. It is the pattern of generating outputs quickly without deeply understanding them. For an individual working on a bounded problem, this can be fast and genuinely useful. For outputs that enter a workflow, a review chain, or a client deliverable, it creates problems that are expensive to find and expensive to fix.
Harvard Business Review named the output “workslop” — AI content that appears complete and professional while lacking substantive value. [11] In a survey of more than 1,000 U.S. workers, 40% reported receiving it in the previous month. Each incident takes an average of nearly two hours to resolve. The sender saves 30 minutes. The organization loses two hours. The knowledge base quietly fills with unverified content that won’t be traceable later. This is the information governance problem I have been writing about since January, wearing a different hat.
In legal, the consequences are professional. U.S. courts recorded 487 instances of AI errors or hallucinations in court documents in 2025 alone — more than ten times the 2024 total. [12] Deloitte found that only 24% of legal departments have any formal quality assurance framework for AI. [1] The tools are scaling faster than the checks on the tools.
84% haven’t redesigned anything
Deloitte found that 84% of organizations have not redesigned roles around AI. Only 17% have incorporated adoption incentives into performance frameworks. [1] Most are running a two-job model: employees generate AI output and then validate it extensively, without any structural change to how those roles are defined or supported. The savings promised by the technology are being partially consumed by the labor required to verify it.
As AI absorbs routine legal work, what remains becomes more complex — potentially requiring more senior professionals to backstop it, not fewer. That is not in most ROI models.
Deliberate beats fast
Deloitte’s case studies show a €6 million annual reduction in external legal spend at one insurance company and 5–8% productivity gains at a financial services department in early deployment. [1] The MIT analysis found that back-office AI deployments with clear process boundaries delivered $2–10 million in annual savings from document review automation alone. [2] What sets those results apart is not which tool was chosen.
Deloitte frames it plainly: “The most important investment decision for legal leaders is not which platform to procure. It is ensuring the full balanced portfolio of investment across technology, people, data, and transformation.” [1]
Define the problem before you fund the solution. Set milestones, not missions. Build accountability in from day one. I wrote about this in May in the context of technology implementation broadly — the principles are the same. AI is not a special exception to how change actually works.
The bill for AI adoption in legal is already large and growing in ways we cannot fully predict. The question worth sitting with right now is not whether your organization is using AI. It is whether you treated adoption as a transformation problem — or a procurement event. Because one of those paths leads somewhere. The other leads to a very expensive mess that is harder to unwind than it was to create.
More isn’t always more. Sometimes it’s just more to clean up.
Sources
[1] Deloitte Legal, The AI Imperative: Reshaping of the Legal Industry, June 2026
[2] MIT / Legal.io, MIT Report Finds 95% of AI Pilots Fail to Deliver ROI, 2025
[3] Bain & Company / Legal Dive, Gen AI and Attorneys Aren’t (Yet) Working Well Together, 2024
[4] Cornell Chronicle, AI’s Hidden Cost Problem: Companies Need Literacy, Not Token-Maxxing
[5] Forbes / Janakiram MSV, Uber Burns Its 2026 AI Budget in Four Months on Claude Code, May 2026
[6] Oplexa, AI Inference Cost Crisis 2026
[7] Goldman Sachs, Decoding the Agentic Economy, May 2026
[8] Campbell Robertson / Substack, Tokenmaxxing: The Hidden Cost That Eats the Savings You Were Promised, March 2026
[9] Stanford CodeX / TermScout, Navigating AI Vendor Contracts and the Future of Law, March 2025
[10] METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, July 2025
[11] Harvard Business Review, AI-Generated ‘Workslop’ Is Destroying Productivity, September 2025
[12] LawSites / Damien Charlotin, Legal Industry Reaches AI Tipping Point, March 2026
From AI Risk to AI Readiness: The Blueprint for Your Legal Department
What you’ll find in this blueprint:
- From AI adoption to real AI value: what makes the difference
- The 3 problems no software can solve on its own
- A 90-day plan for AI implementation any legal team can act on
Frequently Asked Questions About the Cost of AI Adoption in Legal Departments
Stop measuring adoption by consumption and start measuring it by outcome on a defined use case. When usage volume becomes the KPI, employees who are anxious about headcount learn that burning tokens signals relevance, so consumption inflates independent of value. This matters because agentic workflows can fire 10 to 20 separate LLM calls per task and only 15% of companies can forecast their AI costs within 10% accuracy, so a usage-based scoreboard sets you up for a budget blowout like the technology firm that exhausted its entire 2026 AI budget in four months. Tie incentives to a bounded process (document review, matter triage) with a milestone and a verified quality bar, then scale only after that use case earns it. A platform that logs cost and output per matter, rather than raw token counts, is what makes this measurable.
Negotiate explicit portability of your institutional context (precedent libraries, corrections, risk thresholds) and data-usage limits before signing, because that accrued learning is the real switching cost and it usually is not covered in the standard agreement. Stanford Law found that 92% of AI vendor contracts claim data usage rights beyond what service delivery requires and only 17% commit to complying with applicable laws, so the default terms favor the vendor. For legal specifically, tools built as workflow harnesses on top of a foundation model can charge a premium subscription plus consumption API fees plus integration and training on top of a model you could access more cheaply, and when you leave, months of accrued matter context evaporate with unresolved privilege implications. Insist on export rights, defined data-handling obligations, and pricing transparency across all layers so exit is a decision, not a trap.
You need a formal human-verification checkpoint at every point where AI output leaves an individual and enters a review chain, deliverable, or filing, because unverified output creates liability that is expensive to find later. U.S. courts recorded 487 instances of AI errors or hallucinations in court documents in 2025, more than ten times the 2024 total, yet Deloitte found only 24% of legal departments have any formal QA framework for AI. The gap is professional, not just operational: a hallucinated citation is a sanctions risk and a reputational one. Build defined review responsibilities, traceability of what was AI-generated, and a verification step into the workflow itself rather than relying on individual diligence, and use tooling that keeps a clear audit trail of AI-assisted work product.
They fail when organizations bolt AI onto existing roles without redesigning them, so employees run a two-job model (generate output, then extensively validate it) and the promised savings get consumed by verification labor. Deloitte found 84% of organizations have not redesigned roles around AI and only 17% have built adoption into performance frameworks. Compounding this, as AI absorbs routine work, what remains is more complex and may require more senior professionals to backstop it, driving up the seniority of judgment needed downstream, a cost that sits in no standard ROI model. The organizations that actually capture value (one insurer cut external legal spend by €6 million annually) treated adoption as a transformation problem by redesigning process, data, and roles together, not as a procurement event.
