Running the numbers
Law firms are by no means alone in running numbers on AI return on investment (ROI) and getting different answers depending on what they ask. Forrester, reporting in 2026 on a 2025 survey of more than 1,400 AI decision-makers, found 73 percent reporting a return below 50 percent on money spent to date, and only 2 percent above 75 percent. MIT’s NANDA project, in a preliminary study, put the share of enterprise pilots delivering no measurable P&L impact at about 95 percent. In Deloitte’s 2025 research, 84 percent of those investing in AI said they were gaining a return. PwC’s 2026 study of 1,217 executives found that about 20 percent of companies capture 74 percent of all the economic value AI creates.
Those figures rely substantially on what executives say they have seen or believe, though, rather than on formal testing of substantive performance metrics.
The legal profession’s numbers are thinner still. Thomson Reuters has found that only 18 percent of professional services organisations collect ROI metrics for AI at all, and that 82 percent of legal departments either do not measure it or do not know whether they do. Axiom’s 2026 survey of 528 in-house legal leaders found that every AI-using team planned to raise its budget, and that 17 percent tracked whether results were achieved.
This means that the debate about AI’s ROI in law firms is being conducted largely without proper data. Strategic decisions built on these numbers might be wrong.
The typical management committee response to such uncertainty is to slow spending and demand better proof of ROI, even if that proof is anecdotal or indirect – for instance resolving client pressure to act or lose the work. That response is of course sound. A firm that cannot say whether an investment returns more than it costs is not running a proper business, and the partners who insist on questioning ROI are doing their job.
AI’s ROI in a law firm vs elsewhere
How should ROI on AI be properly estimated in a law firm, though? As most firms apply it, the ROI metric was built to assess the economic value of capital expenditure. It assumes a fixed sum invested, a stable asset, and a return in the same currency over a known life. None of those conditions fit well with AI, which is constantly advancing, changes what the firm sells to clients and the pricing models by which it earns fees. And which, ultimately, transforms the firm’s whole business model into a new paradigm where different value assumptions and metrics apply.
A method intended to appraise a photocopier, a building or a mine cannot properly appraise the economic benefit of radical change in a professional partnership’s business model. Applied wrongly, it gives a precise, wrong answer to a question the firm did not even mean to ask.
Interrogating ROI more elaborately through discounted cash flow (DCF) methods that model the net present value (NPV) of the future profits to be derived from AI, based on what has been achieved generally in other contexts, typically fails in a similar way. In the context of AI and law firms today, the danger lies in applying correct arithmetic to bad assumptions – perhaps even poorly informed guesswork. Admittedly, good evidence does exist for the value that AI-native or AI-fluent operating models generate, but that sits in software, financial services and other sectors that were already heavily digital before generative AI arrived. Other evidence from capital-intensive industries where AI-fluency shows up in robotics and other features has little relevance to a knowledge-intensive partnership such as a law firm.
Transposing these numbers from other industries onto a leveraged professional partnership selling bespoke judgement requires assumptions about pricing, client procurement behaviour, regulatory constraints, verification costs, and much else besides, that do not typically translate well. DCF methods inherit all the uncertainties and ambiguities inherent in the approach, multiplied by translation errors, to return confidently precise but false conclusions that are no more reliable than the weakest, most flawed assumption in the model.
DCF also rewards near-term predictable savings and penalises distant uncertain gains. This is not a flaw of DCF but is by design. In this context, though, it therefore not only fails to value the benefit of business model transformation but actively argues against it. A firm that models AI’s ROI this way will approve a pilot project but decline the build-out and be able to show convincing business case workings for that being the “correct” decision. As the time horizon lengthens and the complexities increase, the investments that matter most often become the hardest to get approved.
The partnership governance model exacerbates this dissonance yet further. Partnerships hold little capital and distribute their profit each year, so no institutional discount rate really exists. The effective discount rate is set more by the age profile of the equity partners rather than the cost of capital. A partner two to five years from retirement (i.e. the most senior) applies different rules to the younger partners who know that investment in AI today drives the competitive advantage of their practices tomorrow.
In essence, estimating AI’s value in a law firm is best done by framing the investment as an option rather than an asset. A firm investing in AI is buying today the capability to operate a completely different kind of legal services business in the future. The ROI sits in what it learns, in the decisions it then makes about what to do with that, and in the outcomes of those decisions.
An evolutionary marathon, not a transformational sprint
Acquiring an AI-fluent business model is an evolutionary marathon, not a transformational sprint. The way that value is measured evolves along the way, too. Firms that take the time and trouble to cost this properly now are looking well beyond the costs of “software as a service” (SaaS) licences. The all-in cost of AI adoption also includes implementation, security review, training, forgone fee-earner hours, workflow build, ontological preparation, supervision, management attention, and the pilots that failed. Some of that outlay is one-off and some recurs as the tools develop and people churn.
Verification is also crucial to accurately assessing AI’s ROI in a law firm. This means the extra checking that must be done because the output is not yet trusted, which is a recurring charge that rises with volume. Stanford’s RegLab study found purpose-built legal research tools wrong (or mis-grounded) on between 17 and 33 percent of queries. An error rate below one percent on a single item or workflow can still become an institutional failure. Rare indeed are the firms that properly price the risk of verification failure, despite the lesson of so-called O-ring economics, named after the inexpensive component whose failure led to the loss of the space shuttle Challenger. This is where a positive business case can easily turn negative.
Verification costs fall as reliability improves and frontier capability is doubling on some measures every seven to eight months, now. A firm measuring that today is measuring a substantively different technology to the one it measured twelve or eighteen months ago, and each advance brings both new value opportunity and new risks with it.
Assessing AI’s ROI is further complicated by the firm adopting AI across its existing workflows, within its existing time-based, human-leveraged model, while shifting to new workflows that live inside the wholly different business model that AI-fluency entails, with the new model being incompatible with the existing one. In today’s model, revenue is a function of hours, rates and headcount. The firm captures value through input volume. Efficiency, whether AI-induced or otherwise, reduces that revenue.
Decoupling revenue and profit from headcount and billable hours
In a results-priced, AI-fluent model, revenue is decoupled from time and headcount. AI-induced efficiency tends to be accompanied by the ability to take on more complex, larger-scale work, delivering better results for clients at a lower price while generating more profit for the firm. These are not just two settings on a dial but radically opposing logics. Each one rewards what the other punishes. It is very hard indeed for both logics to co-exist in the same system, such as a law firm.
A common approach to trying to resolve such a paradox is to assume that everything can be worked out incrementally, perhaps by shifting incentives to influence behaviour. It probably can, but only up to a point. That point is reached when rivals move ahead in the shift from the human-leveraged, time-based model to the AI-leveraged, results-based one. At that point, clients shift business to them and the lagging firm is poorly equipped to win it back. ROI assumptions collapse. So too does economic performance, perhaps existentially.
The market appears already to have accounted for some of AI’s value-add under the first logic. Worked rates rose 7.3 percent across US law firms in 2025, against inflation near half that, and profits followed. The temptation that follows, logically, is to value AI’s ROI in terms of uplift in hourly rates. That gain is real, but transient. It relies on clients continuing to accept the billable hour as the unit of value, which assumes that AI-fluency will not deliver better options for them. Thomson Reuters indeed records that most of law firm profit growth still comes from rate rises rather than operational improvement, but it also shows that clients have begun moving work downstream as rates escalate.
Thomson Reuters’ 2026 Future of Professionals report finds 78 percent of corporate clients rating AI-enabled quality improvement as very important or essential, but only 6 percent saying most of their providers deliver it. Thirty-two percent of respondents are reconsidering relationships with firms falling behind. Once an hour delivers ten times what it used to, that hour stops describing value. Firms cannot simply multiply their rates by ten, because clients will not tolerate it. Value that would accrue to the firm under the second logic is therefore lost to it, and AI’s ROI shows up at perhaps a radically lower value than it should.
Deloitte’s polling of around 3,700 professionals puts rough numbers on the split between the two logics across industries generally. Forty-eight percent have introduced AI without redesigning the work it sits inside, and 12 percent have rebuilt the operating model around it. My own modelling, based on simple assumptions, suggests likely profit uplift in law firms to be about 15 percent for those employing the AI-assisted first logic, against 1.5× to nearly 5× the baseline, or more, for those properly employing the AI-enabled second logic.
Another item usually missing from the AI ROI ledger is the medium to long-term time horizon. Erik Brynjolfsson has estimated that the organisational reconfiguration required around a new technology can cost up to ten times the technology itself, and his productivity J-curve finds that measured productivity falls before it rises. This will resonate with the many law firm leaders currently viewing their AI returns with dismay. Firms that under-invest in that organisational reconfiguration see the dip, respond by cutting staff and other investment before the curve turns up, then forfeit the rebound.
Cognitive offloading, de-skilling and “never skilling”
The second item missing from this ledger is the cost of cognitive offloading and de-skilling (or, at the extreme, even “never-skilling”) among junior lawyers. Traditionally, juniors learned their judgement by doing the hard research, review and drafting work that AI now does. How, then, do today’s eager juniors become tomorrow’s skilled, wise seniors? Again, the challenge is by no means unique to law firms. BCG’s survey of C-suite leaders generally found around half already observing this talent-pipeline effect and a clear majority treating it as a three-to-five-year threat.
This particular risk is mitigable, though, and knowledge managers are already offering novel solutions. Some formative work can be allocated to juniors specifically for development purposes. Supervision can be redesigned around orchestration and verification of agentic machine output, which becomes critical in any event as firms move towards AI fluency. Aided by AI, juniors can take on more complex work, at larger scale, earlier in their careers.
All these issues affect the assumptions on which AI’s ROI must be based if it is to be a sound basis for strategic decision-making and budgeting. But the disappointing current reality is that where numbers exist at all, they have typically been built around short-termism and benchmarked against nothing sensible. Firms have established reasonable estimations of what AI returns inside the status quo. What it would return inside an AI-fluent business model remains unmeasured because most firms have no clue what that entails in practical terms. They hope that the answer will come in due course from vendors offering new SaaS tools. To the extent it does, the answers will fall short of what is needed to support the best decisions. Given the pace and scale of change, it is now imperative for law firms to properly learn what AI-fluency really means for them, and for AI’s ROI, and what they must do to respond strategically to both.
References
Axiom Law (2026) Legal AI is everywhere. Now comes the hard part: The 2026 Axiom In-House Legal AI Report. 29 June. Available at: https://www.axiomlaw.com/resources/articles/legal-ai-survey-report (Accessed: 25 July 2026).
Brynjolfsson, E., Rock, D. and Syverson, C. (2021) ‘The productivity J-curve: How intangibles complement general purpose technologies’, American Economic Journal: Macroeconomics, 13(1), pp. 333–372. doi: https://doi.org/10.1257/mac.20180386.
Challapally, A., Pease, C., Raskar, R. and Chari, P. (2025) The GenAI Divide: State of AI in Business 2025. MIT Project NANDA, July. Available at: https://nanda.media.mit.edu/ai_report_2025.pdf (Accessed: 25 July 2026).
Deloitte (2025) AI and tech investment ROI. Deloitte Insights. Available at: https://www.deloitte.com/us/en/insights/topics/digital-transformation/ai-tech-investment-roi.html (Accessed: 25 July 2026).
Deloitte (2026) Enterprise AI trends in 2026: AI transformation strategy opportunities and predictions. 2 June. Available at: https://www.deloitte.com/us/en/what-we-do/capabilities/applied-artificial-intelligence/blogs/pulse-check-series-latest-ai-developments/ai-transformation-predictions-2026.html (Accessed: 25 July 2026).
Forrester (2026) The Human Secrets to AI Success. Available at: https://www.forrester.com/technology/human-secret-to-ai-success/ (Accessed: 25 July 2026).
Goel, S., Martin, D. and Kaffe, C. (2026) ‘When Everyone Uses AI, Companies Risk Losing Critical Skills’, Boston Consulting Group, 17 June. Available at: https://www.bcg.com/publications/2026/when-everyone-uses-ai-companies-risk-critical-skills (Accessed: 25 July 2026).
Kremer, M. (1993) ‘The O-Ring Theory of Economic Development’, The Quarterly Journal of Economics, 108(3), pp. 551–575. doi: https://doi.org/10.2307/2118400.
Kwa, T., West, B., Becker, J. et al. (2025) Measuring AI ability to complete long tasks. METR, 19 March. Available at: https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/ (Accessed: 25 July 2026).
Magesh, V., Surani, F., Dahl, M., Suzgun, M., Manning, C.D. and Ho, D.E. (2024) Hallucination-free? Assessing the reliability of leading AI legal research tools. Stanford RegLab, May. Available at: https://reglab.stanford.edu/publications/hallucination-free-assessing-the-reliability-of-leading-ai-legal-research-tools/ (Accessed: 25 July 2026).
PwC (2026) AI Performance Study. 13 April. Available at: https://www.pwc.com/gx/en/news-room/press-releases/2026/pwc-2026-ai-performance-study.html (Accessed: 25 July 2026).
Thomson Reuters Institute (2026) 2026 AI in Professional Services Report. Available at: https://www.thomsonreuters.com/en/reports/2026-ai-in-professional-services-report (Accessed: 25 July 2026).
Thomson Reuters Institute (2026) Future of Professionals Report 2026: When the AI strategy doesn’t reach the desk: a profession under pressure. Available at: https://www.thomsonreuters.com/en/institute/future-of-professionals-2026/report (Accessed: 25 July 2026).
Thomson Reuters Institute and the Center on Ethics and the Legal Profession at Georgetown Law (2026) 2026 Report on the State of the US Legal Market: Peak prosperity and the fault lines below. 7 January. Available at: https://www.thomsonreuters.com/en-us/posts/legal/state-of-the-us-legal-market-2026/ (Accessed: 25 July 2026).
Waikar, S. (2021) ‘How Has COVID Affected the AI Economy?’, Stanford Institute for Human-Centered Artificial Intelligence, 3 March. Available at: https://hai.stanford.edu/news/how-has-covid-affected-ai-economy (Accessed: 25 July 2026).
