Why AI pays off only when the work is redesigned

Most lawyers now use generative AI in some part of their work. In a LexisNexis survey of UK lawyers published in September 2025, 61% said they use it day to day, up from 46% in January that year [1]. Organisations are seeing much less from it. In PwC's January 2026 survey of 4,454 chief executives, 56% said AI had brought them neither higher revenue nor lower costs in the previous 12 months [2].

Our view is that AI saves individual lawyers time, and the organisation only sees that saving when the steps around the tool change too. We tested that view against the research from McKinsey, BCG, Deloitte, Bain and Thomson Reuters, including the findings that cut against it and the fact that several of these firms sell the redesign they recommend.

Why do lawyers save time while their organisations see little?

Individuals gain and organisations mostly do not. McKinsey's 2026 survey of 1,719 respondents shows the gap in two figures. 80% said AI had improved their own productivity. 37% said AI contributed anything to their organisation's earnings before interest and taxes (EBIT, the operating profit measure McKinsey uses), a share McKinsey describes as essentially unchanged from the previous year [3].

Both figures can be true at once. A lawyer who drafts a first version of a memo in twenty minutes instead of two hours has saved real time. If the memo then waits three days for a partner's review, goes back for changes and reaches the client on the day it would have reached them anyway, the client and the firm see the same result as before. The saving stayed with the individual.

What did the organisations that see value do differently?

They changed the work. McKinsey's March 2025 report tested 25 attributes of organisations against whether they saw EBIT impact from generative AI. Redesigning workflows had the biggest effect of the 25. At the time, only 21% of organisations using generative AI had fundamentally redesigned any of their workflows [4]. A workflow here means the sequence of steps, people and handovers that turns a request into a finished piece of work.

The 2026 survey shows the same pattern. About 6% of respondents qualify as high performers, meaning they attribute 5% or more of EBIT to AI. Nearly three-quarters of them had fundamentally redesigned workflows, against about one-quarter of everyone else. They were also twice as likely to have defined processes for measuring the impact of their AI initiatives [3].

Employees report the same thing. BCG's 2026 survey of 11,749 workers compared companies that redesign workflows with companies that only roll out AI tools. In the first group, 67% of respondents saw measurable business improvement from AI; in the second, 43% [5]. BCG also found that better tools on their own moved measurable impact by about 5 percentage points, while a clear strategy for how the work should change moved it by about 25 [6].

Exhibit 1

Workers in companies that redesign the work see results more often

Share of workers who saw measurable business improvement from AI, by company approach, 2026, %.

• Redesign workflows67%
• Roll out AI tools only43%

Survey of 11,749 workers in 14 markets. Source: BCG, AI at Work 2026 [5].

Most organisations have yet to make that change. In Deloitte's survey of 3,235 business and IT leaders, 30% were redesigning key processes around AI, 37% were using it at a surface level with little or no change to their processes, and 84% had redesigned no jobs around it [7].

What happens to a faster draft inside an unchanged process?

The time saved on one step is lost in the steps around it. Bain's 2025 technology report gives a useful comparison from software teams. Teams using AI coding assistants saw productivity gains of 10% to 15%. Writing and testing code, which is the part the assistant speeds up, accounts for only 25% to 35% of the time from idea to launch [8]. The rest of the time goes on specification, review, approvals and release, and those steps run at the same speed as before.

Legal work has the same structure. A contract review request in an in-house team usually passes through intake, a wait in a queue, a first review, questions back to the business, a second review, sign-off and a reply. If AI cuts the first review from two hours to thirty minutes, but the request still waits two days in the queue and another day for sign-off, the business sees almost no change in turnaround. The lawyer's time saving is real, but the turnaround the business sees depends mostly on the steps that did not change.

Lawyers already know this from managing matters. The client measures the matter by when the advice arrives, and the hours inside it are only part of that.

Does the legal evidence point the same way?

Legal-sector surveys point the same way, with one difference: they measure strategy rather than redesign.

Teams with a named AI strategy report better results. In Thomson Reuters' Future of Professionals survey of 1,816 professionals in law, tax, audit and related fields, 66% of those in organisations with a named AI strategy said AI met or exceeded their expectations for value. Where there was no active strategy, the figure was 22% [9]. Among law firm respondents, half said their firm's AI strategy is invisible in day-to-day work [10]. The survey does not show why, but a strategy that people cannot see in their daily work is unlikely to change how that work is done.

Exhibit 2

Professionals with a named AI strategy were far more likely to say AI met expectations

Share of professionals in law, tax, audit and related fields who said AI met or exceeded their expectations for value, by whether their organisation had an AI strategy, 2026, %.

• Named AI strategy66%
• No active strategy22%

Survey of 1,816 professionals. Source: Thomson Reuters Institute, Future of Professionals Report 2026 [9].

Few legal teams measure the result. In an ACC and Everlaw survey of 284 chief legal officers, general counsel and legal operations professionals, 12% of in-house teams tracked the return on their technology, and 9% measured its impact on business outcomes [11]. A team that has no record of how long a process took before AI was introduced has no way to show what changed afterwards.

Where does AI make legal work worse?

Redesign also has to decide where AI should be kept out and how its output is checked, because the research shows real failure rates in legal and expert work.

A Stanford-led evaluation of leading AI legal research tools, including Lexis+ AI and Westlaw's AI products, found that each produced hallucinated answers between 17% and 33% of the time [12]. In a preregistered experiment with 758 BCG consultants, AI improved speed and quality on tasks within its capability. On a task outside its capability, consultants using AI reached the correct answer about 19 percentage points less often than consultants working without it [13]. Ashurst's trials across 23 offices found large time savings on first drafts, with AI drafts scoring 3 out of 5 for accuracy against 3.5 for drafts by lawyers [14].

These results do not argue against using AI. They mean a redesigned process has to name the steps where a lawyer checks the output, and what the lawyer checks for, before the tool is used on live work. That is the check a supervising lawyer already gives a junior's first draft, written into the process.

What does redesign look like for a legal team?

Suppose a fictional in-house team at a mid-sized distribution company receives about 40 requests a month to review a non-disclosure agreement (NDA). The figures below are invented for the example.

1. Map the process as it runs today, with times. Requests arrive by email, often without the counterparty's name or the purpose of the disclosure. They wait about two days for a lawyer. The lawyer reads the whole NDA, emails the business to ask for the missing details, reads it again and replies. From request to reply takes an average of four working days and about 90 minutes of lawyer time.
2. Remove the steps that add nothing. A short request form that asks for the counterparty, the purpose and the term removes the email exchange about missing details.
3. Decide which steps a tool takes and which stay with a lawyer. A tool compares each NDA with the team's standard positions and lists every clause that departs from them. The lawyer reviews only the listed clauses and decides what to accept. Any NDA with more than a set number of departures goes to a full review by the lawyer.
4. Set the measures before starting. Turnaround from request to reply, lawyer time per request, the share of NDAs that go to full review, and errors found when a second lawyer checks a sample.
5. Run it on a small number of requests and compare. Ten requests through the new process, measured against the four-day, 90-minute baseline from step 1.

Two findings from the research apply directly. McKinsey found that tracking well-defined measures for generative AI was the practice most linked to EBIT impact, and that fewer than one in five organisations did it [4]. BCG found that 66% of frontline employees who use AI regularly get limited or no guidance on what to do with the time it saves them [5]. Step 4 deals with the first. Deciding in step 3 what the lawyer does with the time released deals with the second.

Comparing a clause with a standard position and marking up what departs from it is work lawyers do every day. The redesign writes that habit down so a tool can take the first pass.

Where the skeptics are right

Caution about building AI into legal work is often sound. A team that keeps a tool off live work until it knows where a lawyer checks the output is applying the same judgment it applies to any new process, and the failure rates above support it.

The evidence also has limits we want to state plainly. Most of the figures above come from surveys in which respondents describe both what their organisation did and the results it got. They show that redesign and reported results go together. They do not prove that redesign causes the results, because well-run organisations may simply do both. Each respondent also decides for themselves what counts as redesign.

Much of the research comes from firms that sell process redesign (McKinsey, BCG, Bain, Deloitte and PwC) or legal technology (Thomson Reuters and Everlaw). That does not make the data wrong, and readers should know it.

We found no published, independently measured before-and-after study of a legal team that redesigned a process with AI. The Ashurst trials measure single tasks, and most other legal examples are accounts given in interviews.

One finding counts against building your own tools. MIT's 2025 study of enterprise AI found that tools bought from specialist vendors and then customised reached deployment about twice as often as tools built internally [15]. The same report is the source of the widely repeated claim that 95% of AI pilots fail. That figure applies to custom enterprise tools, comes from 52 organisations interviewed and 153 leaders surveyed, and is described by its authors as preliminary [15]. It says little about AI in general.

On this course, the first version a participant builds is a quick way to test a redesign with fictional data. Whether the team then keeps it, buys a product or asks its IT team to build one is a separate decision, and a clear redesign makes that decision easier to take.

What we'd ask

If you lead a legal team or a firm:

• Choose one process that happens often, takes longer than it should and can be measured.
• Agree the measures and record the baseline before any tool is introduced, so you can show what changed afterwards.
• Decide what lawyers do with the time a tool releases.
• Make the AI strategy visible in day-to-day work, in the form of processes that have actually changed.

If you work in a legal team:

• Map the process as it runs today, with times, and remove the steps that add nothing.
• Decide which steps a tool takes and which stay with a lawyer, and name where a lawyer checks the output and what they check for before the tool is used on live work.
• Run the redesigned process on a small number of requests and measure the result against the baseline.

Why we built the course

The research keeps returning to the same point: the value comes from changing how the work is done, and someone in each team has to map it, redesign it and measure it. We built AI Women in Law so that more of the people doing that are women lawyers. Over five live Saturdays, participants take one problem from their own team, redesign the process and build a first version of a tool for it. See the course.

References

1. LexisNexis UK, Two-thirds of UK lawyers now use AI, yet firm culture slows progress, press release, 1 September 2025.
2. PwC, 29th Global CEO Survey: Leading through uncertainty in the age of AI, 19 January 2026. 4,454 CEOs in 95 countries and territories.
3. McKinsey & Company, The state of AI in 2026: On the road to ROI, 25 August 2026. 1,719 respondents in 97 countries, surveyed May to June 2026.
4. McKinsey & Company, The state of AI: How organizations are rewiring to capture value, 12 March 2025. 1,491 respondents, surveyed July 2024.
5. BCG, AI at Work 2026, 3 June 2026. 11,749 workers in 14 markets.
6. BCG, AI at Work: Why strategy matters more than tools, 2026.
7. Deloitte, State of AI in the Enterprise 2026: The Untapped Edge, 21 January 2026. 3,235 business and IT leaders in 24 countries.
8. Bain & Company, From Pilots to Payoff: Generative AI in Software Development (Technology Report 2025), 23 September 2025.
9. Thomson Reuters Institute, Future of Professionals Report 2026, June 2026. 1,816 professionals.
10. Thomson Reuters Institute, Future of Professionals 2026: Legal Report, 2026. 736 law firm and 203 corporate legal department responses.
11. ACC and Everlaw, survey of chief legal officers, press release, 17 November 2025. 284 respondents.
12. Magesh, Surani, Dahl, Suzgun, Manning and Ho, "Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools", Journal of Empirical Legal Studies 22(2), 2025.
13. Dell'Acqua et al., "Navigating the Jagged Technological Frontier", Organization Science, 2026. 758 BCG consultants.
14. Ashurst, Vox PopulAI: Lessons from a global law firm's exploration of generative AI, 10 June 2024. 411 participants in 23 offices.
15. Challapally, Pease, Raskar and Chari, The GenAI Divide: State of AI in Business 2025, MIT NANDA, July 2025. 52 organisations interviewed, 153 leaders surveyed. Copy read at mlq.ai.