Author: Mark Buck

Last year KPMG International put an awkward proposition to its auditor. Grant Thornton UK was now using AI across its audit work, KPMG argued, so some of the saving should come back to the client, and KPMG was prepared to look for another auditor if it did not.

Reports in February, drawing on Companies House filings, showed that the fee for auditing KPMG International's 2025 accounts fell by about 14%, from $416,000 to $357,000. KPMG also argued that its accounts were straightforward, so not all of that reduction can be put down to AI.

The sum is small - but look at who asked. A Big Four firm that sells AI-driven efficiency to its own clients took a share of its supplier's AI saving at the first opportunity.

When AI lets a firm do the same work with fewer people, the saving can appear first in operating profit. That makes it tempting to capitalise the improvement at the firm's existing multiple, as though it were as durable as the rest of its earnings. The saving may not stay there.

Clients, employees, AI suppliers and, more slowly, governments all have ways of claiming part of it, and capitalising the whole gain assumes that none of them succeeds.

Three questions

When management credits AI with a better margin, an analyst has three questions to answer.

• Has AI raised productivity, meaning the same output from fewer labour hours or more output from the same?
• Has that productivity become profit, or has it gone straight to customers through lower prices?
• And how much of any profit will shareholders still hold once everyone else with a claim has made it?

Answering the first does not answer the other two. A firm that needs 80 people instead of 100 for the same work has made a real productivity gain. Whether its owners keep the proceeds is a separate matter.

The claimants are already queuing

Clients are moving first, and professional services is where they are moving fastest. The FT reported in September that Goldman Sachs has been asking its law firms to quantify their AI savings and share part of them, and using the answers in fee negotiations. Morgan Stanley and Citi have been moving towards competitive bidding and alternative fee arrangements.

Yet the savings are hard to find. Bloomberg Law's Roy Strom has pointed out that AI has not yet reduced the hours clients buy from large law firms; demand for those hours rose 4.2% in the first half of 2026.

That means clients are asking for their share before the savings have clearly appeared, and they are unlikely to ask less firmly once the savings show up in the numbers.

Staff come next, through the pay round. People who can make AI tools produce reliable work are scarce, and they will expect to be paid accordingly. Then there is the AI supplier. A vendor whose tools a firm's processes now depend on is in a strong position at renewal, and today's introductory pricing is a poor guide to what the software will cost once switching has become expensive.

The state moves slowest. In July 2026 the Office for Budget Responsibility published an illustrative scenario in which AI shifts income from wages to profits. The wage share of GDP falls from 40% in 2030–31 to 20% by 2075–76. GDP does not shrink, but tax receipts fall from about 38% of GDP to under 35%, because the OBR puts the average effective tax rate on wages at 46% and on profits at 27%.

The US Congressional Budget Office has described the same risk, and Yale's Budget Lab estimates that under a rapid AI-growth scenario, the extra US federal revenue in 2030 would be roughly half what it would be had capital and labour shares stayed at their 2026 levels, because capital income is taxed more lightly.

All three are scenarios rather than predictions. But a government that watches its largest tax base shrink while profits rise has obvious ways to respond, and investors should not assume the response will favour shareholders.

That response is unlikely to show up in a three-year earnings forecast. Its effect falls on the long-run margin and growth assumptions that feed terminal value.

Professional services as the test case

Labour-intensive firms with little physical capital are where this should appear first. Their largest cost is people, and a saving on salaries passes almost directly into operating margin.

The early numbers point that way without proving it. PwC's UK chief cut graduate hiring from 1,500 to 1,300 places, citing AI alongside a weaker economy, which he described as the bigger factor. A third of the UK's top 100 law firms told PwC they expect to automate at least 16% of current chargeable work, much of which is now done by junior staff.

The Law Society's 2026 benchmarking survey found partner profitability up 10.5% in 2025. Against that, PwC's own Youth Employment Index found no statistically significant link between AI adoption and UK youth unemployment, with early signs confined to the most exposed sectors.

The evidence is not yet strong enough to call a trend. It is strong enough to tell investors what they should start measuring.

What the assumption is worth

Take an illustrative professional-services business with £100m of revenue, staff costs of £60m and an operating margin of 15%. AI allows it to reduce headcount by a fifth, saving £12m a year. After £2m of new AI spending, operating profit rises by £10m and the margin reaches 25%.

Capitalise that £10m at 12 times operating profit and the firm is worth £120m more. Now assume that clients, staff and suppliers take it back in a straight line over five years, so the extra profit runs at £10m, £8m, £6m, £4m and £2m before disappearing. The straight line is chosen for simplicity and is not an estimate of how fast such margins actually erode.

Discounted at 8%, with each year's profit received at the year end, that stream is worth about £25m. Roughly four-fifths of the re-rating depends on an assumption nobody has tested: that the saving stays with shareholders.

The real answer will usually sit somewhere between the two cases, and the gap between them is large enough to deserve its own line in the model.

Pricing it

Has revenue per employee risen faster than costs have fallen for reasons other than headcount? Has the margin held for a year against new entrants and rising software bills? Most of all, has it come through a client repricing and a pay round without giving ground?

Where the margin is rising mainly because wage costs are falling relative to output, the productivity may be real, but the investor still needs evidence that the surplus will stay put.

Until those tests are passed, an AI-driven margin gain belongs in the model as a probability-weighted addition to earnings with a shorter persistence period than the established base, and with some allowance in terminal value for a tax system that three fiscal bodies have modelled as exposed to it.

The discount should be larger for labour-intensive professional services than for businesses that already earned very high revenue per employee before AI arrived. Once a gain has survived those tests, capitalise it.

Grant Thornton has already found this out. A margin that came from somewhere can be asked to go back.