Professional Services AI: Who Pays, Who Checks, Who Signs

Beyond the accountability research driving this week's feature article, the wider professional services cycle produced three separate stories: finance chiefs approving budget on a steepening curve, a regulator that has already answered the human-review question for one profession in writing, and two opposite bets on who ends up capturing the savings.

Related: Professional Services AI Now Has to Show Its Work

Finance Chiefs Stopped Playing Defense

Deloitte's Q2 UK CFO Survey, reported by Consultancy.uk, found 73% of respondents optimistic that AI will materially enhance business performance, up from 59% at the end of 2025 and 39% in 2024. That curve is steep, and it reads differently alongside what else moved in the same quarter.

Corporate priorities rotated between Q1 and Q2. Cost reduction fell from 68% to 53%, while new products, services and market expansion rose from 25% to 37%. Acquisition-led growth and cash flow both edged up behind them. Finance leadership is moving off balance-sheet defense and onto growth spending, which is a different buying posture than the one most artificial intelligence consulting engagements were sold into two years ago.

Investment intent is close to unanimous, above 90% on both the one-year and the five-year horizon. On returns, 78% expect greater productivity and improved business performance over five years, while only 50% expect productivity gains inside 12 months. That split is unusually candid: the money is being committed well ahead of the payback anyone will publicly claim.

Two caveats belong on this. The sample is 58 CFOs, more than half of them from FTSE 100 and FTSE 250 companies, so it captures UK large-cap sentiment rather than the market. And sentiment is not measurement; it is what finance leaders believe will happen. Darren Graves, CEO of Deloitte UK, called the growing optimism about AI's effect on productivity encouraging, which is a fair reading of a line that has nearly doubled in two years.

One number quietly frames the rest. Geopolitical risk, ranked the top concern in 16 of the last 18 quarters, dropped from 79 to 68 on the survey's 0-100 scale. Attention is coming off external threat and going onto internal capability, which is usually when budget gets approved.

Source: Consultancy.uk

One Profession Already Has Its Answer in Writing

For accountants and tax practitioners, the question of which AI output still requires a human signature is not open. The IRS Office of Professional Responsibility settled it in OPR Alert 2026-19, issued on 24 June, and Thomson Reuters' tax and accounting team has since walked through what it demands. The alert creates no new rules. It maps AI use onto Circular 230 obligations practitioners already carry, which is precisely why it arrived without a comment period or a phase-in.

The mapping is specific: §10.22 on due diligence in preparing and filing, §10.27 on fees, §10.35 on competence, §10.36 on firm procedures and supervision, §10.37 on written advice, and IRC §§6713 and 7216(a) on taxpayer data. The operative instruction is that AI output should be treated as a starting point, never a finished product, and that practitioners independently verify facts, citations and calculations before anything reaches a client or the IRS.

Competence now extends to the tools themselves. A practitioner is expected to understand how an AI system generates content, where its limits sit, and where errors, bias or unreliable output are likely to surface. On data, only secure, enterprise-approved systems qualify, and §7216-protected return information generally requires specific signed client consent before it goes near an AI tool.

The billing provision is the one most likely to change behavior quickly. Cost savings from AI are to be passed on openly, and charging a client for manual labor time that AI performed is treated as a potential violation of the fee rules. That is the same pressure client surveys keep describing, arriving as a professional standard rather than a preference, with financial sanctions, public censure, mandatory ethics courses and referral to state licensing authorities standing behind it.

Firm leadership carries obligations of its own: staff training on AI risks, documented protocols for data handling and accuracy monitoring, and vetting procedures for third-party tools. Documented accuracy monitoring is the requirement most firms cannot satisfy today, because it presumes a per-run record of what each tool actually did rather than a policy asserting that somebody checks. AgentPMT's audit feed exists for exactly that gap, showing every agent action in real time and filterable down to the full request and response payloads, so "we monitor accuracy" becomes something a firm can produce on demand.

Source: Thomson Reuters Tax & Accounting

Whether the Savings Reach the Client Is Not a Technology Question

Richard Tromans, writing in Artificial Lawyer, took up the question the client surveys keep raising and argued that the answer has almost nothing to do with AI capability. It depends, in his framing, on whether clients classify legal work as a utility or as a luxury good.

His historical case is that technology reliably lowers end prices. A 1900 transatlantic steerage ticket cost between three and seven weeks of British wages; the equivalent today runs under half a week's. A functional computer costs £200 against a £1,200 premium option. Email and video conferencing have fallen to nearly nothing. Technology changes the means of production, he writes, and eventually drives down the final cost to the customer.

Legal resists the pattern because of how buying works at the top end. If elite legal services are positioned as luxury goods, AI raises firm profitability without touching rates, because a client buying prestige does not want the discount. Part of a luxury good's value, in his words, is that not everyone else can have it. He points to the general counsel who complains about pricing publicly while accepting rising hourly rates privately, and to the NewMod firms that understand the AI pricing advantage and still struggle, because clients will not cross-shop between the two tiers.

A perfume comparison carries the argument: a £500 Mayfair fragrance does not objectively outperform a £25 alternative, and both prices hold. Tromans lands on this being a sociological problem about how a stratified society measures value, rather than a legal or technological one.

This is argument, not survey data, and reads best as such. Its practical use to anyone building an artificial intelligence business strategy for a services firm is the reminder that efficiency and price are joined by market structure rather than by arithmetic. The categories of work a client already treats as a utility are the categories where automation shows up on the invoice first, and they are the sensible place to start.

Source: Artificial Lawyer

A Firm Betting That Senior Judgment Is the Constraint

General Legal, profiled by Bob Ambrogi on LawNext, is six months out of Y Combinator with roughly 400 clients and a newly launched venture financing practice. Its three founders came from Casetext, which Thomson Reuters acquired for $650 million: Ryan Walker, formerly Casetext's CTO and a PhD mathematician rather than a lawyer, is CEO; Javed Qadrud-Din, who ran AI at Casetext, is CTO; J.P. Mohler, an LLM engineer there, is chief product officer and managing partner.

The structural choice is a management services organization, which lets an investable technology company operate alongside the law firm itself. The premise is that retrofitting AI onto a traditional practice is slower than rebuilding the practice around it.

The staffing decision is the part worth studying. Walker's claim is that AI can automate 95% of routine legal work, and the firm's response has been to hire highly experienced lawyers rather than to give junior ones more leverage. If routine drafting is automated, what remains is judgment, and judgment does not scale by adding associates. That inverts the leverage pyramid law firm economics have run on for a century, and it is a concrete answer to where review capacity sits once the drafting is machine-produced. The firm describes a "second brain" approach that keeps work product aligned to each client's strategy, and it has stood up an MCP server so clients' own AI agents can engage the firm directly, which LawNext reports as a first among law firms.

Walker expects real pressure to arrive when large clients simply stop paying for work AI can do. Whether the 95% figure survives contact with practice is unresolved, but the staffing choice is testable within a year, and it is the sharpest available claim about which half of professional services automation is worth paying for. A firm that automates the routine and then buys more senior judgment with the savings has made a specific bet, in public, that others can now measure themselves against.

Source: LawNext

Sources

• CFOs more optimistic about AI's impact on business performance, Consultancy.uk (Deloitte Q2 UK CFO Survey)
• 2026 IRS rules on using AI for tax preparation and filing, Thomson Reuters Tax & Accounting (IRS OPR Alert 2026-19)
• Should Clients Expect Price Cuts Due To Legal AI?, Artificial Lawyer (Richard Tromans)
• LawNext: From Product to Practice, Why Casetext's Former CTO Founded General Legal, an AI-Native Law Firm, LawNext (Bob Ambrogi)