AI at the Firm: What Actually Changes the Profession — and What Stays with the Professional

The most exposed sector is also the most advanced

Law, audit, accounting, consulting: no sector works closer to AI's raw material — text, numbers, documents. The consequence shows in the statistics: 17% of French professional, scientific and technical firms used at least one AI technology in 2024, seven points above the national average (INSEE); at European level the sector reaches 40.4% — second only to information and communication, while every other sector stays below 25% (Eurostat).

Among accountants, declared adoption has become near-universal — 92% of firms report using AI tools — but only 25% have generalized that use (Qonto–Ifop barometer). This wide gap between trying and scaling is the sector's real issue: the question is no longer adoption, it is industrialization — and it turns on client-data confidentiality, result reliability and team training.

The sociology of leadership confirms the sector's head start: business services are over-represented among the "Innovators" — 19% of the sample, running highly digitized companies that train their teams and build AI into their processes and their offerings (Bpifrance Le Lab). The market context makes the subject strategic: the French strategy and organization consulting market was flat in 2024, with a forecast of −2 to +1% for 2025 (Xerfi / Syntec Conseil) — when volume no longer drives growth, productivity and differentiation replace it.

One principle runs through this report, and it is specific to regulated professions of trust: AI industrializes the first draft and the review — judgment, signature and the client relationship remain the profession. The firms that succeed do not delegate their responsibility: they move their teams' time from raw production to control and counsel.

The Ashurst lesson: speed is not accuracy

The sector shares the general paradox — 88% of organizations use AI (McKinsey), barely 5% have scaled it (Capgemini / Microsoft), and about 95% of generative AI pilots produce no measurable return (MIT Project NANDA) — but it has something rare: controlled trials run by the professions themselves, which state precisely where AI helps and where it misleads.

The most instructive comes from Ashurst: trials across 411 participants, 23 offices, 14 countries. First drafts come out markedly faster — up to 80% time saved on regulatory filings, 59% on sector reports, 45% on legal briefings — but the average accuracy of AI deliverables comes out at 3 out of 5, against 3.5 for humans (Ashurst). Fast does not mean right: the net gain only exists if professional review is built into the process — and into the price.

The BCG × Harvard field experiment on 758 consultants maps the geography of the gain: on tasks inside AI's competence "frontier", +25% speed, +40% assessed quality, +12% tasks completed; beyond that frontier, performance degrades (BCG × Harvard). The usage map — what AI does well, what it does badly — is not a deployment detail: it is half the value.

Up to 80% time saved on the first draft, 3/5 accuracy against 3.5 for humans: the whole economics of the augmented firm sits between those two numbers.

Five families already transforming the firm

As in other sectors, value concentrates in five use-case families with three common traits — a repeated task, a measurable cost, a gain that translates into hours, margin or avoided risk. All are documented by real deployments, from the solo practitioner to the global network.

Drafting and first drafts is the most immediate family. The Ashurst trials above set its yield — and its condition. At Bain, generative AI cuts the drafting time of post-merger integration plans and transition agreements by about 80% (Bain & Company). The family's golden rule: the first draft is delegated, the signed version never is.

Document review and due diligence is the oldest family — and the most quantified. JPMorgan's COIN automates the equivalent of 360,000 lawyer-hours per year on credit-contract review (JPMorgan); Zurich saved 40,000 hours on claims document review (Zurich Insurance); Clifford Chance industrializes clause extraction with Kira (Clifford Chance). On legal platforms, A&O Shearman deployed Harvey to 3,500 lawyers across 43 offices — some 40,000 queries during the beta, on contract analysis and due diligence (A&O Shearman / Harvey) — with Slaughter and May, BonelliErede and Harneys following firmwide (Harvey).

Accounting production and review is the family of everyday firms. KPMG Nord moved 650 clients onto an AI production engine in two-to-three-month waves: 20 to 30% of data-entry time saved depending on file type, and up to 98% document recognition (Cegid, KPMG Nord case). At the other end of the spectrum, OB Expert Conseil — an accountant practicing alone — automates review worksheets, anomaly checks and a pre-filled AML questionnaire (Cegid). The profession is organizing itself: the CNOEC publishes use cases, precautions and a model charter for generative AI use (CNOEC).

Audit and tax is the family of the big platforms — and the proof that the movement is structural, not experimental. KPMG embeds AI agents in Clara, its global audit platform serving more than 95,000 auditors (KPMG); Deloitte does the same in Omnia (Deloitte); EY launched EY.ai after 1.4 billion dollars of investment and integrates AI into its assurance platform (EY); PwC is building its AI-assisted audit platform (PwC). Tax follows the same slope, from filing compliance to regulatory watch (KPMG; Deloitte). For a mid-size firm the message is not to imitate these platforms: it is that your clients will be audited and advised by them — the market's productivity standard is moving.

The firm's knowledge, finally, is the family that capitalizes everything else. At McKinsey, the Lilli assistant — built on 40+ knowledge sources and 100,000 documents — is used by 72% of the firm, over 500,000 queries a month, saving up to 30% of research and synthesis time (McKinsey). Linklaters handles about 60,000 queries a week with its internal assistant Laila (Linklaters); Clifford Chance exceeds 60% daily adoption of its secure assistant (Clifford Chance); PwC deployed ChatGPT Enterprise to more than 100,000 employees (PwC), and Deloitte is making Claude available to over 470,000 people with a 15,000-professional certification program (Deloitte / Anthropic). In France, CMS Francis Lefebvre deploys GenIA-L, co-developed with Lefebvre Dalloz on its legal editorial corpus (CMS Francis Lefebvre). The prerequisite, as everywhere: a clean, governed documentary asset.

Five families, one constant: AI produces, the professional signs. The firms that win moved the time — not the responsibility.

Proof at firm scale

The objection "this is a global-network story" does not survive recent French deployments. From the solo practice to the 150-lawyer firm, the winning rules are the same: narrow scope, governed data, a baseline figure, a dated decision.

The signal is clear: training and governance are being institutionalized. In a regulated profession, that is the best indicator that a technology is leaving experimentation and entering the normal practice of the profession.

True costing — and the billable-hour question

The general economics apply to the firm: 10% algorithms, 20% technology and data, 70% processes and change management (BCG); 55 to 75% of total cost outside the license (YPrtnrs Institute estimate); a ×2 to ×4 factor between announced budget and observed cost when the 70% stays invisible (Gartner); and an average return of 3.7 dollars per dollar invested (IDC).

But the firm has an economic question the factory does not: when production accelerates, what happens to the billable hour? A first draft produced five times faster, billed on time spent, destroys revenue — unless value moves toward fixed fees, counsel and the relationship. In a flat consulting market (−2 to +1% forecast for 2025 (Xerfi / Syntec Conseil)), AI is therefore not just a cost lever: it is a revenue-model choice, and the firms that anticipate it turn a risk of hourly deflation into a margin advantage. Observed timelines frame the decision: weeks for first drafts and data entry, months for internal knowledge, 12 to 24 months for audit platforms and due diligence at scale.

A firm project's journey — and its four breaking points

Firm projects that stall look alike. Laid on top of each other, they reveal a typical journey and four breaking points — two general, two specific to professions of trust.

1. Break #1 — the showcase pilot without a number. The tool is tried "to see", with no measured baseline cost (drafting, entry or research hours): it joins the ≈ 95% with no measurable return (MIT NANDA). The counter: choose within the five documented families, where hours are already counted.
2. Break #2 — the ignored frontier. AI is applied indiscriminately to everything, while performance degrades beyond its competence frontier (BCG × Harvard). The counter: write the usage map — what AI does, what it does not do — and train it into the teams; that is exactly what the CNOEC charters and the CNB training plans are for.
3. Break #3 — the forgotten 70%. The budget describes the license, then discovers the training, the review-process adaptation and quality control — hence the ×2 to ×4 factor (Gartner). The counter: a total-cost budget where the 70% is visible before signing.
4. Break #4 — trust without control. The public lesson came from consulting itself: Deloitte Australia had to partially refund a 290,000-Australian-dollar government report containing AI-fabricated citations (The Register). In a profession where the signature binds, every AI deliverable passes professional review until its error rate is characterized — Ashurst puts numbers on it: 3/5 without review, against 3.5/5 for humans. The counter: signed review is in the process, in the price — and in the firm's charter, along with professional secrecy and client-data protection.

In a profession where the signature binds, review is not an overhead of the AI project: it is the product.

What the managing partner must decide — and when

Neither a multi-year program nor wait-and-see: the trajectory that works holds in five decisions over ninety days, each with its indicator.

1. Choose a family, not a tool. Among the five documented families, take the one where hours are already counted. Indicator: a validated baseline figure (drafting, entry or research hours per file).
2. Audit the documentary asset of the chosen case. Templates, precedents, files: existing, accessible, governed? Indicator: a written verdict within fifteen days.
3. Cost at total cost and settle the revenue model. License + integration + training + quality control (the 10/20/70 rule); and decide what becomes of the hour saved: rebilled, reinvested in counsel, or kept as margin. Indicator: a budget where the 70% is visible and a written pricing position.
4. Write the usage map and the charter before starting. What AI does, what it does not, who reviews, who signs. Indicator: charter circulated and training scheduled (the national CNB plan validates up to 4 hours).
5. Instrument the measurement from day one. Hours per file before/after, same scope, review-rework rate. Indicator: a dashboard active from week 1.

Doing nothing has a sector-specific cost: in a flat market, productivity standards move with the big networks' platforms, and the gap widens silently — file by file, fee by fee. The firms cited in this report had no more resources than their peers: they had a method, and applied it early.