How Generative AI Could Reshape Professional Services

Corporate legal departments employed about 145,000 lawyers in 2024, up from 78,000 in 2008. That 87% expansion, calculated by the Association of Corporate Counsel using federal labor data, far exceeded the 23% growth in law-firm lawyers over the same period.

Generative AI did not cause a trend that began 14 years before ChatGPT arrived. It could accelerate it.

That distinction separates a plausible threat to professional-services firms from the familiar keynote fog. AI can help companies perform research, draft documents and analyze records with fewer billable hours purchased from outside advisers. Yet the available evidence does not establish that AI has reduced professional-services revenue, displaced firms at scale or made expert judgment conveniently downloadable.

The sharper question is narrower: Which tasks can clients bring inside their organizations once software lowers the cost of doing them, and what remains worth buying from a specialist?

The Strongest Version of the Threat

Jonathan Peachey, chief executive of Factory X, described professional services in an AWS interview as selling “people’s brain power for money.” His argument is that AI reaches closer to the product in law, consulting and accounting than it does in a company using AI to improve its marketing.

The logic runs in three steps. Professional firms charge for skilled labor. Generative systems can perform parts of research, drafting, classification and review. If a client can combine those systems with its own employees and internal information, some work previously sent outside becomes cheaper to perform in-house.

Each step is credible. The conclusion still depends on task quality, supervision costs, liability, confidentiality and whether the client has enough recurring work to justify building an internal operation. A general counsel may automate contract comparisons without wanting an AI system, or an internal employee supervising it, to run a bet-the-company lawsuit.

Client demand also complicates the replacement story. A 2025 Thomson Reuters Institute survey of more than 1,700 professionals in the United States, United Kingdom and Canada found that 59% of corporate legal clients wanted their outside firms to use generative AI. Among law-firm clients, 71% did not know whether their firms were using it.

Those results point toward pressure on delivery and pricing rather than immediate abandonment of outside counsel. Clients may continue buying expertise while expecting the supplier to use cheaper tools. If an AI-assisted task takes two hours instead of ten, the economic fight concerns who keeps the saving: the client, the firm or the software vendor.

In-Housing Came First

The long growth of corporate legal departments supplies useful history because it identifies the mechanism AI might strengthen. From 2008 through 2024, in-house lawyer headcount rose 87%, compared with 23% at law firms and 38% in government. ACC derived its estimate by subtracting government and law-firm employment from the total lawyer population counted by the US Bureau of Labor Statistics.

Headcount cannot reveal how much outside work those departments retained, how productive their lawyers became or why companies hired them. Regulation, privacy, ethics and risk all expanded as corporate concerns during those years. The figures nevertheless show that companies were already accumulating the people needed to absorb more legal work.

AI adds leverage to that installed workforce. An organization with lawyers, proprietary contracts and years of internal records has raw material for automating repeatable work. A company without those assets cannot conjure an effective legal department by buying chatbot seats and scheduling a celebratory webinar.

This history also narrows what can responsibly be claimed. Any current shift toward in-housing has multiple causes, and the ACC headcount series ends in 2024. It offers no post-adoption revenue measure capable of isolating AI’s effect.

The Software is Becoming Part of the Supply Chain

OpenAI’s September 2026 launch of Astra for Law illustrates how the market could be rearranged without eliminating professional firms. The product combines a model with a US legal-search index spanning more than 230 million URLs, governance controls and 26 plugins connecting it to products including Relativity, Clio and iManage. OpenAI said selected law firms would receive initial access.

That design bundles a model, legal information and workflow connections while leaving lawyers in the loop. It can strengthen firms that integrate it well, help corporate departments perform more work themselves and shift bargaining power toward the company controlling the model and distribution interface. All three outcomes can coexist.

Performance remains a constraint. OpenAI tested Astra for Law on 200 questions from a private validation set and reported that it passed the overall correctness check on 54% at the highest reasoning setting, versus 38.7% for its general model with web search. The Next Web noted that OpenAI conducted the testing and no independent party had audited the results.

A 54% pass rate may make a system useful for supervised research while leaving it unsuitable for unsupervised legal conclusions. Automation can still alter economics before it reaches perfection. Junior employees and document-review teams do not achieve perfection either, but firms have training, review and accountability structures for their mistakes. AI systems need an equivalent operating layer, and building one costs money.

Adoption Statistics Cannot Settle the Business Case

Only 20% of legal professionals in the Thomson Reuters survey said they were measuring return on investment from generative AI. That leaves a peculiar information gap: clients want firms to use the technology, while most surveyed professionals cannot demonstrate its return through measures such as cost savings or client satisfaction.

Headline failure rates do not repair that gap. Operational-resilience practitioner Jamie Watters traced the repeated claim that 80% of AI projects fail. The citation chain led from RAND’s cautious phrase “by some estimates” to a 2022 magazine article referring to unnamed surveys of executive opinion. RAND itself conducted 65 interviews about perceived failures and produced no project failure rate.

That provenance does not show that AI projects usually succeed. It shows that the famous percentage cannot answer the question. “Failure” might mean cancellation, no organization-wide deployment, missed ROI, poor output or an executive feeling glum about the invoice. Those outcomes require different denominators and remedies.

For professional-services firms, the useful measures are less theatrical: hours removed from a defined workflow, error and rework rates, supervision time, total operating cost, client retention, realized fees and the share of work brought in-house. Without those figures, a pilot can impress a partnership committee while changing little beyond the software budget.

Remember the Corporate-App Rush

Peachey compared today’s AI pressure with the period when businesses responded to the iPhone by asking, “Where’s our app?” The analogy works as a warning against copying a technology before identifying the customer problem. It has limits because generative AI can change internal production, while many early corporate apps were primarily new customer interfaces.

A 2019 account of Forrester’s retail-app research describes usability testing involving nine retailers, 196 shoppers and 52 criteria. The stronger apps helped people navigate stores, check inventory, filter products or coordinate online orders with pickup. A Macy’s feature allowed shoppers to scan an item and buy it in the app, only to require pickup from another area even when the item was already in their hands. Digital transformation, meet the extra errand.

The comparison suggests a practical test. A professional-services AI system needs to remove friction from a valuable workflow, preserve the controls demanded by high-stakes work and change the economics enough for someone to care. Merely adding a chat box to the firm’s knowledge base repeats the glorified-mobile-website mistake with more expensive computing.

Peachey argues that firms need internal platforms capable of encoding their expertise before they sell AI products to clients. “Platform” needs careful handling. Research summarized in a 2019 Harvard Business Review article distinguishes innovation platforms, which support complementary products, from transaction platforms, which facilitate exchanges. An internal collection of models, data and workflows may be valuable infrastructure without becoming either kind of platform. The label can make an integration project sound grander than its network effects.

The evidence supports concern, experimentation and sharper accounting. It does not yet support an obituary for professional services. Firms face the greatest exposure where work is repeatable, information-rich and easy to review; they retain stronger defenses where clients pay for accountability, negotiation, institutional judgment and responsibility when the answer goes wrong.

AI may let clients buy fewer hours. The firms that survive that pressure will have to explain what the remaining hours accomplish, and clients finally have the tools to ask.