Who Actually Deploys Enterprise AI: Services Firms' AI Revenue, Tallied

Tata Consultancy Services now says AI work runs at $3.1 billion a year, more than 10% of its revenue. Infosys puts AI at 8.2% of revenue and HCLTech at about 5.9%. Accenture, the firm that started this kind of AI consulting revenue disclosure, stopped publishing its numbers in December 2025, and IBM's last figure is from January. The firms with the most to prove are now the ones publishing.

The short answer to who deploys enterprise AI is that services firms do, at scale, and only some of them will tell you how much. The firms that publish an AI figure put it at between roughly 4% and 10% of revenue. The figures cannot be added up cleanly, because every firm defines AI work differently, and the two largest US-listed firms have withdrawn theirs.

The AI consulting revenue tally, firm by firm

This table records, for each firm, the latest AI figure it published, what that figure measures, the period it covers and what share of revenue it represents. Where a share is our arithmetic rather than the firm's, the table says so.

Shares marked "our arithmetic" divide the disclosed AI figure by the same period's reported revenue or bookings. HCLTech's Q2 FY27 figures come from press coverage of its release. Infosys's dollar figure is 8.2% of $5,082 million. The rows are not comparable with one another, for reasons set out in the next section.

Add up the four firms that report an AI revenue run rate and you get about $10 billion a year. That is TCS's $3.1 billion, Infosys at roughly $1.67 billion annualised, HCLTech's stated $876 million run rate, and Accenture's last quarter of about $1.1 billion multiplied by four. Against roughly $140 billion of annualised revenue at the same four firms, that is about 7%.

Treat that 7% as directional at best. It mixes four definitions and one figure that is now ten months old. It is still the only cross-firm number the public record supports.

Four firms, four definitions of AI work

A disclosed AI figure is only as useful as its definition. Here the definitions diverge enough that a reader comparing headline numbers is comparing different objects. That is not a criticism of any one firm. It is a property of a metric that no accounting standard governs, which is the same problem we found in the ranking of SaaS vendors' AI revenue metrics.

Accenture's definition was the narrowest, and the most honest

Julie Sweet set out the boundary on the September 2025 and December 2025 calls. Advanced AI meant generative, agentic and physical AI. It did not include data work, classical machine learning, robotic process automation, or AI Accenture uses internally to deliver projects.

That exclusion list matters. Data engineering is most of the labour in a real AI deployment, and Accenture chose to leave it out. A narrow definition produced a smaller number. In our view that made Accenture's figure the most conservative of the set, which is the opposite of what a reader assumes when they see the largest firm with a modest share.

A run rate, a percentage and a cumulative total are three different things

TCS annualises a single quarter. Infosys publishes a percentage and no base. IBM publishes a running total of signings since mid-2023, so its number can only ever go up, even in a quarter with no new AI work.

IBM's own exhibit notes that about one-fifth of the book is software and four-fifths is consulting. So roughly $10 billion of IBM's $12.5 billion is consulting signings accumulated over about two and a half years. That is a large figure. It is also not comparable to anything else in the table.

The percentage with no definition

Infosys led its July release with "AI Revenues at 8.2% in Q1". The body of the release gives no dollar figure and no definition. We could not find what Infosys counts as AI revenue in the release itself, and we would not lean on the number until it does. That is a disclosure choice, and buyers should read it as one.

Accenture and IBM stopped publishing AI figures

The more telling pattern is not who discloses. It is who stopped. The two largest US-listed firms in this tally both withdrew their headline AI number between December 2025 and the first quarter of 2026.

What Accenture said when it stopped

On 18 December 2025, Sweet told analysts that "this will be the last quarter in which we share these specific metrics". Her stated reason was that demand had shifted toward scaled programmes that combine several forms of AI, so isolating advanced AI no longer reflected how clients buy.

The final figures were respectable. Accenture had delivered about $11.5 billion of advanced AI bookings across 11,000 projects and $4.8 billion of revenue since it began counting. In fiscal 2025 alone, AI revenue tripled to $2.7 billion and bookings nearly doubled to $5.9 billion, about 3.9% of $69.67 billion of revenue.

The reason given is plausible. It is also what a firm would say if the AI share were growing more slowly than investors hoped. We cannot tell which is true from the outside, and that is the cost of a withdrawn metric. Since then, Accenture's releases have carried only qualitative AI language, as covered in our reading of Accenture's record fiscal 2026 bookings.

IBM kept the definition and dropped the number

IBM's January 2026 release said its generative AI book of business stood at "more than $12.5 billion". Its April and July 2026 releases still define the metric in the supporting exhibit, but the dollar figure does not appear in the release text we read. Keeping a metric's definition while removing its value is an unusual choice. It leaves the door open to bringing the number back when it flatters.

Why Indian IT firms started disclosing AI revenue

The firms now publishing are growing slowly overall. Infosys cut its fiscal 2027 guidance to 1.5% to 3.0% constant currency growth in July. TCS revenue rose about 0.5% sequentially in constant currency in its September quarter. HCLTech's dollar revenue rose 2.2% year on year in the same quarter, according to coverage of its release.

Against that, the AI lines are the only fast-growing numbers on the page. HCLTech's advanced AI revenue rose from $171 million in the June quarter to $219 million in the September quarter. That is up 78% year on year in constant currency, according to Analytics India Magazine's report on the results. A firm with a slow headline and a fast segment has every reason to publish the segment.

The deflation the headline hides

An AI revenue line shows what is being won. It does not show what is being lost. When a services firm uses AI to deliver a maintenance contract with fewer people, the client often captures the saving as a lower price. That revenue loss lands in the core business, not the AI line.

TCS said on its call that AI work is seeing better pricing and that some AI businesses carry margins above the company average, according to coverage of the September quarter results. No firm in this tally publishes the matching number for revenue lost to AI-driven repricing. Until one does, a rising AI share can mean AI is growing, or that everything else is shrinking. This is the same trap examined in the analysis of the billable-hour problem.

Deloitte's $74.5 billion with no AI revenue line

Deloitte reported aggregate global revenue of $74.5 billion for the year ended 31 May 2026, up 5.7% in dollars and 3.8% in local currency. Its workforce grew to nearly 500,000. That makes it roughly the same revenue size as Accenture's fiscal 2026 total of $74.18 billion.

The release discusses AI at length. It names Zora AI, an agentic platform built with NVIDIA. It describes a $3 billion investment through 2030 to modernise delivery and operations. It does not give an AI revenue figure, an AI bookings figure, or a count of AI practitioners.

This is not evasion in any regulatory sense. Deloitte is a network of private partnerships and owes the public no segment reporting. It does mean that a buyer comparing implementation partners has one of the two largest firms in the market reporting nothing measurable on the work being bought. Our position is that buyers should ask for the number directly in a proposal, and treat a refusal as information.

Where enterprise AI deployment labour actually sits

Revenue tells you who is paid for AI work. Headcount tells you who is doing it. The five firms in this tally with a published workforce figure employ about 2.46 million people between them, by our addition of their latest reported totals.

Accenture is the only firm that has put a number on its specialist AI workforce. Sweet said on the September 2025 call that the firm had grown from 40,000 AI and data professionals in fiscal 2023 to 77,000. By December 2025 it had nearly reached its 80,000 goal. More than 550,000 of its people had been trained in generative AI fundamentals.

The specialist ratio is about 1 in 10

Nearly 80,000 specialists against about 784,000 total people in December 2025 is roughly 10%. That lines up closely with Accenture's AI share of bookings in the same quarter, about 10.5%. It is a coincidence of two independent numbers, not a law. It does suggest that, at the largest firm, the AI specialist bench is sized to the AI order book rather than ahead of it.

Accenture also said in September 2025 that it was exiting, on a compressed timeline, people for whom reskilling was not viable. Its total headcount still rose from about 779,000 to about 814,000 over the following year. Deployment labour is being reshaped, not shrunk, at least at this firm. The bottleneck that creates inside your own organisation is covered in the piece on AI centres of excellence.

Where this tally is weakest

This post makes a cross-firm comparison that the underlying data does not fully support.

The definitions are incompatible, and the sum hides it

The roughly $10 billion combined run rate adds a narrow Accenture definition to undefined Infosys and TCS definitions. If TCS and Infosys count data and classical machine learning work that Accenture excluded, their shares are inflated relative to Accenture's. If they count less, the reverse holds. We do not know which. The ranking in the first chart could change order under a common definition, and the honest reading is a range of 4% to 10%, not a league table.

Several figures also rest on press coverage rather than the firms' own documents. The TCS and HCLTech September quarter numbers come from Analytics India Magazine's reporting, because neither firm's release page could be opened during research. TCS's dollar revenue appears as $7.64 billion in some coverage and $7.56 billion in others. Each tooltip states which tier of source a figure comes from.

The case that the shrinking disclosure means nothing

The strongest counter-argument to this post is Sweet's own. If AI is now embedded in nearly every engagement, a separate AI line becomes arbitrary. A cloud migration with an agent attached is either an AI project or it is not, and the answer depends on who is counting. On that reading, the Indian firms' new disclosures are the less mature practice, and Accenture's withdrawal is what every firm will do once AI is ordinary.

We think that argument is partly right and still incomplete. A metric does not stop being useful because it is hard to draw. Withdrawing a number at the moment investors are most anxious about it, without replacing it with anything measurable, removes the one test of whether the spending is reaching the services layer. That is a loss for buyers as much as for shareholders, and the regulatory risk of loose AI claims is set out in our log of SEC AI-washing enforcement.

Frequently asked questions

How to use these disclosures when you pick a partner

If you are choosing an implementation partner this quarter, do one thing before the next meeting. Ask each shortlisted firm for its AI revenue or AI bookings figure, in writing, with the definition attached. Ask specifically whether data engineering and classical machine learning are included.

The answers will sort themselves. A firm that can state the number and the boundary has measured its own practice. A firm that answers with a client count or a training statistic has not, or will not, and you are entitled to weigh that as heavily as any case study in the deck.

Related reading