IBM's $12.5 Billion Generative AI Book of Business, and What It Counted

IBM's generative AI book of business reached more than $12.5 billion at the end of 2025, and then IBM stopped reporting it. The number was never AI revenue. About four-fifths of it was consulting signings, a measure of contracts won rather than work delivered. This teardown rebuilds the quarterly series from IBM's own filings and shows what the metric could and could not tell you.

Short answer

IBM's generative AI book of business is not revenue. It adds cumulative consulting signings, software transactional revenue and new SaaS annual contract value tied to AI offerings, counted since mid-2023. It reached more than $12.5 billion by December 2025, about four-fifths consulting. IBM stopped reporting it after that quarter.

What IBM's generative AI book of business actually measured

IBM introduced the figure in 2023, after it launched watsonx, its AI and data platform. It then quoted the number on every earnings call through January 2026. The formal definition sat in Exhibit 99.2 of each earnings 8-K.

In its final wording, the book is calculated as inception to date software transactional revenue, plus new SaaS annual contract value and consulting signings related to specific offerings.

Software transactional revenue is IBM's term for one-off licence sales, as opposed to recurring subscription and support revenue. SaaS annual contract value, usually shortened to ACV, is the first-year value of a new cloud software contract. Consulting signings are, in IBM's own definition, management's initial estimate of the value of a client's commitment under a services contract. Extensions and scope increases count only for their incremental new value.

IBM was explicit that the metric is not revenue. Its KPI definitions say the figure is not meant to represent revenue in the current period, or revenue that will be recognised in future periods. That caveat almost never travelled with the headline.

Three ingredients running on three different clocks

Transactional revenue is recognised revenue, booked when the licence ships. ACV is a contract value, a promise of a year of subscription income that has not been earned yet. A consulting signing is an estimate made on the day a contract is signed, for work that may run for several years.

Adding them together produces a number that is part revenue, part contract value and part estimate. None of the three ever leaves the total. A consulting project that is later descoped, delayed or delivered over five years still counts at its signing value, permanently, because the figure is inception to date.

In my view this design was sensible for a launch metric and wrong for a multi-year one. In 2023 IBM needed a way to show traction before meaningful AI revenue existed. By 2025 the same construction was reporting $12.5 billion of cumulative demand, with no published way to see how much had turned into income.

The words IBM added to the definition in 2025

The definition did not stay still. Releases from Q3 2024 through Q1 2025 described the book simply as software transactional revenue, SaaS ACV and consulting signings. The Q2 2025 release added two qualifiers. From Q3 2025 the phrase "inception to date" was written into the software element as well.

Wording taken from IBM's Form 8-K Exhibit 99.1 releases and the January 2026 Exhibit 99.2. None of the changes came with a restated history, so earlier quarters cannot be recomputed on the later definition.

Each change narrows or clarifies what counts. Read the series as a direction of travel rather than a precise measure.

The quarterly series: from low hundreds of millions to $12.5 billion

IBM never published the series as a table. Rebuilt from its releases, chart decks and call remarks, it runs as follows. Every value from Q1 2024 onward is a floor, because IBM stated each one as "more than" a round number.

Changes marked "per IBM" are IBM's own wording. The others are our floor-to-floor arithmetic on stated minimums, so the true step in any quarter may be larger or smaller by up to the rounding in IBM's figures.

The quarterly increments got bigger almost every quarter

Floor to floor, the quarterly addition grew from about $1 billion through 2024 to about $3 billion in Q4 2025. Over calendar 2025 the book grew by at least $7.5 billion, from $5 billion to $12.5 billion. Q4 is seasonally IBM's largest signings quarter, which explains part of the final jump.

IBM's CFO, James Kavanaugh, gave the consulting piece of that step on the January 2026 call. He said the consulting generative AI book of business surpassed $2 billion in the quarter, calling it IBM's largest quarter of generative AI. So consulting supplied most of the Q4 increase on its own.

A cumulative series of signings can only rise. The informative signal was always the size of each step, and that step was still growing when IBM stopped publishing it.

The Q3 2025 release called the figure the "AI book of business", dropping the word generative. The Q4 2025 release restored it. No definitional change was announced alongside either label.

Consulting was four-fifths of IBM's AI book of business

On the January 2026 call, CEO Arvind Krishna split the final figure. He said software was more than $2 billion and consulting more than $10.5 billion, within a cumulative book of over $12.5 billion. That works out to roughly 16% software and 84% consulting.

The split is close to the "approximately one-fifth" software share IBM wrote into its definitions exhibit, and it had been stable for a long time. In Q4 2023, consulting was already about two-thirds of the book, according to CFO Brew's report of that call. IBM's Q4 2024 chart deck carried a separate line, generative AI bookings above $4 billion inception to date, next to the $5 billion total. If that line is the consulting component, as the gap to the total suggests, consulting was near four-fifths a year before the final split.

Why signings grow faster than software revenue

A consulting signing records the estimated value of a whole contract on the day it is signed. A software licence records only what shipped in the quarter, and a new SaaS deal enters at one year of value. So a multi-year transformation engagement enters the book at its full estimated value immediately, while a SaaS contract of the same annual size enters at a fraction of its lifetime worth.

The mix is therefore partly a measurement artefact. I would expect software's real share of IBM's AI economics to be larger than one-fifth once subscription renewals are included, because the book excludes them by design.

Acquisitions flatter IBM's 2026 software growth. How much was organic is worked through in the analysis of IBM software revenue and the Confluent contribution.

IBM AI revenue versus the book of business: the reconciliation IBM never published

The question every analyst asked was how much of the book had become revenue. IBM answered in fragments rather than a table. The Q4 2025 call offered three of them in a single answer from Kavanaugh to a Goldman Sachs analyst.

He said generative AI now represented over a third of consulting bookings, over 25% of a $32 billion backlog, and over 15% of revenue on an exit run rate. The answer was about consulting, so the backlog and revenue in it are best read as Consulting figures.

What the backlog share implies about delivered work

Over 25% of a $32 billion backlog is more than $8 billion of generative AI consulting work signed and not yet delivered. Set that against a cumulative consulting book above $10.5 billion. If the two measures cover the same scope, at most about $2.5 billion of the consulting book had been delivered as revenue since 2023.

That condition may not hold. Backlog penetration could count any engagement with generative AI embedded in it, while the book counts signings related to specific offerings. IBM did not say which. Treat the $2.5 billion as an illustration of how far orders and income can sit apart, not as a measured figure.

The exit run rate is the closest thing to IBM AI revenue

The 15% exit run rate is the only revenue-shaped number IBM offered. Consulting revenue was $5.35 billion in Q4 2025. If the 15% applies to that segment, generative AI consulting revenue was running above $800 million a quarter, or above $3.2 billion annualised.

The $12.5 billion headline was close to 4 times the annualised consulting AI revenue it implied, and even that comparison flatters the headline, because the book also includes software. A cumulative metric hides this gap by construction.

In my view IBM should have published the run rate every quarter from 2024, next to the book. The run rate is the number an outsider can check against segment revenue. The book could only be checked against itself.

What replaced IBM's generative AI metric in 2026

Kavanaugh announced the change on the January 2026 call. "This will be the last quarter in which we report this metric separately. AI is now embedded across our business," he said, according to The Motley Fool's transcript. He added that a standalone generative AI metric no longer reflected the full scope of how AI drives value across IBM.

Since then IBM has talked about shares rather than totals. MarketBeat's summary of the July 2026 call reports that generative AI was about 50% of Consulting signings in Q2 and more than 30% of backlog. We could not open a transcript of that call, so treat both shares as directional. The Q3 2026 earnings release itself contains no generative AI dollar figure, no signings figure and no backlog figure.

Signings and backlog went the wrong way in Q3 2026

The totals beneath the shares are in the 10-Q. Consulting signings were $5.03 billion in Q3 2026, down 3.7% as reported and 2.0% at constant currency, against $5.22 billion a year earlier. Nine-month signings were $15.42 billion, up 3.1%.

Backlog was $30.2 billion at 30 September, down from $30.8 billion at 30 June and about $32 billion at the end of 2025. That is a fall of about 5.6% in nine months, computed from IBM's own figures. Consulting revenue was flat at $5.34 billion, while Software grew 8.2% to $7.80 billion.

A dollar figure is exactly the information the old metric supplied, however crudely. If about half of Q2 2026 signings of $5.03 billion were generative AI, that is roughly $2.5 billion in one quarter. Under the old construction, consulting alone would have added about as much as its record Q4 2025 step. It is a strange moment to stop counting, and I would not accept "AI is embedded everywhere" as the whole explanation.

How IBM's AI bookings metric compares with other vendors' AI disclosures

IBM was not alone in building a bespoke AI number, and it was not alone in retiring one. Accenture reported generative AI new bookings of $1.8 billion in its fourth fiscal quarter of 2025 and $5.9 billion for the year. Press coverage of its December 2025 results then reported that it would stop breaking out advanced AI figures, because AI is now part of most client projects.

IBM followed about 6 weeks later with almost the same rationale. What Accenture's bookings looked like before and after is covered in the breakdown of Accenture's bookings and the services-as-software shift.

A point-in-time recurring metric like ARR can fall. A cumulative signings tally cannot. That single property makes ARR and ACV more informative even when they are smaller numbers. A ranking of AI revenue metrics by rigour is in the scorecard of seven SaaS AI disclosure metrics.

The services firms that still publish an AI revenue share, and the ones that stopped, are tallied in the firm-by-firm review of AI consulting revenue disclosures. Salesforce's figures are tracked in the Agentforce ARR disclosure ledger.

Where this argument is weakest

The case against IBM's metric is easy to overstate. Here is where it is thin.

Retiring a cumulative number can be the honest call

A cumulative tally gets less informative every quarter it survives, because old signings dominate it. Kavanaugh's explanation is also plausible on IBM's own evidence. If half of consulting signings involve generative AI, separating "AI deals" from the rest becomes an accounting judgement that invites gaming.

The signings decline is also a single quarter. Nine-month signings were still up 3.1%, and the trailing twelve-month book-to-bill ratio was about 1.05 at 30 June, per IBM's Q2 2026 10-Q. A book-to-bill above 1 means IBM was still signing more than it billed. A backlog drift of about 5.6% across nine months is a fact worth watching, not a verdict.

Our reconciliation is the second weak point. The $2.5 billion and $3.2 billion figures above assume that the backlog share, the run-rate share and the book all measure the same scope. IBM never said they do, and the Q4 call transcript is a third-party transcript that carries its own error warning.

A third weakness is coverage. The Q2 2026 share figures come from a call summary, not a transcript we opened. We could not open IBM's Q3 2026 chart deck at all, so no Q3 2026 generative AI share appears in this post. If that deck shows a large dollar increase in AI backlog, the drain reading above weakens considerably.

Frequently asked questions

Where to start

If you follow IBM, or you sell into accounts where IBM Consulting competes, rebuild the retired number yourself each quarter. Take Consulting signings and backlog in dollars from the 10-Q, then apply the generative AI share management quotes on the call. The product is a rough AI backlog figure that IBM no longer prints, and it can go down.

If you report an AI metric of your own to a board or to investors, run one test before the next meeting: ask whether your number can fall. If it cannot, pair it with one that can, such as recurring or delivered revenue.

Related reading