IBM sold $9.5 billion of AI, and its consulting arm did not grow a dollar

The gap raises a difficult question: is AI creating new consulting demand, or replacing the work clients previously purchased?On July 22, IBM put two numbers next to each other that most of the industry read separately, and reading them separately is the mistake. The first number: IBM’s…

The gap raises a difficult question: is AI creating new consulting demand, or replacing the work clients previously purchased?On July 22, IBM put two numbers next to each other that most of the industry read separately, and reading them separately is the mistake. The first number: IBM’s inception-to-date generative AI book of business has reached $9.5 billion, and roughly four-fifths of it comes from Consulting signings rather than software.The second number: Consulting revenue for the quarter was $5.33 billion, flat year over year. Software grew 5% to $7.76 billion. Infrastructure fell 7% to $3.84 billion. So the segment that owns four-fifths of the AI windfall is the one segment of the AI story that did not grow at all.I have not seen a cleaner quantified picture of what AI is doing to a services P&L, and I want to spend this essay on the arithmetic rather than the storyline, because the storyline can be told convincingly in either direction and the arithmetic cannot. Billions of dollars of AI work signed. Zero net growth in the segment signing it.Every partner at every firm, at every size, should sit with that pairing for longer than the earnings cycle allowed, because the same pairing is forming inside their own book right now, and most of them are only looking at the first number.A book of business is a promise, and revenue is what the promise nets out toStart with what the $9.5 billion actually is, because the phrase “book of business” is doing quiet work in every press mention of it. It is an inception-to-date figure, cumulative signings since IBM started counting, not a quarterly revenue line. Four-fifths of it from Consulting means something like $7.6 billion of AI-related consulting has been signed over the life of the program.Meanwhile, the consulting segment produces $5.33 billion in a quarter, the same as it did a year ago. The book is a promise about future work. Revenue is what all the promises, new and old, net out to when they pass through the same twelve months. And at IBM, the netting comes to zero.There are two honest readings of that zero, and it matters which one is true.The first reading: AI work is cannibalizing traditional work roughly one for one. Every dollar of AI work signed is a dollar of legacy application management, process redesign, benchmarking, or plain analyst-hours that the client stopped buying, either because the AI engagement replaced it or because the client redirected a fixed budget. Under this reading, the firm is running to stand still.The second reading is grimmer and more flattering at the same time: the traditional base is actively shrinking, and the AI book is the only thing holding the segment flat. Under this reading, the $9.5 billion is not failing to produce growth. It is producing all of the survival.Notice that from outside the firm you cannot distinguish these two cases, and from inside you barely can, because the client rarely tells you which budget your new engagement came out of. What you can say with confidence is what the pairing rules out. It rules out the version everyone wants to believe, in which AI demand arrives as net-new wallet, stacked on top of everything the client already bought.If that version were true at IBM, the largest AI services signer would be printing consulting growth somewhere north of its software line. It is printing zero.The decomposition every partner should run on their own signingsHere is the exercise I would put in front of any leadership team celebrating its AI pipeline this quarter, and it takes an afternoon, not a study. Take every AI-related signing from the past year and force it into one of three boxes.Box one: net-new budget, work this client would not have bought from anyone in the pre-AI world, funded by money that did not previously come to you or to a competitor.Box two: relabeled spend, work that replaces something the client used to buy from you at a similar price, the strategy review reborn as an AI strategy review.Box three: defensive spend, work you won at a discount, or at compressed scope, because the client made clear that the alternative was doing it themselves with a model and a smaller vendor.The box-two and box-three work is real revenue, and you should keep winning it. But only box one is growth, and my experience of these decompositions is that partners systematically overestimate their share, because the signing conversation always feels new.The client is excited, and the statement of work has words in it nobody wrote two years ago. None of that novelty tells you which budget line died to fund it. The IBM print is what it looks like when a very large firm runs this decomposition involuntarily, in public, through its income statement: boxes two and three turn out to be most of the book.The exercise also produces a second-order benefit that has nothing to do with strategy documents. It forces partners to have the classification argument with each other, out loud, engagement by engagement, and the argument is where the self-deception surfaces.Every firm has a deal everyone privately knows is defensive and publicly celebrates as growth. Getting it reclassified in the room does more for planning honesty than any market model.There is a third factor hiding under the flat line that the decomposition will also surface, and it is the one I find partners least willing to look at. AI-assisted delivery compresses duration. If the same signing now converts to delivered work in fewer months, with fewer people, then a growing book and a flat revenue line can coexist even without any cannibalization of demand, because each signed dollar spends less time on the meter.I do not know how much of IBM’s zero is explained by faster delivery versus displaced demand, and I would distrust anyone who claims to know from outside. For your own firm, you can know. Compare engagement durations on comparable work, two years apart. If they are shortening, your book has to grow just for your revenue to hold still, and your growth target is quietly steeper than the one on the board slide.Four-fifths of the AI book is services, and that cuts both waysThe composition of the $9.5 billion deserves more attention than it got. Four-fifths from consulting signings means enterprise clients are overwhelmingly buying AI as accompanied change rather than as a product. They are not licensing a platform and running with it. They are paying people to come in, sit with the workflows, learn where the bodies are buried, and make the technology land.For anyone selling services, this is the encouraging half of the print, and it is a rebuke to the loudest version of the doom case. If AI simply replaced services, the software line would be absorbing this spend. Instead, the spend arrives shaped like engagements.The other edge of the same fact: whatever pricing pressure AI exerts, services is where it lands, because services is where the AI money is. When four-fifths of the AI book runs through consulting, consulting is the segment that absorbs the client’s eventual question about why the fee still assumes human-speed production.That question does not arrive at signing. It arrives at renewal, after the client has watched the delivery team work and formed a view about how much of the output the machines produced. A software vendor never faces that conversation, because nobody asks a license how hard it worked.A services firm faces it every renewal cycle, and each cycle the client’s estimate of the machine’s share goes up. Offshoring ran the same script a generation ago: the client learns the new cost structure slowly, then all at once, and the fee conversation that follows is not gentle.So the four-fifths number is simultaneously the evidence that services firms are winning the AI transition and the mechanism by which the transition disciplines their pricing. I do not think that is a contradiction. It is what it looks like when demand migrates toward you faster than your unit of sale adapts, and the flat revenue line is the two forces cancelling, for now, to the dollar.If the work is going to be eaten, decide who eats itThe decomposition is a diagnostic, and diagnostics are only worth running if you are prepared to act on an ugly answer. Suppose the exercise comes back the way I suspect it does at most firms, with net-new work a modest slice and the bulk of the AI book relabeled or defensive.The options at that point are narrower than the conference circuit implies. Refusing the repricing is not among them, because the client’s alternatives improve every quarter whether you participate or not, and a firm that holds traditional pricing to protect the old book is donating the transition to whichever competitor quotes the machine-adjusted price first.The posture I would argue for is deliberate self-cannibalization, run on the firm’s calendar instead of the client’s. Go to your best accounts before they come to you. Reprice the work the models have visibly deflated and name the deflation out loud, then spend the credibility that candor buys on selling upward, into the judgment-heavy engagements where the fee has not moved and will not soon.This trade costs real margin in the current year, which is why most firms will not make it voluntarily. What it buys is the relationship, intact, on the other side. A client who catches a firm charging analyst prices for machine output does not renegotiate that one line item.They start doubting the invoices generally, the honest ones included, and doubt is the one thing in this business that compounds faster than revenue.IBM, whatever else its flat quarter means, is at least running the transition at full speed rather than defending the old book.Four-fifths of a $9.5 billion book did not accumulate by accident. Somebody chose, engagement by engagement, to sell the cannibalizing work themselves. That choice is available at every scale, and the flat revenue line is what it costs. The alternative costs more and arrives later, with a competitor’s logo on it.The Google Cloud Practice is the quieter tellOne more detail from the same quarter, easy to miss under the headline numbers. IBM created a Google Cloud Practice inside IBM Consulting, pairing its Consulting Advantage AI delivery platform with Google’s Gemini Enterprise. On its face, this is routine alliance housekeeping, the kind of announcement firms make constantly.In the context of a flat print and a $9.5 billion book, I read it as a statement about where IBM believes the next tranche of signings comes from: not from its own origination, but from attaching its delivery capacity to a platform’s distribution.A practice named after someone else’s cloud is a channel, and channels are what firms build when they conclude that demand will be routed by the platforms rather than won in the open. That is a rational conclusion, possibly the correct one, and it comes with a cost structure worth naming. Work that arrives through a platform’s motion is work the platform can eventually redirect, and the fee on it will always carry an implicit toll.The firm becomes very good at delivering into one vendor’s stack and correspondingly less credible as an independent voice about whether that stack was the right answer. For a firm of IBM’s scale, running many such practices at once, the exposure diversifies.A smaller firm copying the move with a single alliance should understand it is trading margin and independence for pipeline, and should price the trade honestly rather than discovering it at renewal.The patient reading, and why I would not wait on itI want to be fair to the reading I have been arguing against. Signed AI work converts to revenue over years, not quarters. A book built fast in the early innings of a technology shift will always run ahead of the revenue it eventually produces, and IBM’s flat line may be the trough of a J-curve, with the growth arriving in 2027 and 2028 as the programs mature and expand.One quarter of one firm proves nothing on its own, and I hold my own conclusion here more loosely than the confident tone of this essay probably suggests. If consulting growth reappears within a few quarters while the book keeps compounding, the cannibalization story weakens, and the patient story wins.I would also want to see the composition of the new signings shift, with the AI-native engagements spawning successors rather than replacing predecessors, because a book that reproduces is the one signature of net demand that cannot be faked by relabeling.But a partner running a firm cannot price her year on a J-curve she cannot verify, and the useful part of the IBM print does not depend on which reading wins. The useful part is the discipline it forces.Gross AI signings are the most intoxicating number in professional services right now, quoted at every partner meeting, and they are the wrong number to steer by, because they contain the relabeled and the defensive work indistinguishably mixed with the new.The steering number is net: what your total book does after the AI work finishes eating whatever it is eating. IBM is big enough that its net is printed in public, to the dollar, and the dollar was zero. Your firm’s version of that number exists too.The only question is whether you compute it yourself this quarter, or let the income statement compute it for you a year from now, the way it just did for the largest AI services business in the world.This story is published on Generative AI. Connect with us on LinkedIn and follow Zeniteq to stay in the loop with the latest AI stories.Subscribe to our newsletter and YouTube channel to stay updated with the latest news and updates on generative AI. Let’s shape the future of AI together!IBM sold $9.5 billion of AI, and its consulting arm did not grow a dollar was originally published in Generative AI on Medium, where people are continuing the conversation by highlighting and responding to this story.

Source: Generative AI Pub — Published August 28, 2026 at 02:49 UTC — Category: Image AI