IBM Built 50,000 AI Agents — Then Kept Just 6,000
By Meta, TechPulse USA
IBM built 50,000 AI agents and then threw away 88 percent of them. That is not a failure story — or at least, it is not one IBM Consulting's boss wants you to read that way. The culling, according to Mohamad Ali, the division's senior vice president, is precisely the point: the consultancy spent months deploying digital workers across its own operations, discovered that most of them were dead weight, and now wants to sell the winnowed survivors — about 6,000 agents woven into more than 150 workflows — to everyone else as the "human-digital" model of consulting.
It is a rare thing in this industry: an admission of bloat, packaged as a product. The Financial Express put Ali's interview online at 10:13 on a Sunday night — not the hour you choose unless you want the story moving before Monday's news cycle — and the pitch is as brazen as it is interesting. IBM's consulting arm re-engineered itself with AI agents first, so the story goes, and it will now re-engineer you. "We've now deployed about 6,000 unique agents that are integrated into over 150 of our workflows," Ali told the Financial Express. The other 44,000, presumably, are no longer with us.
The 50,000-agent experiment nobody was supposed to notice
The most revealing detail is not the 6,000 agents IBM kept. It is the 50,000 it started with — a figure Ali disclosed to the US trade publication CRN, noting that the organisation brought in "some of IBM's finest engineers" to slash the number down. Read that again. IBM Consulting needed its best engineers to figure out which of its own AI agents were actually worth keeping.
This is the part of the AI agent boom that vendors would prefer to skip. The pitch is always scale: thousands of agents, deployed everywhere, doing everything. IBM's confession is that the first wave was mostly filler — agents built because building them was easy, then quietly retired when somebody checked whether they made anything better. If the consultancies themselves have to audit their own agent armies with their finest engineers, imagine what is sitting unused inside the enterprises they are about to sell to.
We read both interviews in full — Ali's Financial Express sit-down and the CRN conversation — and cross-checked every figure against IBM's second-quarter results. What survives the winnowing is a coherent, if self-serving, philosophy: the workflow matters more than the agent.
What the 6,000 survivors actually do all day
The remaining agents are not chatbots with delusions of grandeur. They are embedded in the unglamorous plumbing of consulting work — and, increasingly, in client operations across finance, human resources, and banking, where IBM says it is seeing the strongest traction in transformation projects.
The cybersecurity example is the sharpest. A threat investigation that used to take a human analyst 45 minutes can now be run by a cluster of AI agents in about two minutes, followed by ten minutes of human review, Ali told the Financial Express. The same team, he said, can now handle three times as many threats without adding headcount. In the CRN interview, he pointed to the same idea under a different name: the Autonomous Threat Operations Machine (ATOM), a system where dozens of agent "digital workers" sift low-level threats at machine speed before a human reviews their work — finishing investigations in a third of the time a fully human-led team would take.
Behind the agents sits IBM Consulting Advantage (ICA), the platform IBM built to run its own AI operations. It has a dedicated research and development team of more than 1,000 engineers, runs over 2,000 live projects globally, and has more than 500 client deployments that combine what Ali calls "digital and physical labour" — a telling phrase, the human-digital team in miniature. About 30 percent of IBM Consulting's largest projects now run on humans plus digital labour, he told CRN, adding: "I don't think any of our competitors have that and have that running at this massive scale."
The Bangalore lab that kills bad agents
The least theatrical — and most convincing — part of the story is in Bangalore. IBM's Productivity Measurement Lab, based in the city, tests workflows with agents and without agents before anything goes into production. It is, in effect, the place where bad agents go to die.
Ali's honesty here is worth quoting in full, because it cuts against everything the agent-hype industry says: "There are times where you build agents, put it into the workflow, and interestingly enough, you'll get negative productivity," he told the Financial Express. Negative productivity. Not diminished returns, not delayed payback — actual harm. An agent that makes the work worse is worse than no agent at all, and IBM has now measured that happening to itself.
This is the genuinely interesting intellectual move in IBM's pitch. The measure of success, Ali said, "is not the number of agents deployed or the volume of AI tokens consumed, but whether their use translates into productivity gains and better financial outcomes." Which raises an awkward question about the 6,000 number itself: if agent count is the wrong metric, why is it the headline? Because even IBM knows that in this market, numbers open doors that methodology cannot.
An "uber harness" for everybody else's AI
The platform story is where IBM's positioning gets shrewd. Ali calls IBM Consulting Advantage an "uber harness" that lets clients bring their preferred AI vendors and technology stacks — OpenAI, Anthropic, Google's Gemini, Microsoft's Azure AI Foundry stack, AWS Bedrock — rather than forcing IBM's own models on them. It is vendor-neutral consulting as a moat: IBM rents the models and keeps the workflow know-how.
"The bigger that pool of IP, the harder it is for another consulting company to catch up," Ali told CRN. "Which then makes us an incredible channel partner to the OpenAIs, the Anthropics, the Geminis... This is the moat we're building."
The OpenAI partnership announced on August 13 fits the pattern: GPT-5.6 and other frontier models embedded into the platform, plus a dedicated OpenAI practice and forward-deployed units of engineers certified to work inside client environments. The "forward deployed" units — mixed teams of architects, domain specialists, and AI engineers who go into a client, find the broken workflow, redesign it, and build the agent-based fix — are the human half of the human-digital model. Ali also said IBM has partnered with Pearson to verify what its agents can actually do and credential them, a quiet acknowledgment that clients have started asking the obvious question: how do I know your digital worker isn't a hallucination with a job title?
For readers new to the term, we covered what AI agents actually do in a recent explainer — the short version is that they act on your behalf inside software, and the difference between a useful one and an expensive toy is exactly the kind of workflow plumbing IBM is describing.
The next frontier: a price list for digital workers
Here is where the story turns genuinely disruptive. IBM Consulting is, Ali told CRN, "still working out the right licensing and pricing mechanism" for clients who want to buy digital workers. "That is the next frontier," he said. The experiment on the table: rate-card pricing — renting an AI worker for a period of time, the way you would a contractor.
Let that land. The industry that invented the billable hour is now pricing a competing unit of labour. IBM is also moving more of its own business onto fixed-price engagements, Ali said, because its human-digital teams are confident enough to guarantee completion — and it will take outcome-based pricing where clients want it. This is not cost-plus consulting anymore; it is the consultancy betting against its own hours.
The midmarket angle sharpens the point. Ali told CRN that IBM Consulting's midmarket customer segment has been growing precisely because AI agents let smaller teams do more work faster — which is to say, digital labour is pulling enterprise-grade consulting downmarket, where the old hourly model never fit. "IBM Consulting has always been known for high quality — if you want to get a project done and done right, you go to IBM Consulting," he said. "Where we're going next is high quality at affordable costs."
Rate-card digital workers are real — and the billable hour should be scared
So is this a genuine business model or consultancy theatre? Theatre would be easy to dismiss: a slide deck, a demo, a "digital worker" that is really a junior consultant with a chatbot. But the numbers underneath suggest something harder to wave away. IBM Consulting did $5.3 billion in the second quarter (the three months ended June 30), up 1 percent excluding foreign exchange — hardly spectacular, but signings grew 6 percent year on year for the second quarter running, profit margin expanded 160 basis points, gross margin hit 28.9 percent, and segment profit margin reached 12.1 percent. IBM expects consulting revenue growth to accelerate to low-to-mid single digits for the year. That is the financial profile of a business that is already getting leverage out of its agents, not merely talking about them.
And the threat model is obvious once you see it. If IBM — or Accenture, which has its own AI alliance with Anthropic — can rent you a credentialed digital worker by the hour while guaranteeing the outcome, the remaining pure-hours consultancies are selling fax machines. The honest skepticism is about the credentialing, not the economics: Pearson stamps are only as good as what they test, and "we brought our finest engineers to cull the bad agents" is a confession that quality control here is still manual and expensive. Scale that to every enterprise workflow and the moat starts to look like a staffing problem in new clothes.
Still: the consultancies eating their own cooking is exactly what the industry needed. IBM's most quoted line from this whole saga should not be about 6,000 agents. It should be the one from Bangalore: sometimes AI makes productivity go backwards. A vendor that leads with that sentence is one you can do business with. A vendor that leads with 50,000 agents — well, IBM tried that too. It kept 6,000.
The governance question, incidentally, is running parallel to all of this: IBM's own research recently found that eight in ten companies now escalate AI decisions to the boardroom. The agents are arriving whether boards are ready or not. The only question is who gets to price them.
IBM AI agents FAQ: the 50,000-to-6,000 questions people actually ask
What is a "digital worker"? Ali's own definition to CRN: anything that calls an LLM — "anything that's nondeterministic" — embedded in a specific workflow, not a general chatbot.
Why did IBM throw away 44,000 agents? The first wave was built because building was easy; IBM brought in "some of IBM's finest engineers" to triage and QA them down to 6,000 production agents (CRN).
What are the 155 packs? Ali says the 6,000 agents are organised into 155 packs by division — technology transformation, SAP applications, hybrid cloud, AMS, BPO, cybersecurity — so a client project gets the pack (plus human consultants) as a unit (CRN).
Is rate-card pricing live? Not yet: Ali told CRN IBM is "still working out the right licensing and pricing mechanism" — it is "the next frontier", not a product.
Sources: Mohamad Ali, senior vice president of IBM Consulting, in interviews with the Financial Express (published Oct 18, 2026) and CRN; IBM second-quarter 2026 results as reported by CRN. Relay coverage: PakTechStan.
About TechPulse USA: TechPulse USA is an independent technology publication covering the stories shaping the future — artificial intelligence, robotics, and the companies racing to build them. Our reporting is researched from primary sources and reputable outlets, and every fact we print is verified before publication.
Comments
Post a Comment