Published 21 September 2026
Some technology projects that once required large teams of consultants may increasingly be handled by small groups of specialists working with AI agents.
Earlier this year, IBM introduced Forward Deployed Units, or FDUs, small teams of engineers and experts who work directly with clients to build and deploy AI systems in real-world environments. The company said the units put select teams of specialists close to clients and use AI agents to handle parts of the work. Similar approaches are spreading across the industry as technology companies and consulting firms embed smaller, AI-assisted teams with customers.
Companies are still struggling to turn AI experiments into systems that work across their businesses. Gartner said in a recent survey that only 22% of organizations had scaled AI across multiple business units or adopted an AI-first approach, even though 85% of functional leaders said they planned to increase AI spending in 2026. At the same time, greater spend doesn’t necessarily yield greater payoff. A new IBM Institute for Business Value (IBV) study reported an average AI ROI of 17% for enterprises in 2026.
“The market sentiment has changed from, ‘What can be done with AI?’ to actually, ‘Show me the dividend of your AI investments,’” Javier Olaizola, Global Managing Partner for Hybrid Cloud and Data at IBM Consulting, told IBM Think in an interview.
Clients have moved past “marginal experimentation,” he said. “There is a lot of engineering that needs to be done to get to the outcome, to get to the dividend.”
IBM designed FDUs as small groups that work directly with clients and delegate parts of a project to AI agents, Olaizola said.
A typical unit can include three or four people, he said: someone responsible for the business result, a technical architect or forward-deployed engineer, a data specialist and an industry expert, with 10-20 agents work alongside them.
Those agents might write software requirements, generate code, create test cases, scan software for defects or search for patterns in existing procedures, Olaizola said.
Humans remain at both ends of the process. They define the problem, supply the business and technical context, and decide whether the output meets the client’s requirements, he said.
“What makes them special is basically two things,” Olaizola said. “They deliver on the client’s outcome, but at the same time, they are constantly learning from what they are executing.”
The teams can then delegate more work to the agents as they learn which tasks the systems can handle reliably, he said.
As technology companies try to move technical staff closer to customers, forward-deployed engineers have become more common across the AI industry. At the May launch, IBM experts described the role as a combination of engineering, consulting and business expertise, with specialists building systems directly in the environments where clients will use them.
Amazon Web Services announced a similar push in June, with Reuters reporting that AWS plans to commit USD 1 billion to a new organization built around forward-deployed AI engineers who will work intensively with customers on production systems.
IBM said its own units used IBM Consulting Advantage, a platform that provides agents, reusable methods and other tools for consulting work.
Olaizola pointed to an ongoing data modernization project for a large bank in India to show how the model could change the shape and pace of delivery.
The bank, which Olaizola said IBM did not identify publicly, needed to move hundreds or thousands of data processing jobs onto newer technology, he said.
A conventional approach to a migration of that scale would typically require roughly 300 people working for 12 to 15 months, according to Olaizola. IBM was able to carry out the project on a timeline of less than nine months with a fraction of that team, he said, because the agents absorbed most of the repetitive conversion and testing work.
Instead of having people manually correct every bad result, a team would study the errors and change the instructions or context it gave the system. It would then run the work again and measure whether the results improved.
Olaizola described a hypothetical sequence in which AI agents converting legacy code might initially get about 70% of the work right, then reach 85% and 90% as the team refined the inputs.
Once the process worked on a small batch of legacy code migration jobs, a team could apply the same approach to hundreds or thousands of similar jobs, he said.
“It’s about working on the inputs,” Olaizola said.
The approach could also affect how consulting firms charge for projects.
Traditional consulting contracts often tie revenue to the number of people assigned to a project and the hours they work. Olaizola said a model that requires fewer people and less time makes that structure harder to justify.
“If you are pricing by the hour, the incentive for anybody is basically to maximize the hours, not the outcome,” he said.
Clients have already started pushing consulting firms in another direction. The Financial Times recently reported that companies are pressing consultants for lower fees, fixed-price projects and contracts linked more closely to performance as AI automates some technology work.
As agents take on more routine work, Olaizola said, human expertise will shift toward evaluating the results.
AI can handle parts of jobs traditionally performed by developers, architects and analysts, he said. People still have to determine whether the result fits a client’s architecture, meets security requirements and complies with regulations.
Software can function properly and still be unusable if it violates a bank’s security rules or relies on data the company cannot use, Olaizola offered as an example.
“The value basically has moved into a judgment layer,” he said.
He compared the change with journalism. AI can produce large amounts of text quickly, he said, but a person still has to decide whether the material is accurate, useful and appropriate for the audience.
“It’s not about producing the content,” he said. The value comes from journalists providing “the assurance that this is giving the right perspective.”
The model also raises questions about how consulting firms will train less experienced workers.
Large professional services firms have traditionally used junior employees to perform research, analysis and technical work under more experienced colleagues. IBM said when it launched FDUs that it planned to recruit graduates and develop them for forward-deployed roles.
The company did not say the model would eliminate traditional consulting jobs.
Instead, Olaizola explained, the model will change the type of work people do as agents take on more of the production.
Rather than spending their first years on the job doing the routine tasks that agents can now perform, less experienced consultants will develop inside the units from the start, he said, working next to senior colleagues as they frame problems, evaluate agent output and make judgment calls. Building that judgment will become the purpose of development, not a by-product of it.
Olaizola does not expect every business process to rely on AI agents.
Some tasks follow fixed rules, with known inputs and predictable outputs. Conventional software often handles those jobs more efficiently than a system making probabilistic decisions, he said.
Work involving interpretation, flexibility or judgment offers a better fit for agents, according to Olaizola.
The distinction comes as companies continue to spend heavily on AI without always succeeding at broad deployment. Gartner based its September 1 findings on a survey of 1,303 respondents at organizations with at least USD 50 million in annual revenue and found that successful scaling remained limited.
Olaizola said companies need more than capable AI models. They also need people who understand business problems, know where agents can help and can decide whether the output is good enough to use.
For consulting firms, he said, the change could reduce the importance of team size while increasing the value of judgment and accountability.
“AI will be embedded everywhere,” Olaizola said. “Not all the workloads will be agent-based.”
Get curated insights on the most important—and intriguing—AI news. Subscribe to our weekly Think newsletter. See the IBM Privacy Statement.