Every time companies seem to figure out one aspect of deploying AI agents, another wrinkle appears. The agents have to work within existing — often complicated — applications, business rules and processes that don’t stand still once the agents are deployed.
That's become clearer to IT leaders as they've deployed agents in production. They're learning how much the surrounding systems and workflows matter to what AI can actually accomplish.
Meta recently ran into one such complication. The company had developed an AI agent to manage data-column onboarding across its recommendation systems, a process involving several subsystems. The agent had access to instructions and knowledge from each of them. But in production, it began applying rules meant for one subsystem to another.
Meta responded by dividing the work among 14 specialized agents, with a small, deterministic orchestrator handling routing, sequencing and reliability checks. The company said the redesign reduced preparation for a new data-column configuration from three or four days of manual work to less than an hour.
The experience illustrates something enterprises are learning as they put agents to work: The intelligence inside an agent is only part of what determines whether the work succeeds or fails.
Making the agents work together
A recent OutSystems survey conducted with CIO Dive and KPMG offers another perspective. Of 858 executives surveyed, 82% expressed confidence that their organizations could keep pace with governing, integrating and monitoring agents. Asked about resolving conflicts across agents, systems and workflows, only 26% said they were well prepared.
What happens when the controls that work for one agent aren't enough for several agents doing connected work? For OutSystems, a low-code application development platform provider, that means keeping “orchestration in code, not in agents," Tiago Azevedo, the company's CIO, told TechTarget.
Individual agents are given narrow tasks, Azevedo said. The surrounding code establishes which options are available at each step, even though an agent can decide how to proceed within those limits.
Authority shouldn't sit with the agents themselves.
Tiago Azevedo
CIO at OutSystems
Identity and permission checks have to apply whenever an agent acts or calls another system. High-risk actions need human review, and somebody must be able to stop the work, investigate what happened or reverse it.
"Authority shouldn't sit with the agents themselves," Azevedo noted.
OutSystems hasn't encountered conflicts between agents in the internal operations Azevedo described. He said the company takes a preventive approach, rather than documenting a conflict it has to resolve.
The rules around AI need attention, too
Other enterprises are arriving at variations of the same approach, depending on the work they're asking AI to do.
Australian mortgage provider Lendi Group built an orchestration platform, Aurora, combining its own technology with open source and commercial components. Risk and compliance requirements are built into that platform and can be updated as those requirements change.
"We don't let models determine the rules," CTO Devesh Maheshwari told Computer Weekly. Agents can gather information and communicate with customers, but consequential advice remains with brokers. The company also tests agent behavior in production and has people review exceptions.
Trust Bank Singapore learned a different lesson. Its automated runbooks — predefined procedures for investigating IT problems — became outdated as its systems changed. A subsequent attempt to let AI analyze the data without that structure produced inconsistent results. The bank now uses prescribed runbooks to gather evidence, AI to analyze it and engineers to decide what action to take.
When the runbook is incorrect, CTO Srinivas Patil warned, "it doesn't matter how great the AI is, the answer is always going to be wrong." The bank tests its runbooks against reference data to check whether they're still looking for the right signals.
These aren't identical solutions. Lendi maintains business policies within a platform it controls; Trust Bank combines a managed agent platform, prescribed procedures and engineers' judgment. Meta reorganized where agent reasoning occurs.
Their common lesson is that the systems surrounding AI need to evolve along with the work.
When agents expose existing problems in the workflow
Even when a company knows a process is complicated, it might not understand all the dependencies involved. That matters when agents start taking on the work.
Barclays has that challenge in its corporate customer due diligence process. It can take months, crossing systems and teams with plenty of manual information gathering. The bank is developing specialized agents to help with parts of that work, using deterministic orchestration across the process. Knowing the complexity gives it something to plan around, although it doesn't guarantee every dependency has been discovered.
Other dependencies can be less obvious. EY's global vice chair Julie Linn Teigland pointed out to TechTarget that a change to a CRM record could affect commissions, billing, sales forecasts and potentially HR systems. Agents might communicate and complete their own assignments while the broader business process remains inadequately coordinated.
They could also act on incomplete understanding and make the consequences worse before anyone notices.
That's an old enterprise problem. Integrations, business rules and informal employee workarounds often aren't as clean as architecture diagrams might suggest. Agents could expose those gaps while organizations prepare to deploy them. They could also act on incomplete understanding and make the consequences worse before anyone notices.
Which brings us back to experience. Deploying agents isn't something IT can do once and walk away from. It's an ongoing conversation between IT and the business, informed by what agents actually do in production and what needs to change as a result.
There could be an unexpected benefit. Organizations introduce agents hoping to automate tasks, reduce delays and save money. If the effort also forces them to understand and improve processes they never fully understood, AI could improve the enterprise in ways that go beyond the automation that started it all.
James Alan Miller is a veteran technology editor and writer and lead editor for CIO news at Informa TechTarget. He directs coverage of enterprise technology strategy, AI, software, data, infrastructure and the decisions shaping how CIOs manage increasingly complex IT environments.
