Building AI Agents Is the Easy Part. Operating Them Is the Enterprise Problem.

Building an AI agent is becoming remarkably easy. Operating one inside a business is not.

Sharjeel Ahmad made the same distinction recently from the perspective of deploying agents in banking: a demo can prove that an agent can perform a task, but not that it can do so safely, repeatedly and under clear accountability. That gap is the operating model.

That distinction is getting lost because the visible part of an agent is the impressive part. It can plan, call a tool and complete a task in a demo. The invisible work begins when the same agent has to run every day, for different people, against changing systems, without quietly spending too much money or doing something it was never authorised to do.

IBM has started calling the collection of agents, models, tools and workflows across an organisation its “agentic estate”. That is a useful phrase, largely because most enterprises will discover they have one before they remember deciding to build it.

The agentic estate grows one sensible experiment at a time

Nobody sets out to create agent sprawl. A team solves a small problem. Another team connects a model to a different system. A vendor adds an agent to a product the company already uses. Someone publishes an MCP server because integration suddenly looks easy.

Each decision can be perfectly reasonable on its own. Together they create a distributed software estate with credentials, permissions, data access, model bills and failure modes that nobody sees as one system.

This is shadow AI in its more consequential form. The risk is not only an employee using an unapproved chatbot. It is useful AI becoming operational before the organisation has worked out how to operate it.

A catalogue is not an operating model

Discovering which agents exist matters. So do inventories, owners and risk classifications. But knowing that an agent exists does not tell you whether its last run succeeded, which tools it called, what it cost or whether it stopped for the right reason. An inventory describes the agentic estate. It does not operate it.

IBM’s “agentic control plane” language addresses this operational gap: central visibility, runtime controls, credential monitoring, traces, scheduling and reuse across agents built in different frameworks.

That is an important enterprise layer. It still does not remove the need for orchestration inside the work itself. A control plane can tell you an agent ran. The orchestration layer must know what the workflow was trying to accomplish, what state it reached, which action was valid and when a person needed to take over.

The hard work is mostly around the model

When we built ProdPad Conductor, the difficult parts were not the model calls. They were cleaning tool arguments, validating structured responses, preserving workflow state, reconciling actions taken elsewhere in ProdPad, limiting which tools were visible and deciding whether a failure should be repaired, retried or returned to the user.

None of that produces a dramatic demo. All of it determines whether the system can be trusted with real work.

This is why the industry’s fixation on agent builders is slightly backwards. Construction is becoming a feature. Operation is becoming the architecture.

Production agents need boundaries and evidence

An enterprise agent needs more than a goal. It needs:

• an owner and a defined business outcome
• constrained access to tools and data
• durable state that survives a conversation
• traces showing what actually happened
• evaluations that distinguish fluent output from successful work
• approval and escalation points
• cost, latency and failure limits
• a way to replace the model without rebuilding the workflow

Without those things, autonomy simply makes the uncertainty run faster.

ProdPad gives Conductor a difficult but credible proving ground

Product management is not a neat back-office process. Requests are ambiguous, evidence is connected, priorities change, and many decisions require judgement rather than a mechanical rule.

That makes ProdPad a much better test of orchestration than another travel-booking demo. ProdPad Conductor has to coordinate models, deterministic software, product data and people while preserving the state of work that may continue over days or weeks.

The enterprise question is therefore no longer “Can we build an agent?” It is “Can we operate an agentic estate without losing control of cost, authority and learning?”

The agent is the visible feature. The operating system around its work is where enterprise reliability is created.