A production agent needs a job, not a personality.
The strongest deployments do not begin with a request to “add AI.” They begin with a piece of work that already exists: research an account before a sales call, compare documents before a claim is reviewed, coordinate a customer onboarding, or prepare a weekly operating report. The agent is useful when it can move that work to a clear finish line and show what it did.
That is a narrower promise than a general AI employee. It is also far more likely to survive contact with production.
Production agents are usually bounded
IBM's 2026 study of production agents combined 20 case studies with a survey of 306 practitioners. The result is a useful correction to the market's autonomy narrative: 68% of production agents execute ten steps or fewer before human intervention. The same study found that 70% rely on prompting existing models rather than training their own, while reliability remains the leading technical problem.
The implication is practical. The model is rarely the whole product. The operating design around it matters more:
- what starts the job
- which context the agent receives
- which tools it may use
- which actions are reversible
- when a person must approve
- what happens when a tool fails
- how the final outcome is checked
An agent with a small role and strong operating design can be valuable. A powerful model with a vague role is still a risk.
Four kinds of work fit agents well
Revenue operations
Agents can research accounts, prepare meeting briefs, qualify inbound demand, draft follow up and keep CRM records current. The finish line is not “produce insight.” It is a prepared meeting, a routed opportunity or a completed record.
Customer operations
Onboarding, intake, case updates and service coordination often span an inbox, documents and an internal system. An agent can gather missing information and move routine cases forward while routing unusual situations to an owner.
Document operations
Claims, quotes, contracts and invoices contain repetitive reading and comparison work. An agent can extract fields, compare them against rules, prepare evidence and flag disagreement. Final approval can remain human where the consequence requires it.
Internal operations
Recurring reports, vendor requests, employee support and reconciliation work are often spread across tools that do not share one clean workflow. An agent can coordinate the steps, but only if its access is intentionally limited.
The finish line must be visible
“Help the operations team” is not a production role. “Read the submitted documents, validate the required fields and route incomplete cases to the claims queue” is.
A useful role definition has five parts:
| Part | Question |
|---|---|
| Trigger | What starts the work? |
| Inputs | Which data can the agent use? |
| Tools | Which systems can it read or change? |
| Completion | What observable result ends the task? |
| Exception | Who takes over when the boundary is reached? |
This creates something that can be tested. Without a finish line, teams end up evaluating whether an answer “looks intelligent” instead of whether the work was completed correctly.
Autonomy should be earned
Not every action needs the same control. A sensible release separates work into three levels:
- Prepare. The agent researches, drafts or organizes, but a person performs the action.
- Act with approval. The agent prepares the action and pauses at a named review point.
- Act within policy. The agent completes low consequence, reversible actions and records the result.
OpenAI's 2026 enterprise research describes the broader shift from chat based assistance to delegated work, but also emphasizes governance, measurement and production controls as firms scale it. Frontier firms are distinguished by deeper workflows, not just more messages.
The right target is not maximum autonomy. It is the smallest amount of autonomy that removes meaningful work without hiding risk.
What a production release should include
A production agent is more than its prompt. A responsible release includes:
- the role and operating boundary
- approved tool connections and permissions
- test cases for normal work and known failures
- human approval and handoff rules
- traces that show important decisions and tool calls
- an owner for incidents and future changes
If those pieces do not exist, the team has a prototype. It may be impressive, but it is not yet an operating system.
The market is moving quickly toward longer running agents that retain context, use tools and coordinate work. OpenAI's Agents API announcement reflects that direction. The discipline remains the same: define the work, constrain the environment and measure the result.
That is what lets an agent own a real part of the operation without asking the business to trust a black box.