For the first time, the most important question about enterprise AI is becoming less about what the model can say and more about what the system is allowed to do.

That change sounds subtle, but it is not. A chatbot can draft a summary and wait. An AI agent can search connected sources, use tools, create a file, update a record, or prepare a decision for approval. OpenAI described this transition on 12 August 2026 as a shift “from assistance to execution,” based on its enterprise usage research. NIST, meanwhile, launched an AI Agent Standards Initiative in February 2026 because autonomous action creates a need for security, interoperability, and trust.

The technology conversation often focuses on models. The operational conversation must focus on work: What starts the task? Which data may be used? What decisions can be delegated? Where must a person intervene? What evidence should remain after the task is completed?

This is why I believe Business Analysts are not becoming less relevant in the agent era. Their work is moving closer to the control layer of the enterprise.

A process does not become intelligent by adding an agent

Imagine an organization wants an agent to handle supplier invoice exceptions. The exciting demo might show the agent reading an invoice, checking a purchase order, identifying a mismatch, and drafting a response. But the real process is larger than that happy path.

What if the purchase order was amended after the delivery? Which system is authoritative? Can the agent see bank details? Is a two-percent price difference acceptable for every category? Who may approve a payment hold? What happens if the invoice is duplicated across two legal entities? How does an auditor reconstruct the decision six months later?

None of those questions are primarily model questions. They are business rules, data ownership, access, exception management, and accountability questions.

An unclear manual process does not become clear when automated. It becomes a faster source of inconsistent outcomes. The agent may hide that inconsistency behind a smooth interface, which can make the risk harder to notice.

From requirements gathering to decision-boundary design

Traditional requirements work often describes screens, fields, integrations, and acceptance criteria. Those remain useful, but agentic systems add another object that must be designed: the decision boundary.

A practical agent workflow needs at least five explicit boundaries.

First, define the **goal boundary**. “Handle invoice exceptions” is too broad. “Classify an exception, collect supporting evidence, and recommend the next action without releasing payment” is testable and safer.

Second, define the **data boundary**. The agent should know which sources it may access, which source wins when records conflict, and which sensitive fields must be masked or excluded.

Third, define the **action boundary**. Reading a record, drafting a response, changing master data, and authorizing money movement are not equivalent permissions. Each action needs an owner and a risk level.

Fourth, define the **confidence boundary**. The system needs rules for stopping, escalating, or requesting human review. A confident answer is not the same as a verified answer.

Fifth, define the **evidence boundary**. A completed task should leave a useful trail: inputs consulted, tools used, rules applied, output produced, reviewer identity, and final disposition. NIST’s 2026 work on evaluation probes for agentic AI emphasizes the need for visibility into complex multi-step workflows and the evidence behind agent decisions.

This is familiar territory for Business Analysts. It combines as-is and to-be mapping, business rules, roles and permissions, exception paths, UAT, and change management. The difference is that the “user” executing part of the flow may now be a non-human system with variable behavior.

The best first use case is rarely the flashiest

My view is that an organization should not begin with the workflow that looks most impressive on stage. It should begin with a workflow where value and failure are both measurable.

A good first candidate has a repeatable trigger, accessible data, a limited action set, frequent enough volume to learn from, and a clear human owner. Examples might include compiling a daily operational exception brief, checking whether mandatory fields are complete before a case enters review, or gathering evidence for a service request.

These examples may appear modest. That is an advantage. Teams can compare the agent’s output with an existing baseline, observe its mistakes, refine controls, and decide whether autonomy should expand.

The wrong starting point is a vague ambition such as “automate operations with AI.” It gives the team no stable definition of done and makes almost every outcome defensible after the fact.

UAT must test behavior, not only expected answers

Agentic UAT cannot consist only of ten clean prompts that produce ten correct responses. Testing must explore what happens when the environment is messy.

Use duplicate documents. Remove a required field. Create conflicting values in two source systems. Revoke a permission midway through a task. Introduce an unsupported file format. Ask the agent to perform an action outside its mandate. Test whether it stops safely, explains the limitation in useful language, and routes the case to the right person.

The acceptance criteria should cover more than output accuracy. They should include authorization, traceability, escalation quality, recovery after tool failure, latency, operating cost, and the rate of unnecessary human intervention.

This is also where Product Managers need discipline. A roadmap should not measure progress by the number of agents launched. It should measure cycle time, rework, exception leakage, user trust, and business outcomes—without relaxing controls to make the metric look better.

Human review should be designed, not added as a disclaimer

“Human in the loop” is often treated as a universal safety answer. In practice, a tired employee clicking approve on hundreds of recommendations is not meaningful oversight.

The review step needs context. The reviewer should see why the case was escalated, which evidence matters, what changed since the last check, and what action will follow approval. High-risk decisions may require separation of duties; low-risk, reversible actions may need only sampling and monitoring.

The objective is not to keep a person inside every step forever. It is to place human judgment where its value is highest and to retain a clear route for intervention.

The durable advantage is operational clarity

AI models will continue to improve, and enterprise platforms will keep adding agent features. Gartner predicted in August 2025 that 40 percent of enterprise applications would include task-specific agents by the end of 2026, up from less than five percent at the time. The exact pace will vary, but the direction is clear.

What will distinguish successful organizations is not access to a button labelled “agent.” It will be their ability to describe work precisely, connect reliable data, assign authority, test exceptions, and learn from evidence.

That makes process knowledge more valuable, not less. The Business Analyst of the agent era is part translator, part workflow architect, and part control designer. The central question is no longer simply, “What should the system do?” It is, “Under what conditions may the system act—and how will we know it acted correctly?”

That is a harder question. It is also where responsible transformation begins.

**Discussion:** If your organization could delegate one process to an AI agent tomorrow, which decision would you refuse to automate until the controls were proven?

Sources

  • [OpenAI — From assistance to execution: How enterprises put AI to work (12 August 2026)](https://openai.com/index/how-enterprises-put-ai-to-work/)
  • [NIST — AI Agent Standards Initiative (17 February 2026)](https://www.nist.gov/artificial-intelligence/ai-agent-standards-initiative)
  • [NIST — Building Evaluation Probes into Agentic AI (7 April 2026)](https://www.nist.gov/programs-projects/building-evaluation-probes-agentic-ai)
  • [Gartner — 40% of enterprise apps expected to feature task-specific AI agents by end-2026 (26 August 2025)](https://www.gartner.com/en/newsroom/press-releases/2025-08-26-gartner-predicts-40-percent-of-enterprise-apps-will-feature-task-specific-ai-agents-by-2026-up-from-less-than-5-percent-in-2025)