The workflow is the unit of design
An executive considering AI agents might begin with tasks such as classifying a request, retrieving a record, drafting a response or updating a system. Each looks like a candidate for automation. For the customer, however, those steps matter through the result they produce: whether the organization resolves the problem, honors its commitments and brings in a person when judgment is needed.
That distinction matters when software can act across several systems. An agent may use tools and pursue a task through multiple steps rather than simply produce an answer for an employee to consider. Even within a narrow assignment, its actions can carry one team’s assumptions into another team’s work.
The workflow is a strong starting point when deciding how much authority to give an agent. Goals, data, permissions and handoffs need to support the same outcome. If they conflict, an agent could complete every assigned step while making the overall result less reliable. Leaders therefore need to establish what must be true of the surrounding work before automating a consequential part of it.
How local success can become system failure
Consider a hypothetical customer-resolution workflow. A service agent is instructed to close requests quickly. It reads a customer record, offers a remedy within its permitted limit and marks the case resolved. A retention team measures success differently: preserving the relationship may warrant a larger exception. Finance requires approval for that exception. Meanwhile, the record available to the agent omits a promise made in a previous conversation.
The agent could follow its instructions and still leave the organization with a problem. It might close a case retention would keep open, offer a remedy inconsistent with the earlier promise, or initiate a request no one has authority to approve. The immediate output may be accurate relative to the data and rules supplied. The failure lies in the relationship among those rules, the missing information and the decision rights at the handoff.
Donella H. Meadows examines relationships within systems in Thinking in Systems: A Primer; her essay on leverage points identifies information flows, rules and goals as different places to intervene. Applied to the service example, that distinction separates the missing customer promise, the approval rule and the definition of a resolved case. Fritjof Capra and Pier Luigi Luisi's The Systems View of Life likewise emphasizes relationships, patterns and context. These systems perspectives help frame the organizational problem without supplying evidence about the performance of AI agents themselves.
The risk is that a discrepancy travels with the work. An agent uses the objective, information, tools and permissions available to it, then passes its output to a function operating under different assumptions. If no one can question an ambiguity or settle the trade-off, faster completion of the first task may leave more rework for the next team. The hypothetical workflow makes that possibility visible; testing it requires observing what actually happens at the handoffs.
Anthropic distinguishes predefined workflows from agents that direct more of their own tool use, and recommends starting with the simplest workable approach while monitoring feedback and compounding errors. That is vendor guidance for designing a trial, rather than independent proof of a business result. The National Institute of Standards and Technology's AI Risk Management Framework complements it with attention to tasks, context, documented responsibilities, human oversight and repeated assessment. Neither removes the need to test whether the chosen controls work in this particular workflow.
The practical implication is proportionate preparation. As an agent gains the ability to make consequential changes, the organization needs clearer authority and better ways to detect and correct mistakes. A bounded experiment can begin before every process is redesigned, provided its limits are explicit.
The Agentic Coherence Test
The proposed OutcomesLab review is the Agentic Coherence Test. Its five questions help leaders decide where an agent belongs and what needs to change before its authority expands. It is a design aid for a particular workflow, not a certification of readiness.
1. What outcome must the workflow deliver? Define success from the recipient’s perspective, including the commitments that must be honored. Closing a case is not a durable resolution if the customer has to return. Where speed, retention and cost conflict, leaders must set a decision rule rather than leave the compromise for an agent to infer.
2. Where does the work—and its disagreement—travel? Trace a request from entry to final disposition, including the exceptions people negotiate today. Identify the record used at each handoff and who can resolve a conflict. Service, retention and finance may all use “resolved” while meaning different things. At OutcomesLab, we call these unresolved differences Alignment Debt. A trial should expose where they affect decisions rather than hide them behind a tidy process diagram.
3. Which capabilities can be separated from decisions? Retrieval, classification, drafting and record updates need not carry the same authority. An agent can draft a remedy without issuing it. Separating those steps allows a bounded capability to be tested and replaced while a person retains the consequential decision. Success at one step is not proof that the workflow works.
4. Who may read, recommend, change and commit? Specify those permissions separately. A proposed response and an issued refund are different acts, even if both appear on one screen. Set a threshold for human review when records conflict, a commitment is unclear or a decision exceeds the agent’s authority. Name the owner who can resolve the case—not merely the inbox that receives an escalation. The appropriate level of review should reflect consequence and reversibility: correcting a draft is different from reversing a customer-facing commitment.
5. How will the organization learn and correct course? Before expanding autonomy, decide how an incorrect action will be recognized, who can stop or reverse it and what evidence would change the rule. In the hypothetical service workflow, monitor reopened cases, corrected remedies and downstream exceptions alongside cases closed. Test ambiguous records and conflicting instructions before deployment, then watch for them during operation. If the workflow or its priorities change, revisit permissions and measures rather than assuming yesterday’s design still fits.
A further management question follows from Andrew S. Grove's High Output Management: what does the automated step enable other people to get done? Applied to AI, a faster case-closing step that creates more corrective work has shifted effort rather than improved the result.
Why experiments should not wait for perfect coherence
There is a serious objection to this approach. Organizations seldom have a complete, agreed-upon account of how work happens. Requiring one before any AI use could turn a modest trial into a prolonged architecture project. A bounded agent might reveal an unclear handoff sooner than an exhaustive mapping exercise. And a summary drafted for human review need not depend on every cross-functional decision right being settled.
There is evidence that useful AI assistance need not wait for comprehensive redesign. In a customer-support firm, Erik Brynjolfsson, Danielle Li and Lindsey Raymond found productivity gains after workers received generative AI assistance, particularly among less-experienced workers. Their study concerned assistance to people, so it does not establish that autonomous agents can reconcile conflicting goals across functions. It does support taking bounded opportunities seriously rather than assuming organizational imperfections make every trial futile.
The useful boundary is the consequence of the action. If an agent produces an inspectable draft or a reversible recommendation, leaders can learn from a small trial without settling every organizational disagreement first. If it changes an authoritative record, commits resources or acts across functions, the trial needs explicit permissions, a correction path and an accountable owner. Apply the coherence test in proportion to that scope. Its purpose is to make experimentation informative, not to make certainty a prerequisite.
Experiments can contribute to redesign. If an agent repeatedly encounters conflicting customer records, leaders can ask which record governs which decision. If the same exception is escalated every day, the organization may need a clearer rule rather than a faster escalation queue. Neither observation proves that the agent caused the underlying conflict. The trial becomes valuable when people use what it reveals to improve the workflow.
Build decision velocity, not just task speed
A trial earns broader authority when its results support that next permission. Examine corrections and downstream work alongside completed tasks, and resolve any conflict that would become more consequential at greater scope. The decision is what the agent may do next, supported by evidence from the workflow it already serves.
That sequence protects Decision Velocity—the ability to make sound choices and carry them through—rather than equating speed with throughput. It also offers a check on Strategic Entropy. Priorities may change while old instructions remain embedded in tools, measures and permissions. An agent could keep executing an obsolete choice consistently. Periodic review should therefore ask two questions: Is the system doing what we designed it to do? and Does that design still express what we intend?
Before extending the trial, ask what the next permission would let the agent change and whether the organization could recognize and correct a mistaken action. Resolve the relevant ambiguity, give someone authority over the remaining exceptions, and check the effect on the whole workflow. As agents take on more of the work, those decisions will need revisiting whenever priorities, information or customer commitments change.