Field Notes · Essay 02
An AI Agent Can Be Right and Still Not Be Authorized to Act
Why a correct recommendation is not the same as permission to execute it.
Imagine an AI-enabled system examining two customer records.
The names match. Contact information aligns. The available evidence suggests that both records represent the same person.
The system recommends merging them.
Now, for the sake of argument, assume the recommendation is completely correct.
Should the system be allowed to execute the merge?
Not necessarily.
This distinction has become increasingly important to me while building my Ontology-Grounded Agentic AI Data Steward.
We spend considerable time evaluating whether AI systems can make good decisions. But in enterprise environments, the correctness of a recommendation is only part of the problem.
A system can recommend the right action without having the authority to perform it.
And confusing those two things can have consequences that extend well beyond the AI model itself.
The same recommendation, different authority
Consider two employees reviewing the same potential customer-record merge.
One is a business analyst investigating duplicate records. The other is a data steward responsible for reviewing and approving changes to authoritative customer data.
Both examine the same evidence.
Both agree that the records represent the same person.
Yet their authority may be different.
The analyst might be permitted to investigate and recommend a resolution. The data steward might have permission to approve the merge, subject to the organization's policies.
The evidence hasn't changed. Neither has the conclusion.
What differs is who is permitted to do what.
Now introduce an AI agent into that workflow.
The agent may be capable of evaluating the evidence, explaining its recommendation, and identifying the next appropriate step.
But those capabilities do not automatically grant it permission to execute the change.
The same principle applies even when a human approves the recommendation.
Approval records a decision. Execution changes a system. Authorization must still be enforced at the point of execution.
A correct recommendation is not an authorization. And an approval is not a substitute for authorization enforcement.
Three questions, not one
I find it useful to separate three questions when thinking about consequential AI-assisted workflows.
First: Is the recommendation justified?
This concerns the evidence.
Are the records sufficiently similar? Are there conflicting signals? Is important information missing? How much confidence does the available evidence actually support?
A recommendation can be reasonable without being certain.
Second: Is the proposed action authorized?
This concerns permissions and business context.
Who is requesting the change? Which records are affected? Does the relevant policy permit that person or system to initiate or approve the operation? Are additional conditions required?
A recommendation's correctness doesn't answer these questions.
Third: Can the system executing the action independently enforce those conditions?
This concerns execution.
An AI-generated explanation, a confidence score, or a human clicking an approval button should not, by itself, bypass the controls of the system responsible for making the change.
These questions are related, but they are not interchangeable.
An action can be correctly recommended and still be unauthorized.
Conversely, an action may be properly authorized but based on an incorrect recommendation.
A trustworthy workflow needs to account for both possibilities.
Why the distinction matters in enterprise data
Merging two customer records may sound like a small operation.
In practice, an authoritative customer identity can be consumed by multiple downstream systems.
A merge might affect customer servicing, reporting, communications, or processes that rely on a consistent understanding of who a customer is.
An incorrect merge can propagate errors. Reversing the original database operation may not reverse every downstream consequence.
But even a correct merge could violate a business restriction, cross an organizational boundary, or bypass a required review.
Improving the AI model's accuracy does not eliminate those risks.
Neither does making its explanation more persuasive.
That is why I believe enterprise AI discussions need to distinguish between the quality of an AI recommendation and the legitimacy of the action that follows.
The two require different evidence and different controls.
What this means for the system I'm building
My independent Agentic AI Data Steward project explores how deterministic entity-resolution evidence, governed enterprise context, and bounded recommendations can support data stewardship.
One boundary is deliberate.
The current prototype can evaluate candidate matches and produce recommendations for review, but it has no authority to modify enterprise records.
That allows me to work on the evidence and recommendation problem without assuming that a useful recommendation should automatically become an executable action.
I can examine whether a candidate match is supported by evidence, whether conflicting or missing information is appropriately represented, and whether the recommendation provides useful context for review.
Determining what would authorize a future enterprise action is a separate concern.
I want to be precise about that distinction. The prototype's lack of mutation authority is not evidence that a complete enterprise authorization and execution framework has already been implemented or validated.
It is a boundary of the current system, with further engineering and validation required before greater autonomy would be justified.
That may appear less ambitious than allowing an agent to act immediately.
I see it as a way to develop capabilities without prematurely extending their authority.
Autonomy is a product decision, too
None of this means every AI action should require human approval.
That would introduce friction into workflows where automation might be both useful and appropriate.
A read-only recommendation is different from modifying a temporary working copy of a record. Both are different from changing an authoritative customer identity consumed by multiple enterprise systems.
The consequences differ, so the controls should differ too.
Some operations may be suitable for automation with validation, limited permissions, monitoring, and recovery mechanisms.
Others may require explicit approval or stronger independent safeguards.
The appropriate level of autonomy depends on the operating context, the consequences of failure, and the ability to enforce and recover from actions safely.
This is where the discussion becomes as much about product leadership as engineering.
Every additional approval can slow a workflow. Every additional permission can increase the potential consequences of misuse or error.
The challenge isn't to maximize autonomy or minimize it indiscriminately.
It's to determine which actions genuinely benefit from automation, what authority those actions require, and what evidence would justify granting it.
These decisions affect product scope, user experience, operational efficiency, risk, and accountability.
They deserve attention during product discovery and architecture design—not only when a system is approaching deployment.
The question I keep coming back to
When teams evaluate an agentic AI capability, the conversation often begins with what the system can do.
Can it identify a duplicate? Resolve an ambiguity? Recommend an action? Complete a workflow?
Those are important questions.
But I think another deserves equal attention:
Under what conditions should the system be permitted to act on what it knows?
A system may eventually demonstrate both reliable recommendations and appropriately governed execution.
But progress in one area does not establish progress in the other.
As I continue building the Data Steward, I want to preserve that distinction rather than assume increasingly capable recommendations automatically justify increasingly broad permissions.
An AI agent can be right and still not be authorized to act.
That isn't a failure of intelligence.
It's a boundary that trustworthy enterprise systems need to respect.
Further Reading
- Enterprise AI
- Agentic AI
- Governance
- Product
- Human-in-the-Loop