AuthorityShouldExpandMoreSlowlyThanAssistance

AI can become useful inside an agency long before it should be trusted to act independently.

It can retrieve approved information, compare documents, structure a packet, identify missing fields, prepare a draft, and flag an exception. Those tasks remove administrative friction while leaving accountable judgment visible.

The risk begins when technical capability is treated as permission.

Because the system can draft a client message, leadership assumes it can send one. Because it can identify a coverage difference, the workflow begins presenting the difference as advice. Because it can prioritize requests, employees stop inspecting the routing logic.

Capability and authority are different design questions.

Three levels of work

Prepare

The system gathers, retrieves, compares, structures, extracts, and drafts.

Examples include assembling an approved renewal packet, extracting submission fields, comparing policy documents, or drafting a service response from verified account context.

Preparation can often move quickly because the output remains inspectable and a human still owns what happens next.

Recommend

The system prioritizes, flags, proposes, or suggests a next action.

Examples include identifying a likely material renewal change, suggesting which service request deserves urgent attention, or proposing a producer follow-up based on known account activity.

Recommendation requires more than accurate retrieval. The workflow must show the evidence, uncertainty, business rule, and reason the recommendation deserves attention.

Decide

The system commits the agency to an action or communicates authority.

Examples include binding coverage, interpreting whether a policy satisfies a contractual requirement, approving a final client recommendation, selecting a market strategy, or sending a communication that carries licensed advice.

These decisions remain with licensed or accountable people unless the agency has explicit authority, evidence, controls, and a defensible reason to change the boundary.

A renewal communication example

An AI-assisted workflow can gather the prior policy and renewal proposal, compare premium and coverage terms, identify material changes, and draft a client-ready explanation with citations.

That is meaningful assistance.

The account manager still decides which changes matter most, whether the client needs a call instead of an email, how the relationship affects the framing, and whether the communication accurately represents the agency's advice.

The system can prepare the conversation. It should not quietly become the advisor.

A human checkpoint is not a button

Many workflows claim to keep a human in the loop because someone clicks approve.

Approval proves very little if the reviewer cannot inspect the evidence, does not understand the uncertainty, lacks time to evaluate the output, or assumes the system has already checked what matters.

A real checkpoint presents:

• The proposed output
• The approved sources
• The relevant uncertainty
• The exception status
• The decision being requested
• The route for correction or escalation

The checkpoint should reduce the work required to judge without reducing the judgment itself.

If the reviewer must reconstruct the entire account from several systems, the assistance has not prepared the decision. If the reviewer receives only a polished answer, the workflow is asking for trust rather than review.

Expand assistance through evidence

Assistance can expand from one task to another when the agency understands the boundary.

A system that reliably assembles a renewal packet may next compare approved documents. A comparison workflow that makes sources and exceptions visible may next prepare a draft. A useful draft may support more account segments.

None of those improvements automatically justify autonomous client communication.

Authority deserves a separate decision.

Ask:

1. What consequence follows if the output is wrong?
2. Can the agency trace the decision to approved evidence?
3. Is the relevant uncertainty visible?
4. Who is legally and operationally accountable?
5. Can the action be reversed?
6. Have difficult cases been observed?
7. Does the relationship require human interpretation?

Capture why the reviewer changed the output

Human review creates value only if the agency learns from it.

When an account manager changes the draft, record whether the problem was missing data, poor phrasing, wrong emphasis, relationship context, an unclear source, or a judgment that should never be delegated.

The correction may improve the model. It may also prove that the boundary is working as designed.

Not every recurring human decision should be automated. Some should be made easier to perform consistently.

The operating principle

The goal is not merely to keep a human somewhere in the process.

It is to keep the right human in control of the right decision with the right evidence.

Authority should expand more slowly than assistance.

Related framework: Regesta Board, Bound the Authority.