Back to all articlesHuman and AI collaboration

Decide Which AI Recommendations Need Human Approval

Alactic Editorial11 min read

Prepared with AI assistance by Alactic Editorial. Examples identified as hypothetical are illustrative, not customer results.

A manager reviewing an AI-generated summary, a proposed customer concession, and a disputed delivery promise should not use the same approval rule for all three. Reviewing everything turns the manager into a bottleneck. Reviewing only unusually large requests misses smaller decisions that create expectations, conceal uncertainty, or make exceptions.

Keep decisions with a human when they require resolving disputed facts or making a material trade-off. Require approval before proceeding when AI can prepare a sound recommendation but the consequence or exception needs an accountable person's consent. Delegate handling only when the facts are sufficiently clear, the downside is bounded, and an established rule determines the answer.

The practical distinction is not whether AI can produce convincing language. It is whether the organization has already decided what should happen in that situation. Use consequence, uncertainty, and discretion to make that distinction consistently.

1. Define three decision rights, not one approval queue

Start by giving managers and reviewers a shared meaning for each handling mode.

  • Human decision: A named person determines the outcome. AI may organize evidence, identify missing information, and draft alternatives, but its recommendation is not the default answer.
  • Approval before proceeding: AI prepares a proposed outcome with supporting facts. A person with the relevant authority must explicitly accept or change it before anyone treats it as a business commitment.
  • Delegated handling: A case can proceed without case-by-case approval because a person has already authorized a bounded class of decisions. A named human remains responsible for the delegation and its exceptions.

The difference between the first two modes matters. In approval mode, the reviewer evaluates a proposal against a known standard. In human-decision mode, the person must establish which outcome is appropriate, potentially after gathering more evidence. Asking someone to approve a confident draft of an unresolved dispute can anchor their judgment before they have examined the facts.

Delegation is also not an absence of human judgment. It is judgment exercised in advance: defining the eligible cases, the acceptable outcomes, and the circumstances that return a case to a person. These are operating choices, not claims about any particular product's controls.

2. Assess consequence, uncertainty, and discretion separately

For each recommendation, write a short assessment on three dimensions. Avoid collapsing them immediately into a single score: a low score on one dimension should not cancel a serious concern on another.

Consequence: what changes if this is wrong?

Consider money, customer expectations, disclosure of sensitive information, workload, and the ability to reverse the result. Ask who would bear the cost and whether they would discover the mistake in time to respond.

An internal document may have limited consequences if it is clearly a working note. The same document becomes more consequential if a manager uses it to allocate scarce capacity. Classify the intended use, not just the document type.

Uncertainty: what is not established?

Look for missing records, inconsistent statements, stale information, unclear customer intent, and unsupported inferences. A fluent explanation is not evidence that these gaps have been resolved.

State the uncertainty concretely. “The customer says a date was promised, but the available notes contain only a requested date” is more useful than “medium confidence.” It tells the reviewer what must be checked.

Discretion: has the organization already chosen the rule?

Low-discretion cases have an applicable rule and sufficiently clear facts. High-discretion cases require choosing among competing interests, making an exception, or interpreting an ambiguous commitment.

A familiar task can still involve substantial discretion. Answering a question about standard opening hours is different from deciding whether to keep a team available after hours for an important account. Both may arrive as short customer messages; only one has an established answer.

3. Use a routing rule that preserves the strongest concern

Apply these tests in order rather than averaging the dimensions:

  1. Does the case require deciding what the business should stand behind? If facts are disputed, interests conflict, or a consequential exception lacks an established standard, assign a human decision.
  2. Is there a defensible recommendation, but a consequential commitment or bounded exception still needs consent? Require approval before proceeding.
  3. Are the facts clear enough, the consequence limited, and the answer already determined by an authorized rule? Delegate handling within that boundary.
  4. Is a critical fact missing? Do not treat the missing fact as permission. Seek clarification or send the case to the person who can resolve it.

High consequence does not always mean humans must invent the answer from scratch. A well-defined proposal may be suitable for approval. Conversely, a small financial amount does not make a decision routine if it sets a precedent or implies acceptance of a disputed account.

Use a simple explanation for every assignment: “This needs approval because it changes the customer's commercial outcome, even though the facts are clear.” That sentence is easier to challenge and improve than an unexplained risk label.

4. Fictional worked example: a filled decision table

The following company, cases, amounts, and rules are hypothetical. A service team uses AI to prepare notes and recommend customer responses. Its manager wants routine work to move without giving AI discretion over exceptions or disputed promises.

The compact decision table below compares four cases. The detailed entries immediately afterward provide the decision rights and approval rationale that a team could adapt.

Case                    Consequence  Uncertainty  Discretion  Handling
Internal summary        Low          Low          Low         Delegated
Customer concession     Moderate     Low          Moderate    Prior approval
Disputed commitment     High         High         High        Human decision
Routine clarification   Low          Low          Low         Delegated

Internal summary: delegate the note, not the judgment

  • Facts: AI summarizes a completed customer conversation for an internal handover. The record is complete, and no statements conflict.
  • Decision right: Delegated handling for a factual working summary. The receiving team lead owns the delegation.
  • Boundary: The summary may identify requests and recorded next steps. It may not turn an unconfirmed preference into an agreed commitment or decide which customer deserves priority.
  • Approval rationale: Case-by-case approval adds little when the summary is traceable and is not itself a decision. If it will become the sole basis for a consequential allocation, reassess that use before relying on it.
  • Return to a person when: Records conflict, ownership of a next step is unclear, or the proposed summary omits a material objection.

Suggested customer concession: require approval

  • Facts: A documented service failure occurred. AI proposes a hypothetical $75 credit. A service manager may authorize credits up to a hypothetical $100, but this team has not delegated concession choices.
  • Decision right: Approval before proceeding by the service manager. AI may recommend an amount and explain its basis; it may not present the credit as agreed.
  • Boundary: The proposal must identify the verified failure, the customer's requested remedy, and why the suggested amount is proportionate under the team's concession guidance.
  • Approval rationale: The evidence is clear enough for a proposal, but the amount still requires judgment and changes the commercial outcome. The manager's authority limit is not an automatic allowance.
  • Return to a human decision when: The customer disputes the underlying facts or the concession is being considered in exchange for a broader commitment.

Disputed commitment: keep the decision with a human

  • Facts: A customer says a delivery date was guaranteed. One note records the customer's requested date; another says the team would investigate availability.
  • Decision right: A named account lead decides what the business will communicate after reviewing the evidence and consulting the relevant delivery owner.
  • Boundary: AI may build a chronology, separate documented statements from claims, and list unresolved questions. It should not resolve the dispute by selecting the most plausible account.
  • Approval rationale: This is not merely a draft awaiting a signature. The facts and the business response both require judgment. A polished proposed answer could prematurely frame the decision.
  • Immediate handling: Acknowledge that the discrepancy is being reviewed without affirming an unverified promise.

Routine clarification: delegate a narrow factual answer

  • Facts: A customer asks which published reference number to include in a support request. The current guidance gives an unambiguous answer.
  • Decision right: Delegated handling under the support lead's standing authorization.
  • Boundary: Explain the published requirement without requesting unnecessary information or inventing an alternative process.
  • Approval rationale: The answer has limited downside, clear evidence, and no meaningful choice among business outcomes.
  • Return to a person when: The customer cannot follow the standard process, the guidance conflicts with another instruction, or the request includes a disputed obligation.

5. Watch for the routine case that changes category

A useful decision boundary follows the substance of a case, not its subject line. The clarification above stops being routine if the customer adds that an employee previously promised an exception. The summary stops being low consequence if it becomes the definitive account of a complaint.

Look for these category-changing signals:

  • A new audience: An internal working note will be shared externally or used by someone who cannot inspect its evidence.
  • A hidden commitment: A proposed clarification implies that a date, price, capacity allocation, or exception has been accepted.
  • An unresolved objection: The available record contains disagreement that the recommendation smooths over.
  • Accumulated impact: Repeated individually small concessions create a larger pattern that the standing rule did not consider.
  • A changed purpose: Information prepared for reference is now being used to decide priority or evaluate performance.

Do not respond by requiring approval for every message. Instead, write the trigger that changes the handling mode. “Delegate factual clarification unless it requires interpreting a prior promise” is a more useful boundary than “delegate simple questions.”

6. Make approval a decision, not a reading task

An approval request should let the reviewer inspect the judgment quickly. Include the proposed outcome, evidence, unresolved facts, relevant rule, consequence of being wrong, and the reason this person is being asked.

For the hypothetical credit, that means more than a drafted apology. The reviewer needs the documented failure, proposed amount, applicable guidance, and any previous remedy. They should be able to approve, revise, reject, or request further evidence without reconstructing the case.

Assign approval to someone with authority over the outcome, not merely whoever is available. If the usual reviewer is unavailable, a named substitute can take the decision within their authority. Silence is not approval, and urgency does not turn an exception into a routine case.

Approval has costs: delay, reviewer fatigue, and a tendency to accept a plausible first proposal. Delegation has costs too: errors may spread before anyone notices. Use approval where the reviewer can materially improve or authorize the decision, rather than as a symbolic checkpoint.

7. Recover from mistakes by fixing the decision boundary

When a delegated case goes wrong, first establish what was communicated or relied upon, who is affected, and which accountable person will determine the correction. Preserve the difference between the original evidence and the AI's interpretation.

Then identify the failure type. Was the factual record wrong? Was the recommendation reasonable but outside the delegated boundary? Did a reviewer approve without noticing missing evidence? Or was the boundary itself too broad?

The remedy should match the failure. A missing exception trigger calls for a clearer routing rule. An overloaded approval queue calls for better reviewer capacity or narrower requests. Neither is fixed simply by telling people to be more careful.

Temporarily narrow the affected delegation if necessary, but avoid sending all unrelated work back for approval. Restore the bounded activity only after the owner can explain what changed and what evidence will show whether that change works.

8. Test the classification on real work before expanding delegation

For a hypothetical pilot, choose 20 recent recommendations from one team. Include ordinary cases and known exceptions. Have two managers classify them independently by consequence, uncertainty, discretion, handling mode, and decision owner.

Compare disagreements before changing responsibilities. If one manager delegates a case and another requires a human decision, identify the missing fact or ambiguous boundary. Rewrite that boundary in plain language and classify the case again.

During the pilot, track observable outcomes:

  • Cases where the two reviewers disagreed on handling mode.
  • Approval requests returned for missing evidence.
  • Delegated cases later found outside their authorized boundary.
  • Recommendations materially changed by a reviewer, including the reason.
  • Time spent waiting for approval versus time spent resolving uncertainty.

Do not treat a high approval rate as proof that review is unnecessary; reviewers may be rubber-stamping. Likewise, a high escalation rate may indicate unclear guidance rather than poor recommendations. Inspect the reasons behind the counts.

The immediate next action is to fill this decision table for four recurring cases in your own team and name the accountable owner for each. Expand delegation only where the evidence shows a stable rule, recognizable exceptions, and an observable way to detect mistakes.

Human and AI collaborationDecision rightsManagementAI oversight

Put the next conversation to work.

Explore Alactic Communications for WhatsApp conversations, campaigns and AI-assisted workflows. Choose the product and plan that fit your team.

Explore Communications