Every AI Recommendation Needs an Owner

Across inventory, customer conversations, service and finance and insurance (F&I), the same discipline keeps returning in this series: evidence must reach the person authorized to make the decision, and that person must remain accountable for what happens next. This chapter is about who that person is, and about what the dealer principal owes the store before any recommendation reaches a manager.

“Human in the loop” has become an easy promise. It can also be an empty one. If a manager receives a recommendation with no source evidence, no explanation of the rule and no practical way to change it, clicking approve is not meaningful oversight. The human is present, but the system still owns the decision.

Responsible AI begins farther upstream. Before a recommendation reaches anyone, the dealership should know what the workflow is allowed to do, which data it may use, how uncertainty is handled, who has authority, and what record will remain afterward. That is not a model feature. It is an operating system for accountability, and it costs a few pages of writing per decision.

Borrow the structure, not the paperwork

There is no need to invent a governance vocabulary. The National Institute of Standards and Technology published its AI Risk Management Framework in January 2023 with four core functions, govern, map, measure and manage, and followed it in July 2024 with a profile for generative AI. The framework is voluntary and written for organizations far larger than a dealership. Its structure still fits a store, because it asks the questions a dealer principal already asks about any new process: who is accountable, what could go wrong, how will we know, and what will we do about it.

The figure above translates each function into dealership terms. Govern is a named owner and a review cadence. Map is the decision, the data, the customers affected and the consequence of being wrong. Measure is override rates, time to decision and outcomes. Manage is tuning thresholds, pausing a workflow and testing changes before they reach a manager. None of that requires a committee. It requires writing down what the store already believes.

Start with a decision, not a general AI policy

Broad principles matter, but governance becomes real at the level of a specific decision.

Consider three uses of AI in one store: summarizing yesterday's sales activity, recommending a price review for an aged unit, and flagging a possible identity inconsistency in a finance file. Each has a different consequence if it is wrong. The first wastes a few minutes. The second can cost gross or an aged unit. The third touches a customer's identity, a lender's trust and the store's obligations under the Safeguards Rule. Each needs different data, review, permissions and escalation.

The dealership should define a compact operating contract for every workflow:

• The decision or task the system supports
• The approved source data and its limitations
• The condition that creates a recommendation or exception
• The person authorized to approve, change, reject or escalate it
• The actions the system may prepare and the actions it may never take alone
• The evidence and outcome that will be retained for review

This makes responsibility concrete. It also prevents a successful low-risk use case from becoming justification for deploying the same controls in a higher-risk context. The ownership table above is what those contracts look like once they are collected in one place: six decisions, each with one person who decides, one who reviews, one who is informed, and a plain statement of what the system may do on its own, which in the case of a customer message or a finance file is to draft or surface, and nothing more.

Traceability is the foundation of review

A manager cannot evaluate a conclusion without seeing where it came from. A pricing recommendation should expose the relevant inventory history, comparable vehicles, merchandising condition, and the rule or calculation. A customer follow-up suggestion should show the interaction or event that made follow-up appropriate. A risk exception should identify the fields or behaviors that differed.

Traceability should include source freshness and confidence. Old data can be accurate and still be wrong for the current decision. A weak match can look precise after it has been compressed into a score.

The interface should therefore make uncertainty visible. When evidence is incomplete, the recommendation should ask for review or more information. It should not fill gaps with certainty because certainty is easier to display.

This protects more than compliance. It preserves operator trust. Managers are more likely to trust a system they can interrogate; unexplained answers invite workarounds and weak review.

Human control requires real decision rights

Approval should not be the only available response. A qualified reviewer should be able to edit the proposed action, reject it, defer it for a stated reason, assign it, or escalate it. The workflow should record that choice without treating disagreement as user error.

There are also places where review must occur before any external action. Customer communications, pricing changes, sensitive deal decisions and other consequential steps should follow approved dealership policy and applicable requirements. AI can prepare work and assemble evidence. It should not quietly expand its authority because an earlier recommendation was accepted.

Permission boundaries should match the operating boundary. The person who can review an inventory exception may not be authorized to see sensitive customer or finance information. The person who approves a communication may not be allowed to change the rule that selected the audience. Separation of duties is not bureaucracy when the data and the consequence are different. It is how the dealership keeps a useful tool inside an accountable process.

Measure the recommendation and the system around it

Governance cannot stop at an audit log. The dealership also needs to know whether the workflow is useful.

Track what managers did with recommendations and what happened next. High rejection may mean the threshold is wrong, the evidence is weak, or the workflow does not fit the team's playbook. Automatic approval of nearly everything may indicate trust, but it may also indicate that reviewers are not engaging with the decision.

Useful measures include:

• Recommendations approved, changed, rejected, deferred or escalated
• Time from material exception to manager decision
• Override reasons and recurring missing context
• Outcomes after approved actions, with appropriate limits on attribution
• False-positive or low-value alert patterns
• Changes to data sources, rules, prompts or permissions, and who approved them

No single measure proves responsibility. Together they show whether the control is understandable, used as intended, and improving with review. A store that cannot produce these numbers for a workflow does not yet govern it.

Put magnitudes on this. A fictional store running an aged-inventory workflow sees 40 price-review recommendations in a month. If managers approve 38 without a comment, the controller should ask whether anyone is reading them. If they reject 25, the threshold is probably wrong, or the comps are. If they approve 22, change 10 with a reason and reject 8 with a reason, the workflow is doing what a good manager does, and the 18 reasons are the most valuable output the store produced that month, because they are the raw material for a better rule.

Treat changes as operating changes

AI workflows change when a data feed changes, a rule is revised, instructions are updated, or a model, vendor service or runtime behavior changes. A dealership should not treat those changes as invisible software maintenance when they affect a management decision.

Material changes need an owner, test cases, approval and a rollback path. The team should verify that the workflow still uses the intended sources, produces the expected evidence, respects permissions and routes uncertain cases correctly. The last row of the ownership table exists for this reason: a threshold change is a decision like any other, and it is the one most likely to be made quietly.

The floor has not moved

It is tempting to wait for AI-specific rules before writing any of this down. The record argues against waiting. The Federal Trade Commission's CARS Rule, announced in December 2023, was vacated by the Fifth Circuit in January 2025 and formally withdrawn by the FTC in February 2026. In the same period the FTC kept bringing dealer cases under the law that already existed: a $10 million settlement with the Napleton group in April 2022 over added fees and discriminatory financing, and a $2.6 million judgment against Coulter Motor Company in August 2024 over deceptive pricing and discriminatory practices, brought jointly with the State of Arizona.

The Safeguards Rule did move, toward the store. Since May 2024 a dealer that discovers a notification event affecting at least 500 consumers must tell the FTC as soon as possible and no later than 30 days after discovery. Any workflow that touches customer data inherits that clock. The obligations a dealership carries, under the Equal Credit Opportunity Act, the Fair Credit Reporting Act, the Safeguards Rule and the FTC Act, apply to a recommendation whether a person or a model produced it. Governance is how the store proves it knew that.

Start narrow. Choose one repeatable workflow with a named manager and a measurable result. Review it regularly. Expand only after the team understands how the system behaves in normal and exceptional cases. Responsible AI is not the opposite of useful AI. It is what makes useful AI durable. The dealership keeps control because every recommendation has evidence, every action has an owner, and every workflow is open to review.

Questions for the next governance review

• Can the authorized manager inspect the evidence, uncertainty and rule behind each recommendation?
• Does the reviewer have real choices and clear authority, or only an approval button?
• When data, rules, permissions or instructions change, who tests and approves the operating impact, and where is that recorded?