Autonomy needs a handoff line

Autonomy needs a handoff line

The dealership question around AI autonomy is usually framed too broadly.

Should AI answer customers or should humans answer customers.

That framing misses the real operating problem. The better question is: which turns are safe for the system to handle, which turns must go to a human, and how does the manager know the difference.

Autonomy without a handoff line is risky. Human-only follow-up is slow. The workable middle is narrower and more useful: autonomous execution on eligible threads, inside guardrails, with manager visibility and instant human takeover.

That is the product standard Brevmont has been building toward with Overdrive.

Eligibility is the first control

Not every customer message belongs in an autonomous lane.

A customer asking whether a vehicle is available, whether they can come by, or whether a proposed time works is not the same as a customer asking for an out-the-door number, credit approval, trade value, legal explanation, warranty dispute, or manager escalation.

Those are different risk classes.

The first class can often be worked toward appointment intent. The second class belongs to a person.

Overdrive is designed around that distinction. On eligible supported threads, it can answer in the rep's voice, propose concrete times, recognize text-level appointment agreement, confirm back in the thread, mark the lead as appointment-set, and alert the manager.

That is not the same thing as booking a calendar appointment. It respects active-hours sending guardrails, but it does not validate a proposed slot against dealership hours or a rep calendar. It does not promise a sold car. It does not quote financing, credit, trade dollars, discounts, or legal terms.

The boundary matters because a dealership should not buy vague autonomy. It should buy narrow autonomous execution with an explicit refusal line.

Handoffs have to be sticky

A handoff is not a suggestion. It is a state change.

If a customer asks for a manager, the system should hand off. If the buyer gets hostile, raises legal concerns, pushes finance terms, asks about credit, demands trade dollars, or moves into negotiation, the system should hand off. If a rep manually replies in an active autonomous thread, the system should pause on that thread.

That last piece matters.

The AI should never talk over a human who has stepped in. Manual takeover has to be sticky until someone intentionally resumes the autonomous lane. Otherwise the product becomes unpredictable at the exact moment trust matters most.

The manager and rep both need to understand the state of the thread. Is Overdrive active here. Is it paused because of takeover. Did the customer trigger a handoff. Can the rep resume it. Those answers need to be visible in the workflow, not buried in a settings page.

Discoverability is safety.

The manager needs receipts

Autonomy changes the manager's job.

With manual follow-up, the manager is trying to find out whether work happened. With bounded autonomy, the manager also needs to know what work the system handled, what it refused, and where a human needs to step in.

That is why every automatic action should produce a receipt.

The receipt should show the thread, action, handoff state, rep owner, dealership guardrails, and any appointment-intent marker. It should also show when the system did nothing because a STOP or opt-out gate, active-hours rule, human takeover, manager kill switch, or dependency failure blocked the turn.

Nothing should be silent.

If the system replies, the manager sees it. If the system hands off, the manager sees it. If the system refuses to reply because a safety check failed, the manager should be able to inspect that too.

That is the difference between an autonomous feature and an accountable operating system.

The guardrails are product, not paperwork

Guardrails are often treated like legal copy written after the feature ships.

That is backwards.

In dealership AI, guardrails are product behavior. STOP and opt-out handling, active hours, idempotency, double-reply protection, send budgets, per-thread pause, per-rep toggle, manager master toggle, dealership kill switch, and dependency-failure blocking are not policy decorations. They are the feature.

One useful example is double-reply protection.

A customer should not receive two AI replies because the browser, network, worker, or message detector retried the same turn. Brevmont's Overdrive safety model includes layered protection against that class of failure: duplicate hashing, echo detection, idempotency, in-flight locking, same-turn cooldown, similarity breaking, and send-token control.

That is not exciting language for a pitch deck. It is exactly the language that should matter to a GM asking whether the system can be trusted near live customers.

What dealers should ask vendors

A dealership evaluating autonomous follow-up should ask specific questions.

What exact threads are eligible. Which triggers force handoff. Does manual rep reply pause the system. Can a manager kill it for the store. Can a rep turn it off. Can a manager see every automatic reply. What happens after opt-out. What happens outside active hours. What happens if the guardrail service fails. Does the system retry safely. Does it claim to book calendar appointments or only confirm appointment intent in the conversation.

The answers should be operational, not vague.

"The AI knows what to do" is not an answer.

The right answer names the lane, the refusal line, the handoff state, and the receipt.

The future is bounded

Dealer AI is not moving from humans to machines in one jump.

The useful path is bounded autonomy. Let the system handle the narrow work that burns attention and loses speed. Keep humans in control of judgment-heavy turns. Give managers visibility into both.

That model respects the sales floor. It recognizes that reps still close deals, managers still coach, and customers still need accountable humans when the conversation gets serious.

Autonomy can belong on the floor. It just needs a handoff line sharp enough that everyone knows where the human takes over.

brevmont