Working with agentic super intelligence

Regen puts nine agents on your team, and they are agentic: they text, call, book, route, follow up and log every decision, rather than suggesting what a person might do. This chapter sets the division of labor between those agents and your people: what each agent owns, when it must hand a customer to a person, what a person never hands to an agent, and how managers train, test and inspect the agents every day. The rule underneath all of it is simple. Agents own coverage, speed and consistency; people own judgment, relationships, negotiation and every commitment that binds the store. Run that split well and the store answers every lead in seconds, brings lost leads back to life, and spends far fewer human hours per appointment.

The deal between people and agents

Most stores get AI wrong in one of two ways. Some treat it like a fancy autoresponder and keep the agents on a short leash, so salespeople still send thousands of manual follow-ups a month and the cost savings never show up. Others treat it like a replacement for judgment, let an agent talk numbers or hint at an approval, and pay for it in trust and complaints.

The top 1% store does neither. It gives the agents everything that is repeatable, rule-bound and time-sensitive, and it keeps everything that needs judgment in human hands. Use one test for any task:

If this goes wrong, does it cost a customer's trust or the store's money in a way only a person can fix? If yes, a person owns it. If no, and the task is repeatable and time-sensitive, an agent owns it.

Answering a 10:40 p.m. website lead is repeatable and time-sensitive: an agent owns it. Telling a customer what their trade is worth binds the store: a person owns it. Following up a no-show 15 minutes after the missed time, every time, is what super intelligence is for. Deciding whether to stretch on a deal is what a manager is for.

What each agent owns

A few working rules per agent, learned the hard way:

• Revive never argues with a stall reason. It restarts the conversation with an angle that fits the reason. When a lead says they bought elsewhere, Revive congratulates them, stops sales messages and closes the file honestly.
• Respond sends a personal first message, not an acknowledgement. Personal means it names the vehicle, answers the question that was asked and ends with one easy question. "Thanks for your inquiry, a representative will contact you" does not count, and the speed-to-lead clock keeps running until a real answer goes out.
• Converse collects trade details (year, make, model, miles, condition, payoff) and credit comfort, and sends the store's secure application link when asked. It never says or implies that anyone is approved, likely to be approved, or will get a particular rate.
• Voice opens every call with its AI disclosure and the recording notice, in that first sentence or two. It warm-transfers when a person is free; it never cold-transfers into a voicemail box. Outbound, it calls only contacts with recorded prior express written consent for that kind of call.
• Service sends the technician's photos and video, but the advisor sets the price and the customer approves in writing. Equity conversations go to a sales rep the same day.
• Campaign runs nothing without a manager approving the audience, the copy and the pacing. Guardian checks consent on every recipient.
• Coach coaches; it does not discipline. Managers own consequences.
• Guardian sits in front of every send, human or agent. Nobody overrides an opt-out or a quiet-hours block. Ever.

Hand-off rules

An agent hands the conversation to a person, immediately, when any of these happens:

1. The customer asks for a person, in any wording. "Can I talk to someone," "is there a human," "call me" all count.
2. The customer wants a number that commits the store: a firm price, a payment, a rate, an out-the-door figure, a deposit. The agent may point to the published listing. Any number the store has to stand behind comes from a person.
3. Credit goes beyond collection. Collecting comfort level and sending the secure link is fine. Anything about approval, odds, rates or lenders goes to a person.
4. Trade value as a number. The agent gathers the details and books the appraisal. A person gives the value.
5. Condition questions beyond the record. The agent may share what is documented: the inspection report, history report, photos and listing. Anything else ("has it ever had paint work?") goes to someone who can check.
6. Frustration, a complaint, or any mention of an attorney, a regulator or a dispute.
7. Ready-to-buy signals: "I want it," "can you hold it," "how do I put a deposit down."
8. The agent cannot answer from approved knowledge. It says so plainly and never guesses.
9. A sensitive moment: bereavement, illness, job loss, financial hardship. A person responds.

The hand-off packet

Every hand-off arrives with the same packet, so the customer never repeats themselves: a one-line summary; the buying profile (vehicle, use, budget, trade, credit comfort, timing, decision makers) with gaps marked; vehicles of interest with stock numbers; trade details; anything the agent promised and by when; the suggested next step with a word track; consent status; and the full transcript. Read the packet before you send a word.

The pickup clock

During open hours the assigned rep picks up within 5 minutes; top operators do it inside 2. At 5 minutes without action, Regen pings the rep again. At 10 minutes it routes to the manager on duty. The customer is never left hanging and is never told a person is coming when no one is.

Agent to customer, text, at hand-off during open hours:

Good question, and you deserve a straight answer. I've passed this to {{rep_name}}, who can work the numbers with you. {{rep_name}} has our whole conversation, so you won't need to repeat anything.

Agent to customer, text, at hand-off after hours:

The team is in at [9 a.m.]. I've passed everything to {{rep_name}} so you won't have to start over. Would [9:30] or [11:00] tomorrow work for a quick call?

Why it works: it is honest about timing, it keeps momentum with two options, and it promises only what the store will deliver.

Rep picking up, text:

Hi {{first_name}}, it's {{rep_name}} at {{dealer_name}}. I've read your conversation with our assistant, so no need to repeat yourself. You asked about [the question]. Here's where I'd start: [answer or next step].

Why it works: the customer feels remembered, not processed. That alone separates the store from most of the market.

Any agent, when asked "Is this a real person?":

I'm the {{dealer_name}} virtual assistant, an AI. I can answer most questions and book times, and I can get you to a person right now if you'd prefer. Want me to have {{rep_name}} reach out?

Handing back

A person can return a conversation to an agent with a clear instruction: "Follow up Thursday at 10 a.m. about the second key." Agents follow human notes exactly. That is how a rep keeps 40 conversations warm without sending 40 texts.

What humans never delegate

These stay with people, always. The agents are configured so they cannot do them, and Guardian blocks the attempt.

• Final price and payment commitments. Numbers bind the store and carry disclosure duties.
• Credit decisions and anything said about approval. Only a lender approves; misstatements here are a fast path to a complaint.
• Trade appraisal values as a firm number. A value needs eyes on the vehicle and a manager's sign-off.
• Vehicle condition representations beyond the documented record. If it is not on the report, a person checks before anyone says it.
• An upset customer who asks for a person. They get one.
• Legal and compliance judgment calls. Agents follow rules; people with counsel interpret them.
• Firing a customer. Ending a relationship is a human decision with a human explanation.
• Anything a customer asks a human to do.

Training the agents

Every agent is trainable. Regen super intelligence arrives knowing this playbook; it does not arrive knowing your store. The stores that train their agents weekly pull away from the stores that set them up once. There are seven levers. Use all seven.

1. Instructions

Write instructions like a job description for a strong new hire. Who the store is, who the customer is, what to do first, what matters most. Be specific. "Be helpful" teaches nothing. "When a customer asks about a vehicle, confirm it is in stock from the inventory feed, share one detail they asked about, and offer two appointment times within the next 48 hours" teaches a behavior.

2. Voice

Set the tone the store wants: warm, brief, local. Give sample phrases the store uses and words it never uses. Texts read best at one to three short sentences. If your best salesperson would never write "Greetings," neither should the agent.

3. Knowledge

Hours, address, fees as posted, current programs with their end dates, FAQs, the service menu, the inventory feed. Give every knowledge entry an owner and an expiry date. Stale knowledge is the most common cause of a wrong answer: an incentive that ended Sunday, holiday hours nobody updated. When an entry expires, Regen flags it to its owner.

4. Examples

Approved conversations are the fastest teacher. Collect 10 to 20 strong examples per scenario (new website lead, trade question, no-show, credit worry) from your best humans and your best agent conversations. Add a one-line note on why each is good.

5. Corrections

When a manager edits or flags an agent's message in a live conversation, do not just fix that one message. Write the correction as a rule plus an example: "When a customer asks if a vehicle has been in an accident, share the history report link and offer to have a person walk them through it; do not summarize it yourself." One correction fixes every future conversation like it.

6. Guardrails

Regen ships with guardrails nobody can switch off: identity disclosure, opt-outs, quiet hours, no numbers as fact by text, no approval language. Add your store's own on top: "Never mention the loyalty program after the end date in knowledge," "Never offer a same-day delivery promise on vehicles still in reconditioning." Guardian enforces both layers.

7. Evaluations

Keep a test set of real scenarios with the expected behavior for each: the right answer, the right tone, the right hand-off. Run it before every publish. If a change makes the agent better at trade questions and worse at credit questions, the evaluation catches it before a customer does.

Sandbox and published versions

Every change starts in the sandbox. Test it, run the evaluations, compare against the live version, then publish with a short note on what changed and why. If quality dips, roll back to the last published version. Never edit a live agent on the fly at 4 p.m. on a Saturday: changes batch up for the Wednesday Training Studio review, except compliance or accuracy fixes, which ship the same day.

When you switch an agent on for the first time, start narrow: one lead source, or after-hours only. Review every conversation for the first week, widen the scope when the agent quality rate holds at 90% or better, and keep widening.

Reviewing agent work every day

At 09:00 every morning a manager reviews the overnight agent work. It takes 20 to 30 minutes and it is the highest-leverage half hour in the building.

What to pull:

• Every hand-off made overnight, and whether it was picked up.
• Every conversation with a warning sign: an opt-out, a complaint, an "is this a bot" question, a long silence after an agent message.
• Ten random conversations each from Respond, Converse and Voice.

How to score each one (correct or not, on all five):

1. Correct: facts, inventory and hours were right.
2. Compliant: identity, opt-out handling, no numbers stated as fact by text, no approval language.
3. Complete: it answered the question and asked one that moved the deal forward.
4. Human: it sounded like the store, not a form letter.
5. Handed off right: it handed off when a trigger fired, and not when none did.

The share of reviewed conversations that pass all five is the agent quality rate. Every miss becomes a correction filed that morning. Patterns become instruction or knowledge changes for Wednesday.

Common failures and the fix

How Regen runs this

Respond takes every new lead in under 60 seconds; Converse carries every reply; Voice answers every call the floor cannot; Revive works the dead and aged leads in the background; Service and Campaign run the service lane and the journeys; Coach scores every conversation; Insight writes the 08:15 brief; Guardian checks every send. People take over at the hand-off triggers, with the full packet in hand.

The money shows up twice. First, cost savings: agents absorb first response, follow-up, revival, after-hours coverage and call answering, so manual touches per appointment fall, BDC hours move to closing work, and overlapping tools (a texting app, a chat vendor, an answering service, a review tool) can be retired. Second, lost leads brought back to life: Revive restarts conversations nobody had time to work, and every reply lands with a person who can sell.

What managers inspect

• The 09:00 overnight review: hand-offs, warning-sign conversations and 10 random conversations per agent, scored on the five checks.
• Hand-off pickup time today and this week, by rep.
• Open corrections not yet written up as rules.
• Knowledge entries expiring in the next 7 days.
• Sandbox changes waiting for Wednesday, with their evaluation results.
• Guardian blocks: what was stopped and why, and any record that needs fixing.

Targets

• Agent quality rate. Strong store target: 90%. Top 1% operator target: 97%+. Rationale: below 90% the agents create work for people instead of removing it.
• Hand-off pickup time. Strong store target: under 5 minutes. Top 1% operator target: under 2 minutes. Rationale: a fast agent followed by a slow human still loses the customer.
• Manual touches per appointment. Strong store target: down 50% from baseline. Top 1% operator target: down 70% from baseline. Rationale: this is where the cost savings come from.
• Speed to first response. Strong store target: under 5 minutes during open hours. Top 1% operator target: under 60 seconds, 24/7. Rationale: Respond makes the top target the default, so anything slower means a setup gap.