AI actions, approvals, and Autopilot

The dealer AI can do real work for you: add a vehicle, move a lead, start a deal, even record a loan payment. Every write starts as a reviewable proposal card. In Ask first mode, nothing changes until you choose Approve. In web chat, you can also reply a plain "yes" when the AI's latest reply contains exactly one proposal in total. If the owner has enabled Autopilot for an eligible routine action, the application may try it automatically after the AI finishes its reply and all final safety checks pass.

How it works

For a manual-only proposal reply, the AI explicitly says “Nothing has changed yet.” If one reply contains both manual and Autopilot-eligible cards, that sentence applies only to the manual cards it identifies; the automatic cards are still only conditional attempts until their final checks finish.

When you Approve, the card runs the real action: the same one the manual screens use, with the same rules and audit trail behind it. Autopilot uses that same execution path and rechecks its permission at the last moment. On confirmed success the card flips to an Applied badge, and for a new vehicle, lead, or customer you get a View link to jump straight to the record. (Other actions, like a status change or a loan payment, simply show Applied, with no View link.)

If the application cannot confirm the result after execution may have started, the card becomes Needs review and cannot be approved again. Choose Reload status for a fresh check. If it remains Needs review, inspect the affected record and activity trail before taking another action. AutoDealer also pauses every new AI-proposed change for the dealership until that receipt is reconciled. Other pending cards may remain visible, but they cannot run. Contact support after you inspect the affected record so AI changes can be resumed safely. This prevents an uncertain email, payment, or record change from being repeated.

If you Reject instead, the card collapses to a Dismissed badge. Rejecting only stops a card that's still waiting. It never reverses something you already approved.

Approve in web chat by just saying "yes"

You don't have to hunt for the button. If the AI's last message proposed a single change, you can simply reply yes (or "ok", "do it", "go ahead", "approve") and it runs. Reply no (or "cancel", "skip", "nevermind") to dismiss it.

If the shortcut is refused before an action starts, the card stays Needs approval so you can review and click it directly. Needs review is reserved for the different case where execution may have started but its result could not be confirmed.

Autopilot: skip the approval for routine changes

If you own the dealership, you choose how much the assistant does on its own under Settings → AI assistant. Ask me first: every change waits for you. Do the routine: eight everyday jobs happen on their own (publishing cars, descriptions, tasks, lead assignment and pipeline moves, and remembering your preferences) and everything else asks. Run it: everything the assistant is allowed to apply happens on its own, within the daily limit. Fine-tune under the card lets you pick exactly which changes apply on their own instead. Answers only is the brake: the assistant answers and looks things up but prepares no changes, and cards from before wait until you turn it off. Every change it makes on its own is logged and undoable from the audit log, and the morning digest lists them.

Four things stay true no matter what you tick:

• Money and irreversible actions always ask first. Recording a payment, marking a deal sold, sending an email or text the assistant wrote: these can never be put on autopilot. The setting won't offer them, and the server refuses them independently.

• Reading outside content pauses Autopilot for that response and related follow-ups. If the AI read something written by someone outside your team while working on your ask, any action from that response comes as a normal approval card instead of applying automatically. This includes lead or customer messages, attachments, the names and notes of documents filed from an email, imported vehicle descriptions, outside pricing labels, saved schedule text, activity details, and accounting records when that feature is available. The safety state is saved with the conversation or background job and stays manual while that context is carried forward. Start a new conversation for unrelated work. This keeps outside text from triggering an unattended change, no matter what it says.

• Everything is logged, and undoable from the log. An autopilot change shows in the Audit log as "Assistant (for the staffer whose setting allowed it)", with what it changed and an Undo button while the change can still be put back. Filter the log by "Assistant" to see everything the assistant did on its own. Every row the change itself writes says the same: the vehicle's price change, the lead's new status, the listing it published. A task the assistant created on its own carries an Assistant badge on the board, a call it logged or a templated email it sent on its own says "logged by your assistant" on the timeline, a comment it added to a task reads "Assistant (for Jane Doe)" when Jane's standing permission allowed it, and a title step it took shows "by the assistant" in the title history, so nobody is credited with work they did not do. The work of a job a trigger or a schedule started is the assistant's too. A change you approved by hand is logged as yours.

• Final checks can still require you. A changed setting, role, entitlement, or safety condition can leave the proposal waiting for approval. If execution may have started but the result cannot be confirmed, the card becomes Needs review instead. Do not approve it again; reload its status and inspect the affected record.

• There is a ceiling. Autopilot applies at most 10 changes in one reply or background run, and at most 50 in any 24-hour window per dealership. Past either limit the remaining cards wait for a person and say why. Approving by hand never counts against it. The owner can lower either number for the dealership under the dial card (Fine-tune: how much Autopilot may do without you); neither can be raised above these, and leaving a field empty goes back to the ceiling.

Only the owner can change autopilot settings, and each staffer's role limits still apply: the AI acting for a salesperson can only auto-apply what that salesperson could do themselves.

By area. Under the dial, "By area: hold one part of the business below the dial" lets you keep one area more cautious than the rest. Each area (Inventory; Leads and customers; Deals, titles and loans; Tasks; Customer messages; Books; Website, reports, memory and background work) follows the dial unless you pick a lower level for it. Holding the last one at Answers only also takes away the website copy cards, because website words live in that area. Do the routine allows only the routine changes there. Ask me first keeps the cards but applies nothing there on its own. Answers only takes that area's change tools away from the assistant entirely: it can still look things up there, but it cannot even prepare a change, and it will tell you so if you ask for one. A level at or above the dial changes nothing, and the card says so next to it. A card that was already waiting when you held its area at Answers only shows "Waiting: Answers only" until you release the area.

Evidence and the automatic step back. The dial card carries "Evidence for moving up": for each job that can run on its own, how many times you approved it, rejected it or undid it in the last 30 days, and whether that makes it ready (five approvals or more, nothing rejected or undone). Only what happened since the dial existed counts; the old checkbox autopilot is not treated as your consent. And if you undo something the assistant did on its own, that one job goes back to asking first, with a banner on the card that says so and a Restore button that lets it follow the dial again. Everything else keeps running as before.

The supervisor (off by default). Under the dial card, "Supervisor: pause Autopilot for an hour when something looks wrong" is an automatic safety check the owner can turn on. Every quarter hour it reads what Autopilot applied on its own and pauses those automatic changes for one hour if it sees more than 20 of them in an hour, one price change of 25% or $2,500 in a single step, or 3 automatic changes undone by your team in a day. A pause is not a Stop and it is not you turning anything off: your dial is unchanged, approval cards still work, everything the assistant proposes still arrives, and the pause clears itself when the hour is up. You get a bell notification with the reason, the card shows it with a Resume now button, and the audit log records the pause and the resume. It is off by default while the thresholds are tuned on real activity.

Background jobs: hand it bulk work

For work too big for one reply ("reprice every SUV over 60 days old", "clean up the listings missing photos"), the AI can run a background job. Starting one is itself an approval card showing the job's name and exactly what it will do. Once you approve, the job runs on its own in segments, posting its progress back into the same chat.

A job never gets special powers: every change it wants to make goes through the same approval cards (or your Autopilot settings) as if you'd asked in chat. When a step needs your approval, the job pauses, you get a notification, and it continues automatically after you decide. If a card waits a week with no decision, the job stops and says so in your bell; the card itself stays in the chat. You'll get a notification when it finishes, and asking "how's my job going?" gives you the status any time.

If a run is interrupted after it may have saved or applied work, AutoDealer.io marks it as Needs review instead of repeating it automatically. Check the chat to see how far it got before you intentionally start it again.

Schedules: standing work on a clock

For work you want done on a rhythm rather than once ("every weekday at 7:30, check for leads nobody has touched and brief me"), create a schedule. Ask the AI for something recurring, or manage them under Settings → AI assistant → Schedules: pause, edit, delete, see each run's history, or hit Run now to try one immediately. Times are your dealership's local time.

A schedule is just a clock on the background-job machinery, so nothing changes about safety: each run's proposed changes still arrive as approval cards (or apply via your autopilot settings), and the run's briefing lands in your AI chat marked as an update so you can't miss it. Creating a schedule always needs an explicit approval, even with autopilot on: standing behavior is something you say yes to yourself. Schedules pause themselves and tell you why if their creator leaves the team or three runs in a row fail, and owners and managers can have up to ten.

A schedule also remembers what happened to its last runs. Each run starts with a few lines of notes, read from the schedule's own history at the moment it starts:

• How the last run ended: finished, with its one-line summary; stopped, with the reason; or still open.
• What you did with its cards: applied, applied on Autopilot, rejected (and why), undone, or still waiting. A card you decide after the run has ended still counts.
• What not to propose again: any change from the last three runs that you rejected as "not needed", or undid. The assistant leaves those alone unless you ask for them.

So a weekly brief can say what changed since last week, and a morning schedule won't propose again the price change you turned down yesterday. If you edit the schedule's instruction, the memory starts over. When a run read outside content (a customer's message, a bank line), its summary is left out of the next run's notes and only the facts about its cards carry over, so the caution doesn't follow the schedule forever.

Templated emails: the one way the assistant writes to a customer on its own

The assistant never sends free text to a customer. What it can send is a message template the owner or a manager wrote under Settings → Message templates: a subject and a plain-text body with a few placeholders ({{first_name}}, {{last_name}}, {{dealership_name}}, {{salesperson_name}}). {{salesperson_name}} is the salesperson the lead is assigned to; with nobody assigned, and on every message to a customer, it reads "the team at" and your dealership's name, so an email never goes out in the name of someone who did not see it. Only enabled templates count.

Ask for it the way you would ask a colleague ("send Jane the welcome email") and the assistant proposes Send "Welcome" to Jane Doe with the subject and message exactly as she will read them. Nothing on that card is editable; if the words need changing, change the template. At Ask me first the card waits for you, and at Do the routine it still waits. The assistant may send a template on its own only at Run it, or when the owner ticks "Communication: send templated" under Fine-tune.

A fifth stop applies to borrowers: a customer with a live in-house loan gets no templated message before 8 AM or after 9 PM in your dealership's timezone, whoever proposed it. Four more things stop a send before it is proposed and again before it goes out: the person wrote back and nobody answered (the assistant tells you to reply instead), they already received a templated message in the last day, the dealership has hit its daily limit of templated sends, or the template is not enabled. Every send lands on the person's timeline and in the platform's mail history, so one approval can never send twice. Each recipient's own unsubscribe still applies.

Triggers: work that starts when something happens

For work that should start the moment something happens rather than on a clock, the dealership owner can turn on triggers under Settings → AI assistant, in the Autopilot card's "Start work when something happens" section. Each trigger starts one background job when its event happens:

• A new lead comes in: every new lead counts, whether your team typed it in, a shopper sent a website form (a question, a pre-qualification, a trade-in or a test drive request), the storefront assistant took the shopper's details, or it arrived from a marketplace or an ADF feed. The assistant reads the lead and the car they asked about and proposes one first follow-up task for the salesperson, due within a business day. The AI lead handler is on for every plan, so the job loads that skill and runs its first-touch playbook instead (the welcome template card plus the task).
• An appointment is marked a no-show: it proposes one task to reach out and reschedule. The AI appointment concierge is on for every plan, so the job loads that skill and runs its no-show step instead (one reschedule offer as a card plus the task).
• Title work is kicked back by the state: it reads the title record and proposes the correcting step as a task for the person handling titles. The AI title and compliance runner is on for every plan, so the job loads that skill and runs its kick-back step instead.
• An in-house loan payment is returned (NSF): it reads the loan's standing and proposes one task for collections to contact the borrower. The AI BHPH collector is on for every plan, so the job loads that skill and runs its returned-payment step instead.
• A bank line has sat unreviewed for a week: the one trigger that watches for a condition rather than an event. Each morning the assistant checks the bank inbox for posted lines on your matching account that have sat unreviewed for seven days and, for each one (five a morning at most), starts a job that reads the line and the books and proposes one task for accounting to match or categorize it. Once the AI accountant is on for your plan, the job loads that skill and runs its bank inbox step instead. It needs a connected bank feed; with none, nothing happens.
• A new review arrives on your Google listing: also checked each morning, once a day. The assistant looks at your connected Google listing (Website settings) and, for each review that was not there at the last look (three a morning at most), starts a job that reads the latest reviews and proposes one task for you to answer the review on Google, with its stars and what it said in a few words. The AI marketing manager is on for every plan, so the job loads that skill and runs its reputation step instead. The assistant never writes or posts the public reply. Google shows only five reviews of its choosing and none without text, so when the review count rises with nothing new in view, one job says so and asks you to read them on Google. The first look after you turn it on only remembers what is already there. We keep no review text: only the review count and a fingerprint of each review already seen.

A trigger job runs as the assistant under the owner's standing permission, so the audit log shows it as "Assistant (for Jane Doe)" when Jane owns the dealership, never as a staffer who was not there. It follows the same dial as everything else: at Ask me first every task it prepares waits on an approval card; at Do the routine an eligible task may apply on its own, except in a job that started from words your team did not type. A new lead's name, a bank line's description and a review come from outside the dealership, so those jobs start the way a job does after reading outside content, and their cards always wait for a person. A trigger never contacts a customer or a borrower itself; the one exception is a templated email from the lead handler (a welcome) or the appointment concierge (a reschedule offer), which is still a card a person approves because the job read the person's own messages first. The same lead, appointment, title, payment, bank line or review starts at most one job, at most twenty trigger jobs start in a day, and a dealership that is stopped, at Answers only, or out of this month's AI budget starts none. Everything is off until the owner turns it on.

Editing a card that needs approval

A card waiting for approval is editable. You can change the values the AI suggested before approving: a price, a status, a description, options checkboxes, and so on. Money fields accept "18000", "18,000", or "$18,000.50"; they're cleaned up for you. An Autopilot card may already have a final outcome by the time it appears, so it is not presented as a pending edit form once settled.

What you can't change is which record the action points at. The vehicle, lead, or customer being acted on shows as a read-only line on the card. That's on purpose. It means an approval can never be retargeted to a different (or another dealer's) record.

What the AI can do

There are 50 actions today, grouped by area:

• Inventory: create a vehicle (it goes on your primary lot; to put it on another lot, change its Location on the vehicle's Edit page), edit it (its lot, the auction it came from and its walkaround video stay as they are; change those on the Edit page), change its price, change its status, publish it (the card shows how many public photos the car has and whether it has an asking price; a car missing either is never published on its own, even with Autopilot on, so a person sees the gap first), rewrite its description, book a recon, parts, detail or other cost against it, optionally from a named vendor (money, so it always asks; it posts to the books like a cost entered on the vehicle page), add the photos you sent in the chat to a car, up to 25 on one card (private unless the card's switch says public; undoable), add a new car from its photos (the card decodes the VIN itself, you give what you paid and where it came from, and every photo goes on the new car; undoable while nobody has worked on the car), and make private photos public (they go to the website and Facebook and cannot be recalled once syndicated, so it always asks).
• Leads: create (with the same fields as the New lead form: your lead routing picks who gets it, just as for a lead you type in, and a car the person asked about goes in the notes), change status, assign to a salesperson, convert to a customer.
• Customers: create (a new customer's country is US; change it on the customer's Edit page), edit (the country stays as it is).
• Appointments: book a test drive or sales visit for a lead or customer, move it, confirm it, or record the outcome (showed, no-show, cancelled). These always wait for your approval, whatever the Autopilot dial says: an appointment is a promise to a customer.
• Tasks: create, assign, set status, add a comment.
• Deals: create a deal, move a deal through its stages (including mark sold), record a deposit or payment received on a deal (money, so it always asks; the amount may not exceed the balance due, and refunds stay on the deal page), and attach an F&I product such as a service contract or GAP (it raises what the buyer pays, so it always asks; the card shows the deal total before and after).
• Paperwork and e-signature: choose the forms a deal needs (your packs and templates plus the official state forms, named as the Forms card names them; undoable), generate the chosen packs or a named pack as a dated snapshot of the deal's numbers (generate again after a change; a snapshot is never deleted), and send a generated pack for e-signature (an email to the buyer and co-buyer that cannot be unsent, so it always asks, and a pack is sent once; anyone without an email signs in store, and whoever approves the card signs for the dealership); send a signing reminder to whoever has not signed (a fresh link goes out and an email cannot be unsent, so it always asks); and void a pending signing request when the paperwork has to be redone (every link goes dead at once and it cannot be undone, so it always asks).
• Title: move title work through its stages.
• Compliance: re-run the OFAC sanctions screening on a deal (always with your approval). A flagged match is still cleared or confirmed by an owner or manager on the deal's compliance page, never by the assistant.
• BHPH: record a loan payment, and record a borrower's promise to pay (a timeline entry with a follow-up on the promised date; it books no money and changes nothing on the loan).
• Communications: draft and send an email or SMS after explicit approval, send one of your message templates by email (see Templated emails above), and log a call, visit, email, text or note that already happened (always with your approval, because a logged contact counts as working the lead). From the Inbox, it can add an email's sender as a lead, put an email on a lead's or customer's timeline (its attachments are filed on their documents, as when you link it yourself), or archive it. These three always wait for your approval, even with Autopilot on, because the email came from outside the dealership. When a customer asks not to be emailed, it can turn on No account emails on their record, after which nothing is emailed to them from anywhere in the app. This also always waits for your approval, and only a person turns it off again, on the customer's page.
• Website: new copy for one of your custom pages (title, body, meta title, meta description). A new title or body goes to the page's draft, and shoppers see it only after you press Publish on the Pages screen. The meta title and meta description are not part of the draft, the same as when you save them on the Pages screen: on a published page, approving the card changes them on your live site right away, and the card says so before you approve. It can also propose new words for the home page (the hero headline and the line under it, the About text, any section's heading, small line, body, questions, the cards of Why buy from us or the steps of How it works) and the search titles and descriptions under Search results: the card shows the old and new words, approving saves them to the website's draft, the visual builder opens on them, and Publish there is still yours to press. Both kinds of website card can be undone. A website card can wait: approving it changes only the words on the card, over your site as it is at that moment, so whatever you or another card changed in the meantime stays (a page you renamed keeps its new address). If someone changed the same words after the card was made, approving refuses and changes nothing.
• Documents: file a PDF or a photo of a document you attached to your message (in the chat or on Telegram) on a vehicle, deal, customer or lead, as a title, registration, receipt, license copy or any other kind the Documents card knows, with a name and a note you can fix on the card. Nothing is filed until you approve; Undo removes it again.
• Memory: remember a fact for your own future chats.
• Background work: start a background job.
• Schedules: create recurring work (always requires explicit approval).
• Support: escalate an AI-support issue with its conversation reference (always requires explicit approval).

Who can approve what (role gating)

Each action is gated to your role, and the AI only ever offers you actions your role is allowed to perform:

• Owner / Manager: everything.
• Sales: inventory, leads, customers, appointments, tasks, deals and their paperwork, title work, compliance screening, promises to pay, communications, documents, background jobs, personal memory, and support escalation. Not loan payments, schedules, or website copy.
• Accounting: customers, tasks, title work, communications, documents, background jobs, personal memory, support escalation, recording BHPH loan payments and promises to pay, recording payments on a deal, and recording a vehicle cost. Not the rest of inventory, schedules, website copy, or the sales pipeline, so from the Inbox accounting can link or archive an email but not add its sender as a lead.

Saving a personal Memory isn't restricted. Every role can do that, since it only affects your own future chats.

The gate is checked twice: once when deciding what the AI can even offer, and again on every apply attempt, whether manual or Autopilot. So a proposal for something your role can't do is refused before execution with a message like "Your role can't perform this action."

FAQ

Will the AI ever change something without me approving?

Only when the dealership owner has deliberately enabled Autopilot for that eligible routine action. Even then, the application rechecks the live policy, your role, feature access, action risk, proposal values, and injection safeguards before it tries the change. Money, irreversible actions, communications, background-job starts, schedules, and support escalation still require an explicit approval. The one exception among communications is a templated email: it may go out on its own at Run it, or when the owner picks it under Fine-tune.

Why did a job or schedule not run?

Background work follows the assistant's setting. While the assistant is set to Answers only, new jobs are refused with the reason, schedules skip their run without losing their slot, and a job already under way waits before its next step. Turning the assistant off everywhere goes further: its background jobs are cancelled and its schedules are suspended until you turn it back on. A job still waiting a week after it started expires with a note in your bell, so nothing runs weeks later under a setting you changed since.

To see what ran and why, the owner or a manager opens Settings → AI and scrolls to Background work, under Schedules: every job the assistant ran for your team, newest first, with what started it (asked for in chat, a named schedule, or a trigger), where it stands (waiting to start, working, waiting for your approval, done, stopped or cancelled) and, when it stopped, why, in plain words. Open takes you to the conversation with its work and its cards.

Can I undo an approval?

Usually, yes. An applied card shows Undo for changes the assistant can put back: publishing or unpublishing a vehicle, a description, a price or detail edit, a task or lead assignment or status move, a vehicle or title status move that did not close the record, a comment, a memory, the forms chosen for a deal, new words in your website's draft or on a custom page, an Inbox email it archived, No account emails it turned on, and a task, lead, vehicle or schedule it created. Undo restores exactly what the assistant changed and refuses if someone has changed the record since, so it never overwrites later work. That includes your own changes on the card: if you edited or emptied a field before approving, Undo puts that field back too. For the website, Undo takes the assistant's words back out of the draft and leaves the rest of the draft alone, a scheduled publish included; once those words are published, change them in the visual builder instead. On a custom page, Undo of a meta title or description puts the old words back at once, with no Publish, the same way approving changed them. It works even while the assistant is stopped.

Some changes have no automatic undo and the card says so: creating a customer or a deal, converting a lead, moving a vehicle to wholesaled or disposed, completing or cancelling title work, re-running an OFAC screening, generating a forms pack, sending a templated email, adding an Inbox sender as a lead or linking an Inbox email onto someone, starting a job, and support escalations. Reverse those from the record's own page. Rejecting a card only stops a pending proposal; when you reject one you can say why (not needed, wrong details, not now), which helps us tune what the assistant proposes.

I clicked Approve and it said the action can no longer be applied.

A proposal can become unavailable if AutoDealer.io disabled AI, your dealership stopped its assistant, billing is blocked, your role changed, or the assistant is set to Answers only. The message tells you which state applies. An owner or manager can turn the dealership's assistant back on under Settings → AI assistant; platform and billing blocks must be cleared in the place named by the message.

The pending card stays in the conversation. After the blocking state is lifted, review the card again before approving it. A plain "yes" still follows the latest-message and 24-hour safety rules.

A card also checks its record again when you approve it. If the record has moved on since the card was made (the appointment was cancelled or its time has passed, the signing request was completed or voided, a signer already signed, the deal was cancelled), the card tells you why in red and nothing changes. The card keeps its Approve and Reject buttons, so you can dismiss it or ask the assistant for a new one. The same happens when you reply "yes" in chat: the reason shows under the conversation and the card keeps waiting.

A website card also refuses when the words it changes were changed after it was made, in the visual builder, on the Pages screen or by another card you approved. The message names those words ("The hero headline has changed since this card was made, so nothing was changed"), nothing on the site changes, and your own words stay. Ask the assistant for a new card, which starts from the site as it is now. An older website card may say it is out of date instead: approving it could undo later changes, so ask for a new one the same way.

What does "Edited values are invalid" mean?

A value you typed on the card didn't pass validation, and the message names the field. Fix that field and Approve again; the rest of the proposal is preserved.

What does "Needs review" mean?

AutoDealer.io could not prove whether an action finished after execution may have started. The card is locked so the same effect cannot be repeated. Choose Reload status, then check the affected record and activity trail. If the card still says Needs review, do not approve or re-create the action. New AI changes are paused dealership-wide until the receipt is reconciled, so contact support after you confirm what happened.

A card refused before anything ran never shows Needs review. It says why in red and keeps Approve, as described above. You can also see Needs review when you click Approve on an old copy of a card that was already decided somewhere else, in another tab or on Telegram. Reload status then shows what happened to it. If that is Applied or Dismissed, there is nothing more to do.

Do decisions and Autopilot applications leave a record?

Yes. Approved and rejected actions are written to the audit log. An automatic application is shown as the assistant acting for the person whose standing permission allowed it, for example "Assistant (for Jane Doe)", so you can tell it apart from a manual approval, which is logged under the name of the person who approved it.

Can the AI act on another dealer's records?

No. Everything is scoped to your dealership, and the record an action targets is locked on the card and can't be retargeted.