A8gent
All articles

Industry

Auto Repair Estimate Approval Automation: On the Lift

Declined work costs a shop money slowly. An unanswered estimate costs it money right now, because a bay is occupied, a technician is idle and the parts order has not gone in. This is a latency problem, not a marketing one - and the record you build while solving it may also be your legal authorization.

A8gentAutomation practiceAugust 5, 202613 min read
Contents
  1. This Is a Latency Problem, Not a Conversion Problem
  2. The Messages May Be Your Legal Record
  3. Why Approvals Actually Stall
  4. Escalation Design: Different Channel, Not Louder
  5. Getting the Trigger Out of the Shop Management System
  6. The Bay-Hour Arithmetic, Shown
  7. What to Watch Once It Is Running
  8. FAQ

This Is a Latency Problem, Not a Conversion Problem

The two big follow-up problems in an auto shop look similar and behave nothing alike.

Declined work - covered in our declined work follow-up article - is slow. The customer left, the money went with them, and you have weeks or months to get some of it back. Nothing is on fire.

An unanswered estimate is fast and expensive. The car is on the lift. The technician has already spent time diagnosing and is now either standing around or has moved to another job, which means the first car will need to be picked up again later. The parts order is blocked. The bay - which is the shop's actual production capacity - is occupied by a vehicle producing no revenue. Every hour of silence has a cost, and unlike declined work, that cost is incurred whether or not the customer eventually says yes.

That reframing changes what you build. You are not writing persuasive copy. You are minimising the time between "the estimate is ready" and "we have an answer," and treating a fast no as almost as valuable as a fast yes - because a no releases the bay and lets you get the car off the lift and the next one on.

State it as the metric and everything downstream gets easier: median and 90th-percentile hours from estimate sent to authorization recorded. Not approval rate. Approval rate as a target rewards pressure, discounting and ambiguity; latency as a target rewards clarity, reachability and good escalation. Pick the one that makes your shop better at the thing customers actually complain about.

For where this sits among a shop's other automation candidates, see AI automation for auto repair shops.

Sourcing note. Regulatory text below is quoted from the Cornell LII mirror of the California Code of Regulations with URLs inline; vendor claims are quoted from vendor documentation, and where a vendor's documentation refused our requests we say so. California is used as the worked example because its rules are unusually explicit and well published - your state's requirements will differ and you must check them. None of this is legal advice.

The Messages May Be Your Legal Record

This is the section that separates a proper build from a chatbot, and almost nothing currently ranking for this topic mentions it.

In California, 16 CCR 3353 states that "An estimate shall be provided to and authorized by the customer before any work commences." Authorization is not a courtesy step in the workflow. It is the thing that makes the work lawful and the invoice collectable.

16 CCR 3353.1 then specifies how that authorization must be documented. For oral authorization, "the automotive repair dealer shall record the authorization by documenting on the estimate the date, time, name of the person authorizing the repairs, and the telephone number called, if any." For electronic authorization, the dealer must document "the date, time, name of the person authorizing the repairs, and the telephone number or electronic mail address contacted, if any."

And critically, the regulation permits a dealer to "produce this information on documents relating to the authorization that supplement the estimate," which it says can include "a series of electronic communications between the automotive repair dealer and the customer."

Read that last clause carefully, because it has direct architectural consequences. Your SMS thread is not a marketing artefact that can be rotated out after 90 days by whatever messaging platform you signed up for. It is potentially the supplementary documentation of a legally required authorization. That means:

  • Retention is a requirement, not a setting. Find out your messaging vendor's retention period and whether you can export. A platform that deletes threads on a schedule you do not control is unsuitable for this workflow regardless of how good its automation features are.
  • You must capture identity, not just a phone number. The regulation asks for "the name of the person authorizing the repairs." A "Y" from a number on file is weaker than a "Y" from a named person on a number on file. Design the exchange so a name is recorded, especially where the caller is a spouse, a fleet manager or an employee.
  • Timestamps must be authoritative and preserved. Date and time are explicitly required. Store them from the messaging provider's record, not from your workflow's clock.
  • The automation must never infer authorization. A model that reads "sounds fine, whatever you think" and marks an estimate authorized has manufactured a legal record. Authorization should be recorded only from an explicit, unambiguous customer response, and anything else routes to a human. This is a hard rule in code, not a line in a prompt - see guardrails.

Two honest limits on the above. First, California is one state and its rules are unusually explicit; every state differs and some are far less prescriptive. Second, we are describing published regulatory text, not advising you on compliance - your shop's obligations are a question for your own counsel and your state board. What we are confident about is the design implication, which holds everywhere: build the approval flow as an audit trail that happens to be a conversation, rather than a conversation you hope is an audit trail.

Why Approvals Actually Stall

Before automating the chase, understand what you are chasing. The reasons are boringly practical and each has a different fix.

The customer is at work and cannot look at their phone. The single largest cause, and the one that makes fixed-interval reminders useless. A person in a warehouse, a classroom, a hospital or a cab is unreachable between 8 and 5, which is precisely the window your shop operates in. The fix is not more messages; it is asking at drop-off when and how they can be reached, and honouring the answer.

They do not understand what is being asked. An estimate that lists "R&R front strut assembly, alignment, sublet" to somebody who is not in the trade is not a decision they can make. They will delay rather than admit confusion. The fix is the inspection photo and one plain sentence about consequence.

They cannot approve alone. A spouse, a fleet manager, a business owner, a warranty administrator. The fix is asking at drop-off who authorises, and capturing that person's contact details then - not discovering it three hours into the silence.

The number is wrong or the message did not arrive. More common than anyone expects, and completely invisible from the shop's side. A number that has been ported, a message filtered as spam, a customer who blocks unknown numbers. The fix is delivery-status monitoring and cross-channel escalation.

They are deciding, and deciding takes time. A $2,400 repair on a car worth $5,000 is a genuine deliberation. The fix is a message that acknowledges it and offers the actual decision - repair, defer, or do the safety-critical portion only - rather than a yes/no that forces a no.

Notice how many of these are solved at drop-off rather than at chase time. Two questions at the counter - "what's the best way to reach you between nine and five?" and "is there anyone else who needs to approve this?" - eliminate more approval latency than any messaging sequence will. Automate the capture of those answers before you automate the chasing.

Escalation Design: Different Channel, Not Louder

The default build sends a text, waits an hour, sends another text, waits, sends a third. If the phone is in a locker, that is three failures and one annoyed customer at 5pm.

Escalate across channels and across people, not across volume. A workable ladder:

  1. Immediately: the estimate goes out on the channel the customer specified at drop-off, with the inspection evidence attached and one clear action per option.
  2. After a short interval - shorter for a car on a lift than a car in the lot - a single reminder on the same channel. Confirm the first message was actually delivered before sending it; if delivery failed, skip straight to step 3.
  3. Next: a different channel. If you texted, email. If you emailed, text. This is where most stalls break, because the failure was reachability rather than reluctance.
  4. Then: the secondary contact captured at drop-off, if there is one and the customer authorised contacting them.
  5. Then: a human phone call from the advisor. Not the automation. At this point the value of a person exceeds the value of consistency, and the advisor can resolve in ninety seconds what six messages could not.
  6. Finally: an internal escalation. The vehicle is flagged, the service manager is told, and a decision gets made about the bay - move the car, release the lift, reschedule. This step is the one shops skip and it is the one that protects capacity. The purpose of the ladder is not only to get an answer; it is to stop pretending an answer is coming.

The intervals should be driven by what the car is blocking, not by a constant. A vehicle on a lift with a technician assigned needs the ladder compressed into a couple of hours. A vehicle parked outside awaiting parts can sit on a much gentler cadence. If your data lets you distinguish the two - and the shop management system usually does, via job status - use it. That single conditional is worth more than any amount of copy tuning.

Two hard limits. Respect quiet hours, even when a bay is blocked; texting at 9pm about a lift is your problem, not the customer's. And cap the ladder: after the internal escalation, the automation's job is done and the situation belongs to a human being.

On the channel itself, if you send SMS over a US 10-digit long code you are inside A2P 10DLC. Twilio's compliance documentation describes it as "the standard that United States telecom carriers have put in place to ensure that SMS traffic to US end users through long code phone numbers is verified and consensual," requiring a registered Brand and Campaign, and notes that unregistered senders "will receive additional carrier fees for sending unregistered traffic." Throughput to T-Mobile runs from 1,000 SMS segments per day for a Sole Proprietor brand up to 2,000 for Low-Volume Standard and beyond for Standard brands "depending on your Trust Score" - not a constraint at estimate volumes, but registration is not optional. See our A2P 10DLC guide and the glossary entry.

Getting the Trigger Out of the Shop Management System

The automation needs to know when an estimate goes out and when it is answered. That is a vendor question.

Shopmonkey publishes developer documentation at shopmonkey.dev. Its webhook documentation lists events across Appointment, Customer, Inspection, Inventory, Message, Order, Payment, PurchaseOrder, User, Vehicle and Vendor, delivered as HTTPS POST with a JSON payload containing apiVersion, operation, table and data, plus a diff array on updates, with operations classified as INSERT, UPDATE or DELETE. That diff is the useful part here - a status transition on an Order is exactly the trigger this workflow needs, and the diff tells you what changed rather than making you infer it.

The documented operational semantics matter as much as the events:

  • Signing uses HMAC-SHA256 as HMAC256(msgId.timestamp.payload,secret) with a v1,signed header format. Verify it. An unauthenticated endpoint accepting repair order data is an open door.
  • Success requires HTTP 200, 201 or 202; 429 and 500 trigger retries, tracked by an X-Attempt header. Your handler must therefore be idempotent - see idempotency.
  • The response timeout is 55 seconds. Acknowledge immediately and process asynchronously; never call a model and a messaging provider inline from the webhook handler. See queue.
  • "After a certain number of redelivery failures, the Event will be marked failed and your Webhook will be temporarily disabled." A bad deploy can silently switch your integration off, and a disabled webhook looks exactly like a slow day. Alert on the absence of events. See observability and error handling.

We could not get the base URL, authentication specifics, rate limits or pagination to render from the Shopmonkey docs homepage - the navigation loaded, the technical detail did not - so confirm those on their Quickstart and Overview pages rather than from this article.

Tekmetric is gated. Its API documentation at shop.tekmetric.com/api/docs returned HTTP 403 to our requests and we found no public developer reference. The independent api-evangelist catalogue describes an OAuth 2.0 client credentials API at shop.tekmetric.com/api/v1 with a sandbox, states that "Tekmetric does not publish a self-serve developer portal or a public API reference," and characterises the available endpoints as read-only - a third-party summary, not a Tekmetric statement. If read-only is accurate, the consequence for this build is serious: you could detect that an estimate is awaiting approval and be unable to write the authorization back, which for a workflow whose output is a legal record is a significant limitation. Confirm write access with Tekmetric directly before scoping.

If there is no usable integration, the advisor triggers the sequence with a button when they send the estimate. That is a legitimate v1: the value is in the escalation ladder and the recorded thread, not in the trigger being automatic. Ship it, measure the latency improvement, and buy the integration later with the evidence in hand. See polling vs webhook for the middle option.

The Bay-Hour Arithmetic, Shown

Illustrative. We report no client results - replace every figure with your own, all of which your shop management system can report.

What an hour of bay time is worth. Take effective labour rate, say $145/hour, and the shop's realistic productive hours. This is not the same as revenue per hour, and using the billing rate overstates the case - what an unblocked bay actually earns you is the gross profit on the labour and parts it would have produced, less nothing you were going to pay anyway. Suppose gross profit per productive bay hour is $85.

The delay. A four-bay shop writes 40 estimates a week that require customer authorization. Measure current median time-to-authorization - assume it comes back at 4.2 hours. Of that, suppose 2.6 hours is genuinely blocking: a bay or a technician is idled rather than the car simply sitting in the lot.

Weekly blocked bay hours: 40 x 2.6 = 104 hours. That headline is misleading on its own, because a shop absorbs some of it - the technician moves to another job. Suppose only 30 percent of blocked time is truly unrecoverable: 104 x 0.30 = 31.2 genuinely lost bay hours a week. At $85 gross profit per hour, that is 31.2 x $85 = $2,652 a week, or roughly $132,600 a year on 50 weeks.

What automation can realistically remove. Not all of it. The customer at work until 5pm is unreachable no matter how good your ladder is. Suppose better routing, cross-channel escalation and drop-off contact capture cut median blocking time from 2.6 hours to 1.7 - a 35 percent reduction, which is meaningful and not heroic. Recovered: 31.2 x 0.35 = 10.9 bay hours a week, at $85 = $927 a week, or $46,350 a year.

Against cost. Build at $3,500 plus $80/month gives a first-year cost of $3,500 + $960 = $4,460 - replace with your own quotes; automation build cost covers the drivers. Payback: $4,460 / $927 per week = 4.8 weeks.

Now the honesty. That figure is dominated by two assumptions you must measure rather than accept: the 30 percent unrecoverable share, and the 35 percent latency reduction. If your technicians are good at switching jobs and only 10 percent is truly lost, the annual figure drops to $44,200 and the recovered amount to $15,470 - still a good build, on a very different scale. If your shop is small enough that the advisor already chases every estimate within twenty minutes, the improvement may be near zero and you should spend the money elsewhere.

Measure median time-to-authorization for two weeks before you build anything. It costs nothing, it is the metric the project will be judged on, and it is the only way to know whether there is a problem to solve. Our ROI of AI agents piece covers why unmeasured baselines are where automation business cases go wrong, and payback period the framing. For general budget ranges see AI automation cost.

What to Watch Once It Is Running

Report four numbers. Median and 90th-percentile time-to-authorization; the proportion of estimates resolved at each rung of the escalation ladder; delivery failure rate by channel; and the count of vehicles that hit internal escalation. That third one is a data-quality metric masquerading as a delivery metric - a high SMS failure rate almost always means bad phone numbers at intake, which is a counter problem you can fix cheaply.

The characteristic failure modes:

  • Inferring approval. The worst one, for the reasons in the regulatory section above. Any ambiguous reply must route to a human and must not write an authorization. Test this deliberately with replies like "ok whatever you think" and "yeah do what's needed" before go-live.
  • Chasing a customer who already approved by phone. Happens when authorization recorded verbally by the advisor does not stop the automated sequence. If your shop management system is read-only to you, you need an explicit way for the advisor to kill the sequence - a button, and one they will actually press.
  • The ladder that never ends. Without a terminal internal escalation, vehicles sit in an automated loop for days while everyone assumes the system is handling it. Cap it and hand it to a person.
  • Losing the thread. Changing messaging platforms, or a retention policy quietly rotating out the record. Given the authorization documentation point, export and archive.
  • Quiet-hours violations from a compressed ladder crossing into the evening. Clamp it.

Two adjacent builds worth doing next, deliberately and separately. Declined work follow-up - the slow sibling of this problem - is in declined work follow-up for auto repair shops. And if the underlying constraint turns out to be that nobody answers the shop phone, a missed-call text back is far cheaper than either; what a full voice layer costs is on our AI receptionist cost for auto repair shops page.

If you want a second opinion on which of these to build first, the free automation audit works through it, and automation services explains how we scope work like this.

FAQ

Can an automation record a customer's authorization for repairs?
It can record it; it must never infer it. California's 16 CCR 3353.1 requires that electronic authorization be documented with "the date, time, name of the person authorizing the repairs, and the telephone number or electronic mail address contacted, if any," and expressly contemplates "a series of electronic communications between the automotive repair dealer and the customer" as supplementary documentation. So the thread can be the record - but only an explicit, unambiguous customer response should ever write an authorization. Anything vague routes to a human. Your state's rules will differ; check them.
Does that mean my text messages have retention requirements?
Potentially, yes, and it is worth establishing before you choose a messaging platform. If the message thread is functioning as supplementary documentation of a legally required authorization, a platform that rotates threads out on a schedule you do not control is unsuitable regardless of its automation features. Check your vendor's retention period, confirm you can export, and archive the threads yourself. Confirm your actual obligations with your own counsel.
How quickly should I chase an unanswered estimate?
Drive the interval from what the car is blocking, not from a constant. A vehicle on a lift with a technician assigned justifies compressing the whole escalation ladder into a couple of hours. A vehicle parked outside awaiting parts can sit on a far gentler cadence. Most shop management systems expose job status, and that single conditional does more for your results than any amount of copy tuning.
Why not just send more reminders?
Because the most common reason for silence is that the customer is at work and physically cannot look at their phone, and a fourth text does not solve that. Escalate across channels and across people instead: same channel once, then a different channel, then the secondary contact captured at drop-off, then a human call from the advisor, then an internal escalation that makes a decision about the bay. Volume on one channel is the least effective response to a reachability problem.
What should I capture at drop-off?
Two questions that remove more approval latency than any messaging sequence: what is the best way to reach you between nine and five, and is there anyone else who needs to approve this. A spouse, fleet manager or warranty administrator discovered three hours into the silence is three hours you cannot get back. Automate the capture of those answers before you automate the chasing.
Which shop management systems can I trigger this from?
Shopmonkey publishes usable webhook documentation covering Appointment, Customer, Inspection, Inventory, Message, Order, Payment, PurchaseOrder, User, Vehicle and Vendor, with a diff array on updates that makes status transitions a clean trigger. Tekmetric's API documentation returned HTTP 403 to us and has no public developer reference; a third-party catalogue describes it as partner-gated and read-only, which if accurate would prevent writing an authorization back. Confirm write access directly with Tekmetric before scoping a two-way build.
Can I build this without any integration at all?
Yes, and it is often the right v1. The advisor presses a button when the estimate goes out, and the automation runs the escalation ladder and keeps the record. Almost all of the value in this workflow is in the ladder and the audit trail, not in the trigger being automatic. Ship that, measure the change in median time-to-authorization, and use the evidence to justify the integration later.
What metric should the project be judged on?
Median and 90th-percentile hours from estimate sent to authorization recorded, measured against a baseline you collected before building. Not approval rate - targeting approval rate rewards pressure and ambiguity, and pressure in a workflow that produces a legal authorization record is precisely the wrong incentive. Track the escalation rung at which estimates resolve and the delivery failure rate by channel as your diagnostics.

If this resonated

Module 02 is the long version of this post.

Eval set, refusal pattern, kill switch. Build it in a weekend.

All postsLast verified 2026-08-05

Done for you

Want this built for your business?

Tell us the workflow. We scope, build, and ship the agent with guardrails and the numbers to prove it worked. The scoping call is free.

Free scoping call. You own the code.

Request a scoping call