Contents
"Declined" Is Almost Never a Rejection
The word does real damage. A shop management system records the line item as declined, the advisor experiences it as a no, and the follow-up - if it happens at all - is built as though it were trying to overturn a decision.
That is not usually what happened. What happened is one of a small number of things, and they need different responses:
- "Not this paycheck." The most common by far. The customer agrees the work is needed and cannot fund it today. The correct follow-up is timed to their circumstances and offers a way to pay, not a discount.
- "I don't understand why I need that." The advisor explained it in a busy lane in ninety seconds while the customer was thinking about getting back to work. The correct follow-up is the inspection photo and the measurement, not a repeat of the price.
- "I want to check that." The customer intends to ask a friend, a forum, or another shop. The correct follow-up arrives after they have done so, and it is the one that decides where the work gets done.
- "I'm selling the car." A genuine no. The correct follow-up is none, and your system needs to be able to record that so it stops.
Notice that three of those four are timing and information problems, and only one is a real rejection. A follow-up system that treats them identically will convert the easiest cases and irritate the rest. A system that captures why at the moment of deferral - one tap by the advisor, four options - is dramatically more useful than one with better message copy, and it costs almost nothing to build. If you take one thing from this article, take that: capture the reason at the counter, because it is unrecoverable afterwards.
The equivalent problem in the trades - a quote sent, no answer, everybody assumes a no - is covered in our unsold estimate follow-up guide and automate quote follow-up, and the underlying psychology is identical. For where this sits among a shop's other automation candidates, see AI automation for auto repair shops.
What Your Shop Management System Actually Exposes
Everything here depends on getting the deferred line items out of the shop management system, and the vendors differ sharply in how much they will help.
Shopmonkey
Shopmonkey publishes genuine developer documentation at shopmonkey.dev, and its webhook documentation is detailed enough to design against. Read in August 2026, it states that webhooks are available for the following resources: Appointment, Customer, Inspection, Inventory, Message, Order, Payment, PurchaseOrder, User, Vehicle, and Vendor. Subscriptions are created in Settings by naming the webhook, supplying an endpoint URL and selecting "one or more events for delivery."
The operational semantics are documented, which is the part that matters for reliability:
- Payloads are HTTPS POST with JSON containing
apiVersion,operation,tableanddata, with adiffarray on updates. - Operations are classified as INSERT, UPDATE or DELETE.
- Signing follows the Webhook Standard specification using HMAC-SHA256, computed as
HMAC256(msgId.timestamp.payload,secret), with a header format ofv1,signedwhere the signature is base64-encoded. - Success requires an HTTP 200, 201 or 202. 429 and 500 trigger retries. An
X-Attemptheader tracks delivery attempts. - The response timeout is 55 seconds.
- "After a certain number of redelivery failures, the Event will be marked failed and your Webhook will be temporarily disabled."
Three design consequences fall straight out of that. First, verify the signature - an unsigned endpoint accepting order data is an open door. Second, acknowledge fast and process asynchronously: 55 seconds sounds generous until your workflow is calling a language model and an SMS provider inline, so return 200 immediately and queue the work. Third, that last quoted line is the one that will bite you - a run of failures does not just lose events, it disables the subscription, and a disabled webhook looks exactly like a quiet week. Alert on the absence of events, not only on errors. See webhook, queue and observability.
The Shopmonkey docs also list documented resource groups including Appointment, Customer, Order, Vehicle, Payment, Invoice, Message, Workflow and various inventory types. We could not get the base URL, authentication specifics, rate limits or pagination to render from the docs homepage - the navigation loaded but the technical detail did not - so confirm those on the Quickstart and Overview pages rather than taking a shape from here.
Tekmetric
Different posture, and you need to know before you scope. We attempted to read Tekmetric's API documentation at shop.tekmetric.com/api/docs and received HTTP 403 Forbidden. There is no public developer reference we could retrieve.
The independent api-evangelist catalogue on GitHub describes a Tekmetric API using an "OAuth 2.0 client credentials grant" against https://shop.tekmetric.com/api/v1 with a sandbox at https://sandbox.tekmetric.com/api/v1, states that "Tekmetric does not publish a self-serve developer portal or a public API reference," that access "requires prior approval from Tekmetric," and characterises the documented endpoints - shops, customers, vehicles, repair orders, jobs, employees, appointments, inventory, canned jobs - as read-only. That is a third-party catalogue, not a Tekmetric statement, and we present it as such. But if it is even approximately right, the implication for your build is significant: you may be able to read declined jobs and unable to write anything back, which makes the follow-up system a separate record-keeper rather than an extension of the shop management system.
Confirm two things with Tekmetric directly before quoting any work: whether you can get access at all on your plan, and whether write operations exist for the objects you need. Do not design a two-way build on a third-party summary, including ours.
If you have neither
Shops on Mitchell 1, Protractor or an older system may have no practical API path. The fallback is a scheduled report export, or the digital inspection tool - which is frequently a separate product with its own integration surface and is, as the next section argues, where the valuable data lives anyway.
The Digital Inspection Is the Asset, Not the Estimate
Most declined-work follow-ups send a price. That is the weakest possible version of the message, because price is the thing the customer already said no to.
What changes the conversation is evidence, and the shop already produced it. A digital vehicle inspection generates a photograph of the actual component on the actual car, a measurement, and a technician's note. A message built around that is not a sales follow-up - it is a status report on the customer's own property, and it is received completely differently.
Consider the difference between these two follow-ups on the same deferred item:
"Hi Dave, following up on the brake service we quoted at $480. Can we book you in?"
"Hi Dave - checking in on the Accord. When you were in on 14 July we measured the front pads at 3mm; the photo is here [link]. New pads normally start around 11mm and we recommend replacing at 3. Nothing has changed since, but it will keep wearing. Want me to hold a slot?"
The second one costs the same to send. It requires that your automation can retrieve the inspection item, its measurement and its photo URL - which is exactly why the Inspection resource in the Shopmonkey webhook list is more interesting than the Order resource for this build.
Two cautions. Never let a language model invent or restate a measurement. Pull the number from the record as a literal value and template it in. A hallucinated brake measurement is a safety claim about a customer's vehicle made by a machine, and there is no version of that which ends well. The model, if you use one at all here, should be writing the sentence around the fixed values, and even that is optional.
And never imply an urgency the technician did not record. If the note says "monitor," the message says monitor. Escalating severity to drive conversion is both a trust problem and, for safety-critical items, considerably worse than that. Our when not to use an AI agent piece covers the general boundary.
Timing: The Part Everyone Gets Wrong
The default build sends a follow-up seven days after the visit and another at thirty. It is better than nothing and it is not thinking.
Key the clock to the reason, captured at the counter.
- Cost deferral - the customer said they could not afford it now. Following up in seven days insults them; nothing has changed. Following up shortly before their likely next pay cycle, or two to four weeks out, is a different message. This is also the only bucket where a financing option is genuinely useful rather than a discount in disguise.
- Not convinced - follow up quickly, within two or three days, while the visit is still recent, and lead entirely with the inspection evidence. Speed matters here because you are competing with the friend who says "they always try to sell you something."
- Wants a second opinion - follow up in about a week, after they have had it, and make it easy to come back without loss of face. This is the highest-stakes bucket and the one most worth a human phone call rather than a text.
- Genuine no - suppress permanently, and make sure the suppression survives the next system migration.
Then key the interval to the item. Wear items have a rate. A pad at 3mm has a knowable remaining life in miles; a weeping shock does not change on a weekly clock. Where your inspection data supports it, schedule the follow-up around when the item plausibly becomes urgent rather than around a calendar constant. This is the version that stops feeling like marketing and starts feeling like a service, because it is one.
Cap the total. Two or three touches on a single deferred item, then stop and move the vehicle to the ordinary service-reminder cycle. A shop that texts about the same brake job six times has taught that customer to ignore the shop's number, which costs more than the brake job.
Suppress on any subsequent visit. If the car came back and the work was done, or the item was re-inspected, the follow-up must die immediately. Nothing destroys credibility faster than a message about work the customer already paid you to do. This is the single most common embarrassing bug in these builds, and it is caused by the follow-up living in a separate system from the repair orders - which, per the Tekmetric read-only question above, may be exactly your situation. Check for a newer order on the vehicle before every send.
A Worked Example, With the Arithmetic Shown
Illustrative. We report no client results. Replace every input with numbers from your own shop management system - all of them are reportable.
Baseline. A four-bay shop writes roughly 60 repair orders a week. Suppose 55 percent of them carry at least one declined or deferred line: 60 x 0.55 = 33 orders a week with deferred work. Suppose the average deferred amount on those orders is $340 - your figure, from your own reports.
Weekly deferred value: 33 x $340 = $11,220. Annualised over 50 working weeks: $561,000 of quoted work that walked out of the building. That number is large, it is real, and it is also the number every vendor in this space will wave at you. Do not treat it as the opportunity. Most of it was never going to convert.
What the shop already recovers. Some of that comes back anyway - the customer returns in three months and does the brakes. Suppose 18 percent returns without any follow-up system: $561,000 x 0.18 = $100,980. The automation is not competing for the $561,000. It is competing for an increment on top of the $100,980, and any business case that does not subtract the existing recovery rate is overstating itself by a factor of five. Measure your existing recovery rate before you build. It is the single most important input and almost nobody knows it.
Break-even. Suppose the build costs $3,500 once plus $80/month, replaced with your own quotes - automation build cost covers the drivers. First-year cost: $3,500 + (12 x $80) = $4,460. Shop gross profit margin on the work, say 55 percent, so recovered revenue converts to gross profit at $0.55 on the dollar.
Required incremental revenue to break even: $4,460 / 0.55 = $8,109. Against the $561,000 of annual deferred value, that is 8,109 / 561,000 = 1.45 percent incremental recovery. Not 1.45 percent conversion - 1.45 percentage points on top of whatever the shop already recovers unaided.
That framing is the whole point of doing the arithmetic. "Will an evidence-led, reason-timed follow-up on deferred work move recovery by one and a half percentage points?" is a question a service manager can answer with confidence. "Will this software make you $561,000?" is not a question at all.
One sensitivity that matters more than the rest: if your advisors already call every deferred job, the incremental case is much weaker and the automation's real value is consistency and advisor time rather than recovery. Say so in the business case. Our ROI of AI agents piece covers why this distinction is where most automation cases quietly cheat, and payback period the framing.
A Build Order That Ships
1. Capture the reason. Before any integration, before any messaging: four buttons on the advisor's screen or four options on a paper sheet. Two weeks of this data will tell you more about your shop than the rest of the project combined, and it is the input every later stage depends on.
2. Get the list out, read-only. Whatever path is available - webhook, API, scheduled export. Produce a weekly list of deferred items with vehicle, customer, item, measurement, amount and reason. Send it to a spreadsheet. Have the service manager mark every row that should not be there. Fix the query.
3. Manual send, automated list. The system drafts; a human sends. This is a real, shippable stage. Some shops should stop here, particularly small ones where the advisor's relationship with the customer is the product.
4. Automate the send, with the pre-send checks. Before every message: has the vehicle been back since? Has the item been re-inspected? Is the customer suppressed? Has this item already had its cap of touches? Is there a newer order? Any yes means no send. Key each send on order ID plus item ID plus attempt so a retried run cannot double-send - see idempotency.
5. Handle the reply properly. A yes should produce a booking or a callback task with the item pre-attached, so the advisor is not hunting. Anything ambiguous goes to a human. A "no, I sold it" should permanently suppress. Route on keywords deterministically rather than trusting a model to classify - the cost of a misclassified opt-out is not symmetric with the cost of a false alarm.
6. Only then consider a voice layer. The second-opinion bucket genuinely converts better by phone, and inbound replies create call volume the front counter may not be able to absorb. What that costs is broken down on our AI receptionist cost for auto repair shops page. If the shop's underlying problem turns out to be unanswered calls rather than un-followed-up work, fix that first - a missed-call text back works on demand that is already trying to reach you and is a fraction of the build.
What to measure. Incremental recovery rate against your measured pre-build baseline, gross profit on recovered work, and the opt-out rate. That last one is your early-warning system: a rising opt-out rate means the cadence or the tone is wrong, and it will show up weeks before the revenue does. Do not report messages sent.
If you want a second opinion on whether this is the right first build for your shop, the free automation audit works through the alternatives, automation services explains how we scope it, and AI automation cost gives budget context. The sibling build - chasing quotes that were never answered at all - is in estimate approval automation.
FAQ
- Does Shopmonkey have a public API and webhooks?
- Yes. Shopmonkey publishes developer documentation at shopmonkey.dev, and the webhook documentation lists events for Appointment, Customer, Inspection, Inventory, Message, Order, Payment, PurchaseOrder, User, Vehicle and Vendor. Payloads are JSON over HTTPS POST containing apiVersion, operation, table and data, with a diff array on updates, signed with HMAC-SHA256 as HMAC256(msgId.timestamp.payload,secret). We could not get the base URL, auth specifics or rate limits to render from the docs homepage, so confirm those on their Quickstart page.
- What are the Shopmonkey webhook gotchas?
- Three documented ones. The response timeout is 55 seconds, so acknowledge with a 200 immediately and process asynchronously rather than calling a model and an SMS provider inline. Retries fire on 429 and 500, with an X-Attempt header tracking attempts, so your handler must be idempotent. And the docs state that "After a certain number of redelivery failures, the Event will be marked failed and your Webhook will be temporarily disabled" - meaning a bad deploy can silently switch your integration off. Alert on the absence of events, not just on errors.
- Can I integrate with Tekmetric?
- Probably, but not on a self-serve basis, and you should confirm the details directly. Tekmetric's API documentation 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, gated behind partner approval, and characterises the endpoints as read-only - but that is a third-party summary, not a Tekmetric statement. Before quoting any two-way build, confirm with Tekmetric whether you can get access on your plan and whether write operations exist.
- How long should I wait before following up on declined work?
- It depends on why it was declined, which is why capturing the reason at the counter matters more than any message copy. A cost deferral should wait weeks and land near a pay cycle. A customer who was not convinced should hear from you within two or three days, led by the inspection photo. Someone seeking a second opinion is best contacted after about a week, ideally by phone. A genuine no should be suppressed permanently. A single seven-day rule applied to all four gets one of them right.
- Should the follow-up include the price?
- Lead with the evidence, not the price. The price is the thing the customer already declined; the measurement and the photograph from the digital inspection are new information, and they reframe the message from a sales chase into a status report on the customer's own vehicle. Include the price by all means, but after the reason, and never restate a measurement through a language model - pull the literal value from the inspection record and template it in.
- What is the most embarrassing bug in these systems?
- Texting a customer about work they have already had done. It happens when the follow-up system lives separately from the repair orders and does not re-check the vehicle before sending - which is a real risk if your shop management system gives you read-only access and no reliable event on completion. Add a pre-send check for any newer order or re-inspection on that vehicle, and treat a failed check as a hard stop rather than a warning.
- Will this work if my advisors already call every deferred job?
- The revenue case will be much weaker, and you should say so up front. Where advisors already chase consistently, the automation's value is consistency on the days they are slammed, advisor time returned, and a record of what was said - not incremental recovery. That can still be worth paying for, but present it honestly as a capacity and reliability gain rather than as new revenue, or the project will be judged against a number it was never going to hit.
- Can an AI agent handle the replies?
- The clean ones, yes - a booking request, a request for a different day, an opt-out. Route anything ambiguous, anything mentioning a safety concern, and anything about price negotiation to a human. Keyword routing should be deterministic rather than model-driven, because the cost of misclassifying an opt-out or a safety complaint is not symmetric with the cost of an unnecessary escalation. Escalate liberally and accept the false positives.
If this resonated
Module 02 is the long version of this post.
Eval set, refusal pattern, kill switch. Build it in a weekend.
