Deloitte Says 84% See a Return. In an Independent Shop It Shows Up on the Declined-Work List, Not the Front Counter

Answer this by Friday: how much did you sell last month off work your own techs already found?

Not new customers. Not the marketing spend. Work that a technician in your building already found, photographed, measured and wrote up — and that the customer said no to, or said "not today" to, at some point in the last ninety days.

Most owners cannot answer that question, and it is the most embarrassing gap in the business, because that work is already sold in every way that matters except one. The vehicle has been on your lift. The measurement is on file. The photo exists. The customer has already paid you once and did not complain. There is no acquisition cost left to spend.

Deloitte's State of AI in the Enterprise 2026 found that 84% of organizations investing in AI report a positive return, and the pattern behind that number is dull and repeatable: take one messy process, put a human in the middle of it, and prove either time saved or errors reduced before you widen anything. In a repair shop, the one messy process to start with is the declined-work list.

Why the declined-work list is the messiest thing you own

It lives in three places at once. There is the "recommended" or "declined" status inside Tekmetric or Shop-Ware or Mitchell 1. There are the red and yellow items on the digital inspection, which nobody looks at again once the ticket closes. And there is a sticky note on the advisor's monitor that says "Call Bill about the rack — after tax refund."

Nobody owns it, and that is not a discipline failure. Your advisor's day is forty phone calls, nine write-ups, three parts problems, one warranty administrator on hold, and a customer standing at the counter at 5:05 asking why the alignment cost what it cost. The follow-up list is the first thing that falls off, every single day, in every shop in America.

So here is the definition to work from: proving AI paid for itself, in a repair shop, means picking one process with a number attached — declined work recovered within ninety days — writing down what that number is today, and comparing it sixty days later with the same advisor, the same door rate and the same season.

Six numbers to write down before you switch anything on

This is the part everyone skips, and it is the reason most owners end up arguing about whether the thing worked instead of knowing. Pull these for the last ninety days, from your shop management system if it exports and from the repair order file if it does not.

• Repair orders written in the period.
• What share of them carried at least one declined or recommended item.
• Average dollar value of the declined items per repair order.
• How many of those declined items were sold within ninety days — your recovery rate.
• Hours your advisor currently spends on follow-up in a week. Ask him; do not guess.
• Your gross profit percentage on that class of work, which is usually higher than your overall number because it is brakes, suspension and fluid service rather than diagnosis.

That is an afternoon of work, once. Without it, you will be comparing May to July and calling air conditioning season a success.

flowchart TD
  A["Export 90 days of declined items from the shop system"] --> B["Write down the baseline recovery rate and dollars"]
  B --> C["Drafts prepared for the advisor each Monday morning"]
  C --> D["Advisor edits, deletes and sends; nothing goes out unread"]
  D --> E["Every callback logged against the original repair order"]
  E --> F["Month end: compare recovery rate to the baseline"]
  F --> G{"Beat the baseline by more than the cost?"}
  G -->|No| C
  G -->|Yes| H["Widen to the second process, keep measuring"]

What "human review" looks like when it is your service advisor

Monday, 7:15 a.m., before the first drop-off. On the advisor's screen is a list of twenty-two customers from the declined pile, each with a drafted message that references the actual item, the actual photo and the actual mileage: "You were at 92,400 in April and the rear pads measured 3mm. You are probably around 97,000 now, which puts them close."

He deletes four. One is a trade-in that already went to the auction. One is a customer who argues about every dollar and he would rather not spend the morning on it. He rewrites three because the tone is off for people he has known for eleven years. He sends fifteen. The whole thing takes eighteen minutes, which is roughly what he used to spend making four phone calls that went to voicemail.

That is the entire discipline, and it is the difference between the shops that get a return and the shops that get a mess. Nothing goes to a customer unread by a human who knows the customer. The machine does the remembering and the drafting; your advisor does the judging and the sending.

The number that settles the argument

Assumptions, illustrative: 280 repair orders a month, 62% carry at least one declined item, average declined value $415 per ticket, baseline ninety-day recovery of 11%. Gross margin on that work 58%. Cost: $120 a month in AI spend plus about three advisor hours a week, which you are converting from time he was already losing to voicemail.

Now the part that makes this honest rather than a sales pitch. Do not run it on the whole list. For sixty days, work only the odd-numbered repair orders and leave the even ones alone. At the end you have two groups from the same season, the same advisor and the same weather, and the difference between their recovery rates is the answer. If there is no difference, you have learned something worth more than $120.

Three ways this measurement will lie to you

Seasonality is the first liar. A/C work in May and no-starts after the first hard freeze in November will swamp any process change you make. Comparing April to July proves nothing about your follow-up and everything about the weather. The odd-and-even split protects you from this; a before-and-after comparison does not.

Attribution is the second. If you started replying to Google reviews the same month, ran a postcard on state inspections, and hired a new advisor, you will never untangle which one moved the number. Change one thing at a time for sixty days. It is boring, and it is the only version that produces a number you can defend.

Cherry-picking is the third, and it is the sneaky one. Given a list, an advisor calls the easy customers first — the ones he likes, the ones with money, the ones who always say yes. Your recovery rate goes up because the list got easier, not because the process got better. Work the list in the order it comes.

Where a human has to stay in this loop

Never let a drafted message invent a price. The number goes out only if it came off the estimate already on the repair order, and even then say "we quoted $415 in April, let me confirm today's parts price" — because a control arm that was $118 in April may not be $118 now, and quoting an old price into a phone call is how you end up eating $60 to keep the peace.

Never sell safety work sight-unseen. Pads at 3mm in April may be metal on metal in July. The message asks them to come in for a re-check; the tech, not the follow-up, decides what gets sold.

And mind the texting rules. Message consent is not a technicality — the federal statute behind unwanted texts carries damages of $500 per message and up to $1,500 for willful violations, per message. Only text numbers where the customer gave you permission on the intake form, honor a stop the first time somebody sends it, and keep the do-not-contact list where the drafts can see it.

Frequently asked questions

How long before I actually know whether it worked?

Sixty days with a split list gives you a defensible answer for one process. Ninety days is better because your recovery window is ninety days. Anything you decide in three weeks is noise, and anyone who tells you otherwise is selling.

My shop management system will not export the declined items cleanly. Now what?

Then pull one month by hand — one afternoon with the closed repair orders and a spreadsheet with four columns: date, customer, declined item, dollar amount. That single month is a good enough baseline to start, and it will also tell you whether your advisor is even writing declined items into the system properly, which is a separate problem worth finding.

Does this replace my service advisor?

No, and if it did, your recovery rate would fall. The customer buys the rack because Danny has known them since the Sienna, remembers the daughter's commute, and tells them straight when something can wait. What comes off his plate is the remembering, the typing, and the four voicemails.

What number do I put in front of my accountant?

Recovered declined-work gross profit per month, next to total AI spend per month, on the same page. If the first is not several times the second by day sixty, do not widen the scope — fix the process or stop.

The Monday step: export ninety days of declined items and write the recovery rate on a sticky note. That one number is the whole argument. And when those follow-ups start working, the calls come back into a phone that is already busy — which is the next bottleneck. CallSphere builds AI voice and chat agents that answer shop lines and web chat, book the appointment and capture the lead when both advisors are already with customers, so the work you just recovered does not die in a voicemail box.