AI in Restoration

Two years ago, AI in restoration was a chatbot on a website. Today it’s writing your supplement candidates, calling on your aged AR, and grading your scope against IICRC S500. Most owners have no idea what’s already deployable, and most vendors are happy to keep it that way.

Here’s the operator-grade read.

What AI is actually doing in restoration shops right now

Five real use cases have emerged from pilot to production in the last 18 months. Not slides. Not demos. Real.

1. AR collection agents

The clearest win. An AI agent that calls on 30, 60, 90, and 120-day aged invoices, speaks naturally, identifies itself, navigates to the right accounts payable contact, takes a payment commitment or escalates to a human. Cost per call is roughly 10 to 20 percent of a human collector. Throughput is 24/7. The AI doesn’t get demoralized after the eighth voicemail.

The catch: it has to be tuned to your AR data, your carrier list, and your escalation rules. Off-the-shelf voice agents that don’t know what a TPA is, or what AOB means, will burn your reputation in a week. Demand a vendor who has either restoration-specific training or operator review of every script before launch.

2. Carrier comms triage

The other clean win. Inbound carrier emails get classified, routed, and pre-drafted by an AI that has read the file. Adjuster wants a moisture log. AI pulls the moisture log from the file and drafts the response. Adjuster wants a supplement justification. AI drafts the language with the IICRC reference and the photo IDs. Human reviews and sends.

The savings here aren’t headcount. They’re cycle time. The shop that responds to carrier requests in two hours gets paid faster than the shop that responds in two days, and the TPA scorecard reflects it.

3. Scope review against standards

Every shop has the same problem: scopes get written by humans on Friday night and nobody else reads them before they go to the carrier. AI scope review reads the Xactimate file, cross-references the photos, and flags missing or mis-categorized line items. Pad-and-replace on a Cat 2 with no antimicrobial line. Demo without haul. Drying with no psychrometric documentation attached.

This is the use case where Restoration 360’s training matters most. We’ve trained on 200 projects and 20,537 field photos to teach the model what a real restoration scope looks like, not what a generic LLM thinks it should look like. The difference is the difference between catching real omissions and generating noise.

4. AI answering and intake

Voice AI on the front end of the business. After-hours FNOL intake, qualification, dispatch routing. Done well, it captures the loss, gets the address, gets the cause, gets the carrier, and pages the on-call mit lead with a structured loss summary by the time they pick up the phone.

Done badly, it loses the loss. The bar here is high because the cost of a missed water loss at 11 PM on a Friday is your competitor’s truck in the driveway by midnight. Most AI answering products are not yet at that bar for restoration. Some are. Test the failure modes before you trust them.

5. BI dashboards that actually work

Less sexy, more important. Most restoration shops are running on QuickBooks plus their job management platform plus a spreadsheet. The data is there. Nobody can see it. AI-assisted BI doesn’t mean a magical insight engine. It means a dashboard that pulls from all three sources, reconciles them, and shows the owner the three numbers that matter on Monday: recon margin, AR aging, and supplement conversion rate.

For reference, R360’s training cohort runs at a 52% recon margin. If you don’t know what yours is this week, you don’t have a dashboard problem. You have a visibility problem, and AI-assisted BI is the cheapest fix.

What’s still hype

Plenty.

• “AI estimating” that writes a full Xactimate from photos with no operator in the loop. Not real at production quality. The corpus is too thin, the variance in losses is too high, and the carriers will reject anything that looks templated. Anyone selling this is selling a demo, not a product.
• “AI moisture mapping.” The hardware is real. The interpretation layer that decides Cat and Class is not yet good enough to replace an IICRC tech. It’s a useful second opinion. It is not a decision-maker.
• “AI customer experience” chatbots on the website. The 2022 product. Mostly noise. The good ones are voice agents on the phone, not text bots in a chat window.
• “Predictive supplement detection” at scope time. Promising, real in narrow cases (water migration patterns, contents categories), oversold in broad ones. Ask for the precision and recall on real files, not a marketing number.

The honest version of AI in restoration in 2026: it removes the bottom 60 percent of low-judgment work and frees the operator to do the top 40 percent better. It does not replace the operator. Anyone who tells you it does has not run a restoration shop.

What to demand from any AI vendor

Owners are getting pitched constantly and most of the pitches are indistinguishable. Here’s the screen.

1. Show me a real file. Not a demo file. A real, anonymized loss from a real shop, run end to end through your tool. If they can’t, they don’t have customers.
2. What’s your training data? “We use GPT” or “Claude” is not an answer. The question is what restoration-specific data sits between the base model and your output. If the answer is none, you’re paying a premium for a wrapper.
3. Where’s your operator in the loop? Every working AI tool in restoration has a human review point. Where’s yours? If the vendor says “fully automated,” they’re lying or they’re about to embarrass you.
4. What does your tool not do? A vendor who can’t list three things their tool fails at is a vendor who hasn’t shipped to real customers yet.
5. How do you handle a carrier complaint? When the desk adjuster calls and says “this scope looks AI-generated and I’m pushing back on every line,” what does the vendor do? If they have no answer, you have your answer.
6. What’s the integration cost? Setup, data migration, ongoing maintenance, and the human time on your side to keep it tuned. The license fee is rarely the real cost.

The stack that’s actually emerging

The shops that have moved aggressively on AI in the last 18 months are not running one tool. They’re running a stack: voice AI for intake and AR, an LLM-backed scope reviewer, a carrier comms triage layer, and a BI dashboard reconciling everything. The total monthly spend is real, often $3,000 to $8,000 a month for a $5M shop. The return is in cycle time, recon margin, and AR aging, not in headcount reduction.

Owners who treat AI as a headcount play are usually disappointed. Owners who treat it as a visibility and cycle time play are quietly compounding.

What to do Monday

Pick the function in your shop with the worst cycle time. For most owners that’s AR follow-up or supplement turnaround. Don’t shop for an AI vendor yet. Spend one hour timing the actual cycle: how long from invoice issued to first follow-up, how long from supplement identified to written. Whatever number you get is your baseline. Any AI tool you evaluate has to beat it by 3x or it’s not worth the integration cost. That’s your screen, and it’s the same screen every serious operator should be running.