Cold Fries and an Allergen Complaint Hit the Same Inbox. Route Them to Two Different Models.

It is 2:10 on a Tuesday afternoon in March. The lunch rush at the Broad Street store broke fifteen minutes ago, the general manager has finally sat down in the back office with a fountain drink, and the guest-recovery inbox has 41 unread cases sitting across your six stores.

Thirty-four of them are the same four sentences. Fries were cold. Missing the side of ranch. The app charged for a large and the crew handed over a medium. Order 4471 never got the extra pickles. Each one costs the GM eight to twelve minutes: open the case in Punchh or Paytronix, pull the ticket in Toast, decide whether to comp, type a reply that does not sound like a form letter, push the loyalty credit, code the reason, close it.

Two of them are not that at all. One says a nine-year-old with a peanut allergy ate a cookie from your counter and ended up at urgent care. One says a crew member said something in the drive-thru lane that you are going to have to pull camera footage to understand.

Those two are sorted by arrival time, which means they are sitting nineteenth and twenty-eighth, behind the ranch dressing.

The workaround every multi-unit operator pretends is fine

Almost nobody triages that inbox properly, because triage is itself a job. What actually happens is that the GM works top to bottom until the 4:30 truck arrives, then stops. Whatever is left rolls to tomorrow. The franchisor's brand standards say guest contacts get a response inside 24 hours; your real average is closer to 31, and the two cases that mattered were the slowest, because they were the longest to read.

The 2024 version of automating this was a canned-reply macro: keyword matches on "cold" and "missing" that fired a template and a $5 offer. Operators switched those off within a period, because the macro cheerfully sent a coupon to the woman describing her child's emergency room visit. That is a screenshot on a local news Facebook page, and it costs more than every cold fry you have ever comped.

So the trade settled on all-or-nothing: a human reads every case, and the hard ones wait in line with the easy ones.

What changed in 2026: cheap model for the ordinary case, strong model for the hard one

Through the first half of 2026, model routing stopped being a clever trick and became the default way serious rollouts get built. Cisco is putting a personal AI agent in front of roughly 90,000 employees with routing designed in for exactly this reason: run the ordinary work on a fast, inexpensive model and spend the expensive one only where it earns its keep. Frontier AI pricing is down roughly tenfold from 2025, which is what makes the cheap lane cheap enough to run over every single case that arrives.

Model routing means the fast, cheap model handles the ordinary guest complaint and only the genuinely hard case is escalated to the expensive one — the same way a shift lead handles a remake and the GM handles the call from the health department.

The important part for a restaurant operator is not the saving. It is that routing forces somebody to write down, in advance, what counts as hard. That one page is worth more than the software.

flowchart TD
  A["Guest case lands: app, phone, DoorDash or comment card"] --> B{"Names an allergy, illness, injury, minor or attorney?"}
  B -->|No, and comp under $25| C["Fast model drafts reply and loyalty credit"]
  B -->|Unclear, or third case from same guest| D["Strong model reads full order and visit history"]
  B -->|Yes| E["Straight to the Director of Operations"]
  C --> F["GM approves the batch at shift change"]
  D --> F
  E --> G["Incident form opened, insurer and franchisor notified"]
  F --> H["Case closed, reason coded to store and daypart"]
  G --> H

Where the line actually sits in a fast-casual complaint queue

In this trade the routine lane is broad and boring, and that is the point. A cold-item complaint on a delivery order that sat on the shelf 40 minutes. A missing side. A modifier the kitchen dropped. A duplicate charge from a card tapped twice on a Toast Go handheld. A guest who cannot find the loyalty points from last Thursday. A drive-thru receipt showing the new limited-time sandwich rung at the pre-promo price. Every one of those can be settled from the ticket, the refund policy and the guest's own order history, and none of them will ever turn into anything else.

The hard lane is narrow, and in a restaurant it is not defined by how angry the guest sounds. It is defined by exposure. Any mention of an allergen, a reaction or an epinephrine pen. Any mention of illness, vomiting, or more than one person from the same party being sick — that one is a possible foodborne illness cluster, and your county health department may be about to call you before the guest does. Anything foreign found in the food. Anything involving a minor. Anything about crew conduct, a slur, a service animal or a wheelchair. Anything naming a lawyer, an insurance company or a news station. And any third contact from the same phone number inside 30 days, which is either a real operational failure at one store or a refund abuser working your app.

Everything in that second list goes to the strong model, which reads the whole thread, the full visit history and your written incident policy — and drafts nothing to the guest. It writes a summary for a person.

Who draws that line, and it is not the software vendor

The line belongs to your Director of Operations, and two other parties should sign it before it goes live. Your franchisor's brand standards team, because most franchise agreements reserve control of guest communications and require notification of illness or injury claims within a stated number of hours. And your general liability carrier, because the first thing an adjuster asks after a claim is who said what to the guest, and when.

Write it as a one-page list, not a policy essay: trigger words, dollar thresholds, and the name of the person who gets paged. Then review it every quarter with your Area Coaches, because the line moves. The summer you add a peanut-sauce bowl, the allergen trigger gets stricter. The period after you launch a $6 value menu, the refund-abuse trigger gets tighter.

The arithmetic on 1,100 cases a period

An illustration, not a case study. Assume six fast-casual stores, 1,100 guest-recovery cases per four-week period — about 46 a day across the group — and 82% of them routine. Assume a loaded GM cost of $28 an hour, ten minutes per routine case today, and thirty seconds of approval per case once routing is on.

Call it $3,900 a period, or roughly $50,000 across thirteen periods. Be honest about the shape of that saving: GM hours do not turn into cash unless you put them somewhere — on the line at peak, on hiring, on the brand-standards audit score. If the hours simply evaporate, the real win is the other one. The allergen case gets read at 2:11, not at 9:40 the next morning.

Where this still needs a human

Routing decides nothing about a sick guest. It decides who reads the message first. The health-department call, the incident report, the decision to pull a product code from all six stores, the conversation with the parent — human, all of it, exactly as in 2019.

It also will not fix a store. If the Route 9 location throws four times the cold-item complaints of the other five, the answer is a holding-time audit and a look at the drive-thru timer report, not a faster refund. Routing makes that pattern visible sooner, which is uncomfortable, because it is usually one manager and one shift.

And do not let anything write to a guest about a refund above the number your controller can live with. Most operators land between $20 and $30. Above that, a person presses send.

Monday's first step: run it in shadow for one period. The models draft, nobody sends, and your Director of Ops compares the drafts against what the GMs actually sent. You are not testing the software. You are testing whether your one-page hard-case list catches the cases you would have wanted caught.

Frequently asked questions

Will it comp things I would not have comped?

Only up to the limit you set, and only under the rules you write. The realistic surprise runs the other way: it is consistent, and your six GMs were not. Expect the comp line to move in the first period simply because six people stopped applying six different standards.

Does this work for complaints that come through DoorDash and Uber Eats?

Partly. Those platforms own the guest conversation and charge the adjustment back to your store, so what you are triaging is the chargeback rather than the guest. The sorting logic is the same; the recovery is a dispute filed inside their window, not a reply.

Our franchisor runs the guest feedback system. Can I still do this?

Ask first. Many brands require guest replies to go out through their platform in approved wording. Some will let a franchisee triage internally and answer inside the brand tool; some will not. Get the answer in writing from your Franchise Business Consultant before you spend a dollar.

What does it cost to run each period?

Less than you expect on the routine lane and more than you expect on the hard one, which is the entire reason routing exists. Watch it with a spend cap and alerts — the enterprise governance update that shipped on 2 July 2026 lets you set limits by team and get warned at 75% and 90% of budget instead of at the invoice.

One related note: a good share of these cases never reach an inbox at all, because the guest calls the store during peak and the phone simply rings out. CallSphere builds AI voice and chat agents that answer the restaurant line and web chat around the clock, take the details and pass the case on — which at minimum means the allergen complaint gets logged at 2:10 rather than lost because both handsets were busy.