Intent classification · Per-vertical NLU · Franchise + multi-unit service

The complaint names billing + access + intent to dispute in one message. Route it before it ends up in a Google review.

You run customer messages across 50-1,500 franchise locations. A single 2am SMS can carry account-trouble + billing-dispute + escalation-threat at once. Single-label classifiers force one label, the night-shift agent reads it as a routine how-do-I, and Tuesday morning the 1-star review with the franchise name is the first surface that shows the gap. Multi-label intent classification + per-location escalation routing closes that window.

Audit your urgency-routing surface (Tier 1)Explore Fractional CMO operation

Published May 30, 2026

Why per-location reputation costs more than the routing fix

A franchise customer SMS at 2:14am reads: “Tried 3 times to login. Charged me anyway. I want a refund or I’m disputing.” That message carries four intents: account-access-blocked, billing-event-disputed, refund-requested, payment-dispute-threatened. A single-label classifier picks one — probably refund-requested. The night-shift agent receives a routine refund ticket. The chargeback risk is undetected. The escalation risk is undetected. The account-access root cause is undetected.

Tuesday morning the customer leaves a 1-star Google review naming the specific franchise location. Three other prospects within the territory see the review during their decision cycle. The per-location GBP rating drifts down by 0.1-0.2 stars depending on review volume — enough to affect local-pack ranking on competitive queries. Your franchisee P&L feels it across 4-12 weeks. The whole loop ran because four labels collapsed into one at the routing moment.

The gap is not the chatbot vendor. The gap is that one message can carry multiple intents at multiple urgency levels, and the routing layer needs to see them all.

We’ve built the classification layer for multi-unit operators. Here’s what we know.

You probably already use a foundation-model API or a managed NLU service. Each is good at the classification primitive. The gap at multi-location scale is the per-vertical taxonomy that names which intents matter for your business, the urgency aggregation that converts multiple labels into a single per-location actionable score, the per-location threshold tuning that respects per-store staffing and queue depth, and the routing wiring that converts the score into a real-time escalation event.

We have built this layer for franchise and multi-unit service operators. We know which label families recur across verticals (account-access, billing, product, NPS, escalation-threat) and which are vertical-specific (gym freeze-policy, salon booking-conflict, food per-store order-error). We bring the taxonomy and the urgency weights as a starting point and tune them against your ticket history.

How we get from one-label routing to multi-label per-location escalation

Step 1 — Tier 1 AI Readiness Assessment (2-3 weeks). We audit your current classification coverage across chat + email + SMS + voice transcripts + form submissions. We sample your last 30-90 days of inbound messages and run a multi-label diff against the labels your current routing uses. Output: the per-vertical label taxonomy spec, the urgency-weight starter, and a per-location threshold recommendation per location-staffing-pattern.

Step 2 — Tier 2 AI Swarm Setup Sprint (4-8 weeks). We build the classification layer end-to-end: foundation- model integration, per-vertical taxonomy, per-channel context overlay, urgency aggregation, per-location routing wire, per-location auto-escalation policy, the typed at-risk event publisher, and the false-negative review workflow. Your engineering team receives the running system, all source code, all credentials.

Step 3 — Tier 3 Fractional CMO with AI Swarm ( 6-month minimum, 1-2 days/wk). We operate the layer in production. Tune per-vertical weights as new intents emerge. Re-calibrate per-location thresholds against staffing changes. Run the weekly false-negative review with your support leadership. Roll up a monthly per-location urgency-routing report.

What changes for you

You stop discovering the Tuesday 1-star review because the Saturday 2am SMS that produced it was routed to a senior agent at 2:15am, with the location manager looped in.

You stop reading every escalation email yourself because the urgency score is tuned and the routing rules are owned. The messages that reach your inbox are the genuine executive escalations.

You can answer the question your franchise owners ask at every quarterly review: which locations have the highest at-risk inbound volume, and which signals are driving it. The classifier’s typed at-risk events roll up per location per week.

You can onboard a new vertical (a new franchise concept, a new product line) without re-training a model from scratch. The taxonomy spec describes how new intents are added.

Frequently asked

What does multi-label intent classification mean and how is it different from sentiment analysis?

Sentiment analysis returns a single polarity score per message (positive, neutral, negative). Useful, but flat. Multi-label intent classification returns the full set of intents a single message carries simultaneously — account-locked, billing-disputed, refund-requested, threat-to-go-public, NPS-detractor — each with its own confidence score. A real customer message often carries three or four labels at once. Sentiment tells you the customer is unhappy. Multi-label tells you which four things they want addressed and which one is most likely to escalate.

Why is per-vertical classification different from the generic models my chatbot vendor ships?

Generic models trained on broad customer-support corpora handle the common-case intents well. They miss vertical-specific signals. A franchise gym customer asking about freeze-policy is a different intent than a SaaS customer asking about plan-pause. A franchise salon customer mentioning "I drove past your location" is a per-location-staffing complaint, not a generic complaint. The per-vertical label taxonomy adds the vertical-specific intents on top of the generic-support intents. The per-vertical urgency weights tune what should fire which escalation in your operational reality. Both layers compose on top of whichever foundation model you use.

How does per-location escalation routing actually work for multi-unit operators?

Each inbound message gets classified, scored for urgency, and assigned to the right person at the right location. Above a per-location auto-escalation threshold, the message routes to a senior agent + alerts the location manager + emits a typed event into the at-risk-customer pipeline. Above a per-location critical threshold, it also pages on-call leadership and creates a per-customer recovery workspace. Thresholds are per-location because per-location queue depth + staffing patterns differ. A 3am gym message in a 24-hour facility routes differently than a 3am message at a 9am-9pm salon.

What does Completions commit to on Tier 3 if we run this layer in production for us?

Tier 3 process commitments include: multi-label classification at sub-second p95 latency for chat + form + SMS channels; per-vertical label taxonomy maintained quarterly as new intents emerge in your ticket history; per-location urgency-threshold tuning monthly with your operations leadership; weekly false-negative review of messages that crossed the at-risk pipeline but did not trigger escalation. We commit to the operating discipline. Per-vertical recall + precision are tuned per your ticket corpus and recorded as engagement KPIs.

Who owns the labels, the weights, and the per-location thresholds?

Your team owns the per-vertical label taxonomy, the per-location threshold config, the routing rules, the agent rosters, and the credentials. Completions owns the orchestration knowledge: the labeling guidelines, the urgency-weight tuning history, the per-location threshold-drift triage playbook. At engagement end we transition operational ownership back to your team over 30-60 days with documented handover.

How does the urgency stream connect to the rest of the marketing + retention stack?

The classification layer publishes an urgency score per message. The per-location routing layer consumes it for escalation. The pre-emptive-intervention-triggers layer subscribes to the at-risk pipeline for retention. The recovery-rate-dashboard joins urgency-scored messages to recovery outcomes. The change-event-emission stream carries message-state changes that drive downstream marketing-automation re-evaluation. Five layers, one urgency contract. The classification layer owns the contract; each consumer owns its own subscription.

Start with the audit

Tier 1 AI Readiness Assessment (2-3 weeks): we audit your current routing surface, sample your last 30-90 days of inbound messages, and produce the per-vertical label taxonomy + per-location threshold spec. If you decide to build, Tier 2 ships the layer. If you decide to operate it with us, Tier 3 runs it in production. You choose the next step at each gate.