The first time someone asks for “an AI agent for each location,” I can feel the future outage forming in the air. Twelve prompts. Twelve knowledge bases. Twelve slightly different crisis paths. Twelve ways to say the same wrong thing about insurance.
Multi-location ops does not need a swarm of personalities living in separate apartments. It needs one brain that knows the company, plus overlays that know the building.
One brain, or you will drift
The shared brain holds the stuff that must be true everywhere:
- who you are and what you do
- what the agent is allowed to say
- what it must never say (clinical advice, legal promises, “you’re definitely covered”)
- crisis and escalation rules
- how a good intake looks: name, callback number, reason, location, urgency
- where the conversation gets written down
If those rules live in twelve places, they will diverge by Thursday. Someone will “just tweak the prompt” at Location B. Location C will still be on last month’s hours. Location D will start giving medical opinions because a well-meaning manager pasted a FAQ into the wrong box.
Treat the shared brain like production code. Changes go through a review. Version it. Test it against a handful of scary transcripts, not just the happy path.
Personalities are overlays, not clones
A facility still needs a voice. Hours differ. Programs differ. Payer mix differs. The person answering for a detox unit should not sound like a suburban outpatient lobby, and neither should sound like a call center in a parallel universe.
The overlay is a short file, not a second agent:
- display name and local greeting
- hours, after-hours rules, time zone
- programs offered and who they are for
- what they do not offer (this saves more pain than any slogan)
- parking, first-visit, and “what to bring” only if it is stable
- insurance notes that are allowed to be said out loud
- the human team that owns the callback
That’s it. If the overlay is longer than a one-page SOP, you are stuffing policy into personality. Policy belongs in the brain.
Prefer a simple routing rule: identify location early, load the overlay, keep the brain. If the caller does not know the location, the agent should help them find it with a few questions, not guess and hope.
The data problem hiding under the prompt
Most multi-location failures are not model failures. They are spreadsheet failures wearing a headset.
Hours in three systems. Program names that don’t match the website. A “we take that insurance” line that is true at one campus and a lawsuit at another. An agent cannot be more honest than the table you gave it.
So before Focus on tone, you want a location table that a grown-up owns. One row per facility. Fields the agent is allowed to read. A last-updated date that is not theater. If a field is stale, the agent should say it will have a human confirm, not invent a confident answer.
This is boring work. It is also the work. Fancy voice on top of rotten location data is how you scale embarrassment.
How to keep them from contradicting each other
Voice, chat, and the website should drink from the same location table. If the site says a program exists and the agent says it doesn’t, you trained the public to distrust both.
Also keep a contradiction test. Pick ten questions that used to cause fights:
- Do you take this insurance at this campus?
- What’s the after-hours number?
- Can I just show up?
- Do you have beds tonight?
- What’s the first step if you’m calling for my kid?
Run them against every location overlay on a schedule. When two answers disagree, you don’t “prompt harder.” You fix the source.
For anything that changes daily, like bed availability, I either wire a live source or I forbid the agent from answering. “Let me have admissions call you back” is a better sentence than a guess that wrecks someone’s night.
Humans still own the weird cases
Multi-location is where edge cases go to breed. The caller is in a different state than the facility. The program they want is at campus 3 and they called campus 1. They are in crisis and also asking about payment plans.
The agent should be good at triage and terrible at pride. Route it. Capture it. Don’t perform expertise you don’t have.
Give the morning team a queue that is grouped by location, not a pile of transcripts. A director should see their house first. If your “one brain” dumps every campus into one undifferentiated inbox, you didn’t build a system. You built a shared headache.
Bottom line
One agent brain is how you stay consistent. Overlays are how you stay local. A clean location table is how you stay honest.
Clone twelve agents and you will spend the next year debugging your own hydra. Build one brain with many facility personalities and the caller gets the right place, the team gets notes they can use, and you get to change a rule once instead of twelve times at 11pm.
