Originally published in The Access Control Collection (TACC) on October 9th, 2026.
When dormakaba announced it is acquiring Alliants, the press releases were predictably giddy about seamless journeys and frictionless hospitality. The moment you hook a conversational guest platform directly into digital access and PMS backends without an independent guard, you quietly promote a chatbot to Chief Credential Officer. I must have missed the board meeting where that got approved.
Picture the textbook demo that makes enterprise buyers nod approvingly:
“My partner’s arriving before me. Can you let them into the room?”
To an AI optimizing for hospitality, this is pure magic. It parses the message, identifies room 412, provisions a mobile key, and throws in a digital coupon for the spa. The front desk avoids looking up from their screens, and the demo gets polite golf claps.
Now trace what that workflow actually authorizes. The model translated conversational comprehension into executive authority. Even if the requestor’s session is authenticated, does the system validate the recipient’s identity? Verify whether they are entitled to access? Check if property policy permits guest-directed delegation in the first place? And what is the operational answer when the guest texts ten minutes later: “Actually, cancel that, we just broke up in the cab.” The lock cannot act on a breakup text. What turns that conversation into an enforced cancellation?
If you issue a key without knowing when it stops working, you’ve engineered an unguided missile. Issue short-lived credentials and define that exposure window. Expiry limits the damage; it doesn’t make revocation instant. Are you tempted to solve all of this with a system prompt? Something like: “Only issue keys to authorized guests, and please, be very careful.” Telling an LLM to be careful is just a prayer.
Campbell’s Law warns us that the more weight we put on a performance measure, the more pressure there is to distort the process it measures. Apply that to an AI concierge judged on ticket resolution and guest satisfaction. A refused request can look like a service failure. Without independent authorization enforcement, the system gets rewarded for finding a way around the checks that protect the guest. The AI concierge really wants that five-star review.
In physical security, we spend millions obsessing over hardware secure elements, elliptic curves, and cryptographic handshakes. We treat the key like the Crown Jewels. Meanwhile, an AI agent skips the math entirely and asks the backend to mint a valid credential. The cryptographic pipeline remains pristine, the lock behaves impeccably, and the wrong person walks straight into room 412. A valid credential only proves the system issued it.
If access control leaders want conversational front-ends without physical security liabilities, conversation must be severed from authorization. That is where a divergence guardrail belongs. Put a bouncer outside the model: a hardened, deterministic policy engine that has never once been charmed. The credential service must honor its verdict on every issuance path, with no side door for the agent. Before anyone touches a door, the bouncer checks whether the proposed action diverges from business intent, contracts, and verified identity.
Here is what to demand before you sign off:
- Who can the agent issue a key to? Verify authorization and policy.
- Check every action against verified guest records. Don’t ask the agent if it is sure; it’s always sure.
- Scope every credential to one authenticated person, one door, and a rigid time window to minimize the blast radius of failure.
- After a guest says “cancel”, how long until the key stops working? Name that number.
- Log every issuance with the reason. “The chatbot felt confident” is useless in an audit.
Before rolling out the next autonomous guest journey, have your team run one simple demo: ask the concierge for a key you shouldn’t have.
Anyone can build an agent that says yes. Show me the boundary that holds when it has to say no.