How to Choose an AI Concierge: 12 Questions to Ask

Every vendor demo looks convincing. The bot answers quickly, sounds polite and never misses a beat. The problem is that a polished demo tells you almost nothing about how the assistant will behave on a Tuesday afternoon when a patient asks about insurance, a guest wants a late checkout and a prospect asks for pricing. This guide explains how to choose an AI concierge by turning the sales conversation into a set of checkable questions, red flags and a short pilot.

The stakes are higher than most vendors admit. Harvard Business Review's study of 2,241 companies found that the average response time to a web lead was 42 hours, only 37% replied within an hour and 23% never replied at all; a reply within the hour made a lead almost seven times more likely to be qualified. Drift's vendor survey of 433 B2B SaaS companies found that only 7% answered a demo form within five minutes and 55% took more than five working days. An AI concierge exists to close that gap — not to add another layer of delay.

How to choose an AI concierge when every demo looks the same

The core problem is that most vendors demo the same happy path. They show a friendly greeting, a couple of canned answers and a handoff to a human. What they rarely show is the bot handling an incomplete question, a language switch, a refund request or a lead that arrives at 11pm on a Sunday.

A useful mental model is to treat the vendor as a system, not a script. You are buying four things: a knowledge layer that stays current, an integration layer that touches your booking or CRM system, an escalation layer that knows when to stop, and a reporting layer that proves what happened. If a vendor cannot explain all four in plain language, the demo is doing the work that the product should be doing.

That is why the questions below are grouped by layer. Ask them in the same order, write down the answers and compare vendors on evidence rather than adjectives. For a deeper look at where the boundaries should sit, see our notes on guardrails for clinics and hotels.

The 12 questions

Knowledge and answers

1. Where do the answers come from? The bot should answer from your documents, pages and policies — not from a generic model's memory. Ask to see the source of an answer live.
2. How does the knowledge base stay current? Find out whether updates are automatic, scheduled or manual, and who owns that task after launch.
3. What happens when the bot does not know? The correct answer is an honest fallback and a handoff, not a confident guess. Ask the vendor to demonstrate a question outside the knowledge base.
4. Which languages are supported, and how well? Test a real question in each language you need, including spelling variants and mixed-language messages.

Integrations and handoff

1. Which systems does it connect to? For clinics this means the practice management or EHR system; for hotels, the PMS and booking engine; for SaaS, the CRM and calendar. Ask for named integrations, not "API available".
2. How does a lead reach a human? Ask whether the bot creates a CRM record, sends a notification, books a slot or all three — and what the rep sees.
3. What is the escalation logic? Look for triggers such as frustration, repeated questions, medical or safety topics, high-value requests and explicit requests for a person.
4. What happens outside business hours? The assistant should capture the lead and set expectations, not promise a callback it cannot keep.

Operations and commercials

1. What is the SLA, and what is excluded? Ask about uptime, response times for support tickets and the escalation path when something breaks.
2. Who owns the data and the conversation history? You want export rights, deletion on request and a clear statement on where data is stored. This matters for GDPR in the UK and for HIPAA-adjacent workflows in the US.
3. How is pricing structured? Per conversation, per resolved query, per seat or a platform fee — each creates different incentives. Model a busy month, not an average one. Our breakdown of what an AI concierge costs shows where the lines usually sit.
4. What does success look like in numbers? Ask for the metrics the vendor will report: containment rate, escalation rate, leads captured, booked appointments and response time.

Red flags to watch for

A vendor who cannot show a live knowledge-base answer is a demo wrapper. If every answer comes from a scripted flow, the assistant will break the first time a customer phrases something unexpectedly. Ask to type your own question into the live bot, not the vendor's prepared one.

Other warning signs: vague answers on data ownership, "unlimited" pricing with no fair-use definition, no named integrations for your sector, and a refusal to run a short paid pilot. Be equally wary of vendors who promise full automation with no human escalation — in clinics and hotels, that is a compliance and service risk, not a feature.

Finally, watch how the vendor talks about response time. If the pitch is built on speed alone, ask what the bot does with the lead after it replies. A fast answer that goes nowhere is just a faster dead end.

A one-week pilot checklist

A one-week pilot is enough to separate a working assistant from a polished demo, provided you define the test before you start. Pick one channel, one knowledge base and one handoff destination, then run real traffic through it.

• Day 1: Load 20–30 real questions from your inbox or chat logs, including awkward ones, and check the sources of the answers.
• Day 2: Test the lead handoff end to end: form or chat to CRM to notification to human reply.
• Day 3: Test escalation triggers, including out-of-hours messages and a request for a person.
• Day 4: Test languages and edge cases such as refunds, insurance, cancellations and complaints.
• Day 5: Review the numbers: containment, escalation, leads captured and time to first human reply.

Run the pilot on live traffic if you can, and compare it with your current baseline. HubSpot Research found that 82% of buyers consider an immediate response important, with "immediate" meaning ten minutes or less. Use that as a benchmark, not a guarantee, and measure your own before-and-after.

Contract terms that matter

Three clauses deserve close reading. First, data ownership and portability: you should be able to export conversations and knowledge-base content, and to delete data on request. Second, the SLA: uptime, support response and remedies should be written down, not implied.

Third, change control. Models, prompts and pricing change; make sure you are told in advance and can test changes before they go live. If the vendor cannot describe how updates are rolled out, assume they will happen without you.

For sector-specific contract questions, the same principles apply to a hotel AI concierge and to AI concierge ROI for SaaS; the difference is which integrations and metrics you prioritise.

What good looks like

A good AI concierge is boring in the best way. It answers from your knowledge base, says "I don't know" when it doesn't, escalates on clear triggers, captures the lead in your CRM and reports what happened. It does not pretend to be human, and it does not promise things your team cannot deliver.

Measure it on outcomes you already care about: time to first response, leads captured, appointments booked and escalations handled. If those improve without a rise in complaints, the assistant is doing its job.

See it in action

You can test the behaviours described here in a live environment: see how the bot answers from a knowledge base, escalates correctly and hands off a lead in our dental clinic demo. Type your own awkward question rather than following the suggested path.

What to do next

1. Write your own list of 12 questions from this article and send it to every shortlisted vendor before the next demo.
2. Ask each vendor to show a live answer from your own documents, not their demo content.
3. Run the one-week pilot on real traffic with a defined baseline and success metrics.
4. Compare contract terms on data ownership, SLA and change control before you sign.