Compare

The buyer's checklist for a restaurant AI phone agent.

Twelve questions that separate serious phone-ordering systems from voice-bot marketing sites. Bring this to any demo.

Get StartedTry the Live Demo
The short answer

Before buying an AI phone agent for a restaurant, verify structured menu capture, server-side pricing, deterministic confirmation semantics, sold-out awareness, modifier enforcement, transfer paths, transcript access, POS write-back honesty, internal queue fallback, per-location scoping, plan and concurrency clarity, and provider transparency. Every 'yes' should be reproducible on a live call.

Updated By Corey Mack — Founder, Fire ItReviewed by Fire It Editorial — product review

The twelve questions

Any 'no' on the first ten is a red flag. A 'yes' on the last one is a red flag.

  • Is every item captured as a structured field bound to my live menu?
  • Is pricing computed on the server against my menu, not by the model?
  • Are required modifier groups enforced, with min/max selections?
  • Are 86'd items automatically excluded from what the agent offers?
  • Is the caller told 'confirmed' only after the authoritative state records?
  • Can callers request a human transfer at any point?
  • Are transcripts and structured events available for QA?
  • If certified POS write-back is unavailable, does the ticket land in a real internal queue?
  • Is the data model scoped per location, with row-level enforcement?
  • Are plans and concurrency limits documented in public?
  • Is the underlying voice provider disclosed, or hidden behind marketing?
  • Does the vendor over-promise perfect accuracy?

How Fire It answers each

Fire It's product pages spell out how we handle each of these. Structured capture and server-side pricing are on the AI phone-ordering feature page. Deterministic confirmation is on the accuracy page. The internal queue and POS honesty story is on the integrations page. Concurrency and plan boundaries are on pricing.

Two questions we hope you ask us specifically

What happens when the POS is down? — the ticket lands in the internal queue with a POS-unavailable flag; nothing is lost. What happens when the AI mishears the address? — the agent reads back the running order and asks for confirmation; the caller can correct any field as a structured update, not a free-text note.

How to run the checklist as a two-hour vendor eval

Print the twelve questions. Book a live call with each finalist. Read each question during the call and score yes, no, or 'demo-only.' A finalist who scores twelve yeses on a live call is worth a pilot. A finalist who scores yeses only when you're on a marketing site is worth skipping. Two hours of this discipline is worth more than a month of RFP responses because it forces vendors to demonstrate the behaviors they claim on paper.

  • Structured menu binding — order a modifier-heavy item and inspect the resulting ticket.
  • Server-side pricing — order the same combo two ways and compare totals.
  • Deterministic confirmation — ask when the caller hears 'confirmed' and confirm the state.
  • Sold-out enforcement — ask for an 86'd item and observe the response.
  • Transfer semantics — ask for a human and confirm the transfer number rings.

Red flags that don't fit neatly on the checklist

Two red flags fall outside the twelve questions but matter as much. First: a vendor who won't disclose their underlying voice provider is often hiding a switch mid-contract that changes pricing. Second: a vendor who promises perfect accuracy is either lying or misunderstanding the technology. Neither disqualifies a vendor on its own, but both should surface in your due diligence. Fire It discloses its underlying provider on the AI disclosure page and never claims perfect accuracy — the accuracy page walks through the honest metrics.

What a good pilot looks like after the checklist

After a passing checklist score, the pilot should run at overflow-only for two weeks against a real number, with daily transcript review and a weekly ROI check. The vendor should provide a lightweight readout of what the AI heard vs what the transcript captured, not a marketing dashboard. If the pilot readout is glossy and vague, that's a fourth red flag. Fire It's pilot readouts are boring on purpose — capture count, exceptions, transfer count, and pickup-time accuracy.

Good fit if

Where Fire It actually helps

  • You're evaluating multiple voice AI vendors.
  • You want a repeatable test on every demo.
Honest limits

What we don't claim

  • A checklist isn't a substitute for a live pilot.
Keep going

Related on Fire It

See it in practice

Start on Fire It or try the live Neon Slice demo — no card required.

Get StartedTry the Live Demo