If you run PAR Brink, the question that decides whether a voice agent earns its keep is narrow: does a completed phone order arrive in Brink as a ticket the line can work from, priced the way the register would have priced it? The voice is the part everyone demos and the least important part of the purchase. A convincing voice that fires a wrong modifier costs more than an unanswered ring, because now you make the food twice and you have annoyed an existing customer.
X1 Voice reaches Brink through Deliverect. Your published items, prices and modifier groups come across, the agent prices each call against them, and the finished order is written back as a normal order. Nobody re-types anything at the counter.
On a Brink estate that detail matters more than it would for a single independent, because Brink stores are usually part of a brand with shared menu governance and very tight peak-hour capacity.
What the Deliverect hop changes
Three consequences, all of them practical.
The store's menu has to be published to Deliverect before the agent can read it. If the brand already pushes to delivery marketplaces, this exists and setup is short. If it doesn't, publishing is the real first task and it belongs to whoever owns menu configuration, not to a store manager.
Menu changes propagate on a sync rather than the instant you save them. Price updates are fine on that cadence. A sold-out item on a Saturday night is not, so settle the 86 path before you rely on the agent to tell callers what is unavailable. Voice AI and POS 86 sync failure modes covers what actually breaks and what to ask a vendor about it.
Item and modifier names arrive as configured. QSR menus are full of names built for a fixed button grid: abbreviations, size prefixes, internal codes for LTO items. The agent inherits them and will read them aloud. That is the most common reason a technically working agent sounds wrong on day one, and it is a naming fix rather than a software problem.
The peak-hour question specific to a QSR
Most operators evaluate a phone agent as a labor question. On a high-volume Brink store it is at least as much a capacity question.
Today, some share of your phone calls go unanswered during the lunch and dinner peaks. Nobody logs those, which is why they feel smaller than they are. Count a typical week's missed calls, multiply by your average ticket, and see what the number looks like.
Turn on an agent and those calls get answered. That is the point, and it also means more tickets hitting the same line during the same fifteen minutes that the drive-thru is stacked. If your kitchen is already at its ceiling at 12:20, the agent does not create capacity, it converts a lost call into a longer wait for everyone.
The fix is a policy decision made in advance, not a surprise in week two. Quote a longer pickup time during peak, cap phone orders per interval, or route phone tickets to a separate prep position. Any of those work. Discovering the problem through complaints does not.
The failure mode nobody demos
Modifier trees built for a trained crew member.
A required modifier group with a default that only makes sense with a human at the screen, a size implemented as three separate items, a combo that assumes the operator taps things in a fixed order. None of that hurts when someone who has worked the station for six months is ringing it in. An agent follows the structure literally, so it will either refuse a combination that is obviously fine or fire a ticket that validates and is wrong at the pass.
The correction is structural. Simplifying modifier trees covers the shapes that break. On a branded estate this is good news, because the fix happens once centrally and every store inherits it.
Ask the vendor to take a deliberately awkward order on your real menu: substitution, a hold, an add-on, a combo, and a promo code, all in one call. Then walk to the printer. Do not accept the agent reading the order back correctly as proof of anything. The repeat-back and the injected ticket are different systems, and only one reaches the kitchen. That distinction is why integration depth decides this purchase and voice quality does not.
Roll it out one store at a time
Even though the configuration would let you flip fifty stores at once, don't.
Pick one store with real phone volume and a manager who will report problems rather than quietly turn it off. Run three weeks. Read a weekly sample of transcripts instead of watching a summary screen, and keep a running list of every phrase the agent handled badly. Nearly every entry will be naming or modifier structure, and nearly every fix applies estate-wide.
Then pick a second store that is deliberately unlike the first, ideally with a different daypart mix. The problems that appear at store two are the ones that would have embarrassed you at scale. Designing a voice AI pilot sets out what to agree on before you start so the pilot ends in a decision.
The four numbers to check after a month
Answered-call rate, which should sit near 100 percent and is the reason you bought it. Order accuracy, measured by reading a sample of phone tickets against the transcripts rather than trusting a dashboard, following the order accuracy measurement method. Escalation reasons, split into "by design" and "by defect," where only the defect bucket is a work list. And phone revenue per store against the same weeks last year.
That last number is the honest one. If answered calls climbed and phone revenue did not, the agent absorbed calls staff were already answering. That is real labor value and a different business case than incremental sales, and on a multi-unit brand the difference is what gets the rollout funded or shelved.
Setup runs under 24 hours per store once the menu is published, and plans start at $250 per month per location, which you can size against your own volume on the pricing page. If your portfolio mixes Brink with other systems, that is ordinary: X1 Voice connects directly to Square, Clover and OrderCounter and reaches the rest of the supported POS list through Deliverect, so one configuration covers a mixed estate.
Start with one store and one hard order. If five test calls produce five tickets you would have been glad to see a new crew member ring in, the integration is doing its job. If three do, the gap is almost certainly in your item names and modifier groups, and that is worth fixing whatever you decide about voice.