Everyone quotes an accuracy rate. Nobody says what it measures.
Two voice-AI deployments reported in the same week publish intervention rates an order of magnitude apart. Neither publishes a definition — and the definition is the only part a franchisee pays for.
Two stories about drive-thru voice AI landed within twenty-four hours of each other this week, and read side by side they expose the weakest number in the whole category. Both describe systems that take orders. Both report how often a human has to step in. The two figures differ by roughly a factor of ten, and nothing published by either party explains why.
This is reference information, not legal, tax or investment advice. Figures below are published by third parties and compiled as-is; the performance numbers are vendor and brand claims, not audited results. Check the sources before acting on any of them.
The two numbers
On 21 September, Restaurant Dive reported that Presto Phoenix had joined Toast's partner ecosystem, letting restaurants on Toast's point-of-sale deploy voice AI without substantial new hardware. Presto puts its product at roughly 95% order accuracy, with worker intervention on about 10% of orders (source: Restaurant Dive, 2026).
The day before, Franchising.com reported that Bojangles' voice assistant, Bo-Linda, is live at 452 of the chain's 800-plus restaurants, and that human intervention is required 1–2% of the time (source: Franchising.com, 2026).
One deployment therefore says a person touches one order in ten. The other says a person touches one order in fifty to a hundred. These are not small differences in tuning. They are differences large enough that the two numbers are almost certainly counting different events.
An accuracy rate without a definition is not a performance figure. It is a marketing figure that happens to have a decimal point.
What the word could mean
There is no industry standard for what a voice-AI vendor means by accuracy, which leaves at least four plausible definitions in circulation, each producing a very different number from the same lane:
—Orders transcribed correctly, measured against the audio — a speech-recognition metric, and the most flattering one.
—Orders rung into the point-of-sale exactly as the customer intended, including modifiers and substitutions.
—Orders completed without a human touching the screen at all.
—Orders that left the window correct, which is the only one the customer experiences.
A system can plausibly score 95% on the first and far worse on the last, because the errors that survive to the window are the ones nobody caught. Neither of this week's reports says which definition is in use, and neither brand is obliged to say.
The number a unit actually needs
For a franchisee, none of the above is the operative figure. The operative figure is what the intervention costs. An intervention rate of 10% at a lane doing 400 cars a day is forty moments in which a crew member stops what they are doing to rescue an order. Whether that is a saving or a cost depends entirely on whether the labour it frees at the headset exceeds the labour it consumes at the recovery — and that arithmetic is unit-specific, which is exactly why no vendor publishes it.
The wider deployments already reported by this publication make the stakes concrete rather than theoretical: Taco Bell's voice AI is running at more than 890 US lanes, and Bojangles is at 452 restaurants of which 280 are company-owned, meaning roughly 172 franchised units are already living with whatever the real number is.
Four questions before a mandate
—What exactly does your accuracy figure measure, and at which point in the order is it measured?
—What is the intervention rate at a unit with my volume and my menu complexity, not the network average?
—Who measured it — the vendor, the franchisor, or an independent party — and can I see the method?
—What is the remedy if my unit's rate is materially worse than the one I was shown?
A brand that can answer all four has run a real pilot. A brand that offers a single percentage and a testimonial has run a procurement process. The distinction costs nothing to establish before signing and is close to impossible to establish afterwards.