What problem does it solve?
At the drive through speaker, a human order taker has seconds to parse an order over engine noise, a crackling speaker and a regional accent, remember dozens of possible customizations, and still offer an upsell, while a line of cars builds up behind. Phone ordering has the same pressure without the noise. Presto's own announcement of its Checkers & Rally's rollout cites a Franchise Times estimate that more than 80% of quick service restaurant sales come through the drive through, which is why speed of service and order accuracy at that single window matter directly to a restaurant's revenue.
On a short shift, a voice agent that takes the order reliably can free the person who would have taken it to bag orders, run the window or coach a new hire instead.
- A Presto Automation press release cites a Franchise Times estimate that over 80% of quick service restaurant sales are generated from the drive through.Checkers & Rally's and Presto Announce Largest Ever Rollout of Drive-Thru A.I. Voice Assistant in the Hospitality Industry (2022)
How does it work?
- Listen through noise and casual speech. The agent runs streaming speech recognition tuned for the drive through environment: engine and traffic noise, conversation inside the car and regional accents. Generative systems built for this job are designed to handle the huge number of ways a real customer phrases an order, casual conversation included, rather than a fixed set of recognized phrases.
- Resolve the order against the live menu. Every item, price, combo rule and limited time offer comes from the restaurant's own point of sale and inventory feed, so the agent never offers an item that is 86'd (out of stock) at that specific store.
- Ask only what is missing. Size, sauce, spice level and combo choices are confirmed with short follow up questions, the same way a trained order taker would, rather than a rigid script.
- Offer one relevant upsell. A single, contextual suggestion (a size upgrade, a drink, a dessert) is offered once, on the items the customer has already chosen.
- Confirm and transmit. The agent reads the order back, sends it to the point of sale and kitchen display exactly as a person would key it in, so the kitchen sees one consistent order format regardless of who or what took it.
- Hand over cleanly. Anything the agent is not confident about (an unusual request, a complaint, a large or clearly abusive order, a payment problem) goes to a team member at the window with the partial order already on screen.
- Audience
- Customer facing
- Autonomy
- Supervised agent
- Adoption
- Early adopters
- Channels
- Phone and voice
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Containment rate | Too few to pool | 86% | 1 | 1 organization |
Value drivers: Lower cost to serve, Customer experience, Revenue growth, Employee productivity.
Indicative value
A quick service chain with 500 corporate and franchised drive through restaurants
USD 1.8 million to USD 18.8 million
Additional annual revenue from consistent AI upselling on automated orders per year
How this is calculated
Formula: restaurants * ordersPerRestaurantPerDay * 365 * automationShare * upsellGainPerAutomatedOrder. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Drive through restaurants on the platform restaurants, restaurants | 500 | 500 | The reference chain. |
| Drive through orders per restaurant per day ordersPerRestaurantPerDay, orders per restaurant per day | 200 | 400 | Editorial assumption for a mid sized quick service drive through. Replace with your own transaction count. |
| Share of orders the agent completes without a staff member stepping in automationShare, fraction of orders | 0.5 | 0.86 | Conservative against the evidence on this page. Wendy's reports an average of 86% of pilot orders handled without team member intervention; the low end covers a first launch, a noisier market or the phone channel. |
| Extra revenue per automated order from a consistent upsell offer upsellGainPerAutomatedOrder, USD per order | 0.1 | 0.3 | Editorial assumption. Replace with your own average check data from an A/B test of the upsell prompt. |
What it leaves out: Upsell revenue only, counted on every automated order as if staff never upsold before, which overstates the incremental gain wherever staff already upsell well. It also leaves out the technology, integration and support cost, any labor savings from freeing staff to run the window, and any change in average speed of service.
Who already uses it?
2 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
The Wendy's Company
United States · Travel and hospitality · 2023
Wendy's partnered with Google Cloud to build FreshAI, a generative AI voice assistant for drive through ordering, built to handle casual conversation and the more than 200 billion ways Wendy's says a customer can order a Dave's Double, something Wendy's says a traditional rule based system could not do. After months of testing in Columbus, Ohio, FreshAI was active across four restaurants owned by the company in the Columbus, OH market at the time of this update, with franchisee pilots planned for 2024. Wendy's measures success as the share of orders completed without a team member stepping in.
- Containment rate: 86%, during the Columbus, Ohio pilot
"Our accuracy during the pilot, measured as the percentage of orders successfully handled by Wendy's FreshAi without restaurant team member intervention, averaged 86% and we would expect the average to only to increase."
Claimed by: organization
Checkers Drive-In Restaurants
United States · Travel and hospitality · 2022
In January 2022 Checkers Drive In Restaurants, operator of the Checkers and Rally's drive through chains, selected Presto as the exclusive automated voice ordering provider for its restaurants owned by the company, and Presto's systems were scheduled to be deployed across all of them during 2022, a rollout the companies called the largest of its kind in the hospitality industry at the time. The decision followed a four month, multi location pilot in 2021 in which Presto's voice ordering completed most orders with minimal staff intervention, a figure Presto itself reported. In January 2025 the US Securities and Exchange Commission found that this same press release was one of the statements that materially misled investors, because it did not adequately disclose that the voice AI technology powering Presto Voice was owned and operated by a third party ("Supplier A", referenced in the release only as a partner, Hi Auto), not Presto's own technology. This record therefore no longer carries the vendor's automation figure as a metric; see `verification.note` and this use case's risk section for the SEC order.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Live menu, pricing, combo rules and limited time offers per restaurant
- Real time stock or "86 list" feed so the agent never offers what is unavailable
- Recorded or transcribed real orders per market, including accents and slang, to tune recognition
Systems to integrate
- Point of sale system, to receive the finished order in the restaurant's normal format
- Kitchen display system, so cooks see one consistent order regardless of who took it
- Digital menu board, to show what the agent heard back to the customer
- Payment terminal, when payment is also taken at the speaker
Complexity: Medium
The speech recognition is the visible challenge, but the real work is the same as any point of sale integration: a live feed of menu, price and 86'd items per restaurant, and a reliable path to hand the order to the kitchen exactly as if a person had keyed it in.
- 1
Pilot in one market before any rollout
Run the agent in a small number of company owned restaurants first, with a live audio feed a supervisor can listen to, before offering it to franchisees.
- 2
Define exactly when it hands over
Write down the triggers for handover: repeated misunderstanding, a complaint, a clearly abusive or joke order, a payment failure, and any request the menu data does not cover.
- 3
Tune per market, not once
Accents, ambient noise and local slang differ by restaurant. Collect real failed orders from each market and retrain or reprompt against them before wider rollout.
- 4
Cap the upsell
Offer one relevant upsell per order, never repeat it if declined, and measure whether it lifts average check or lifts complaints before turning it up.
- 5
Keep a human always reachable at the window
The team member at the window should see the order building in real time and be able to take over mid order without the customer repeating themselves.
- 6
Verify the vendor's own numbers
Ask exactly how a reported automation rate is measured, whether any off site human reviews or corrects orders, and get it in writing before it goes into a business case.
Guardrails
- Hard stop and handover on abusive, absurd or clearly prank orders (for example implausible quantities)
- No item, price or combo offered outside the live menu feed for that specific restaurant
- Upsell frequency capped at one offer per order, never repeated if declined
- Masking of payment card data at the gateway when payment is taken by voice
KPIs to instrument
- Containment or automation rate by daypart, defined the same way every time it is reported
- Order accuracy verified against what the customer actually receives at the window
- Average check size on automated orders versus staff taken orders on the same menu
- Handover reasons, so recurring gaps in the menu data or recognition get fixed
Human in the loop
A team member is always available to take over at the window without the customer repeating the order, and operations staff review a daily sample of completed and handed over orders to catch misheard items before they reach a customer.
Common failure modes
- Misheard items reaching the window
- Noise, accents or menu ambiguity produce a wrong item that is not caught before pickup. Mitigate with an order readback the customer can correct and a final visual check at the window.
- A reported automation rate that hides human labor
- The US Securities and Exchange Commission found that one voice ordering vendor's proprietary units required substantial off site human order takers, and that its reported automation rate in fact excluded that off site help. Treat any vendor's automation rate as a claim to verify, not a fact.
- Upselling that annoys rather than lifts revenue
- A pushy or repeated upsell prompt lowers satisfaction even when it lifts average check in the short term. Track satisfaction and complaint mentions alongside check size.
- A pilot that never scales
- Results in one quiet, well maintained pilot restaurant do not predict a noisy, understaffed one. Expand market by market and re measure before each expansion.
What are the risks and rules?
EU AI Act
Limited risk (transparency)
Article 50(1): customers must be told they are dealing with an AI system, unless that is obvious from the point of view of a natural person who is reasonably well informed, observant and circumspect, given the circumstances and context of use. Taking a food order is not an Annex III use, so this stays a transparency obligation, not a high risk one.
Guidance
- Article 50, transparency obligations for providers and deployers of certain AI systems (European Union, Europe). People must be informed that they are interacting with an AI system unless this is obvious from the context, which is relevant to how a drive through voice agent identifies itself.
Controls to put in place
- Clear, consistent disclosure that the customer is speaking with an AI voice system
- Written, verifiable definition of any automation rate before it is used internally or externally
- Daily human review of a sample of completed and handed over orders
- Change control before any new upsell prompt or menu category goes live
- PCI DSS scope review wherever payment is taken by voice, and a check for US state biometric or voice privacy laws (for example Illinois BIPA) before storing or matching a caller's voice
When it went wrong elsewhere
- SEC cease and desist order against Presto Automation over drive through voice AI claims. In January 2025 the US Securities and Exchange Commission found that, from 2021 to 2023, Presto Automation made materially misleading statements about its Presto Voice product for drive through ordering: units powered by its proprietary technology required substantial human order takers based abroad, and its reported automation rate in fact referred only to orders completed without on site restaurant staff, not without any human help.
Frequently asked questions
- What share of drive through orders can a voice agent complete without staff help?
- Wendy's reports that during its FreshAi pilot, orders handled without a restaurant team member stepping in averaged 86%. Results depend heavily on menu complexity, market noise and how the rate is defined, so ask any vendor exactly how they measure it.
- Is drive through voice ordering high risk under the EU AI Act?
- Usually not. Taking a food order is not one of the high risk uses listed in Annex III, so it falls under the Article 50 transparency duty: customers must be able to tell they are speaking with an AI system.
- Can a vendor's reported "automation rate" be trusted at face value?
- Verify the definition first. The US SEC found that one vendor's reported automation rate excluded substantial off site human order takers, so the same words can describe very different levels of automation between vendors.
- Does voice ordering replace the person at the drive through window?
- Not on Wendy's own account. Wendy's describes FreshAi as an assistant, not a replacement, meant to let crew members focus on preparing and serving food and building the relationships that bring customers back. A sensible design still keeps a team member able to take over when the conversation goes outside what the agent can handle, but that handover point is editorial advice here, not something Wendy's states.
How to cite this page
Blits.ai AI Use Case Library, "AI voice agent for drive through and phone order taking", last verified 30 September 2026, https://www.blits.ai/ai-use-cases/drive-thru-voice-ordering-agent. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 30 September 2026: Published after review by an automated review workflow (independent skeptic review).
- 30 September 2026: Fixed adversarial review blockers: dropped the problem section's unsourced superlative; narrowed the Blits.ai section's PII masking, card detection, test suite and handover claims to what the platform actually covers for a phone call (transcribed text, DTMF or payment link, call transfer, live takeover), and to the analytics the dashboard actually reports (per bot dashboards, conversation logs, custom statistics, not containment or handover reason metrics). Also corrected the EU AI Act Article 50 basis to the regulation's own "obvious" test, added GDPR to risk.regulations, corrected Wendy's product name to "FreshAi" throughout, trimmed the metaDescription and FAQ 4 to what Wendy's source states.
- 30 September 2026: Unpublished by an automated review workflow (independent skeptic review).
- 29 September 2026: First published