Last updated: July 14, 2026. Financial data reflects Wendy’s FY2025 Form 10-K (fiscal year ended December 28, 2025) and the Q1 2026 Form 10-Q (quarter ended March 29, 2026). Franchise economics reflect the Quality Is Our Recipe, LLC 2025 Franchise Disclosure Document, issued March 28, 2025. Where a figure changes frequently, we state its as-of date. Where the public record is silent, we say so rather than guessing.

Advertisement

Key Takeaways

  • FreshAI is the most durable voice-AI drive-thru program in the quick-service restaurant industry. McDonald’s shut off its IBM-powered order taker in every test restaurant by July 2024. Presto Automation, which supplied Checkers and Del Taco, collapsed after the SEC brought its first AI-washing enforcement action against a public company. Wendy’s system is still running, still expanding, and still serves as Google’s flagship reference deployment. Surviving is the achievement here, and it is not a small one.
  • The economics are knowable, and almost nobody has done the math. Wendy’s 2025 Franchise Disclosure Document prices FreshAI at $1,667 per restaurant per month — $20,004 a year. Against the FDD’s average franchised restaurant sales of $2,108,454, that fee equals roughly 0.95% of sales. FreshAI must therefore generate approximately 95 to 100 basis points of incremental restaurant profit just for the average franchisee to break even. This is the real story, and it reframes everything else.
  • The fee is regressive, and that — not technophobia — is the most likely explanation for slow franchise adoption. A flat $20,004 costs a $1.5 million restaurant 1.33% of sales but a $3 million restaurant only 0.67%. Wendy’s rolled this out to operators during a period of falling traffic. The 500–600 unit target was never going to survive contact with that pricing structure.
  • The 86% and 99% figures everyone quotes are from a four-restaurant 2023 pilot, and the 99% number counts human-rescued orders as successes. They are not current, not system-wide, and not what most people think they measure.
  • Wendy’s has never quantified FreshAI in an SEC filing. The FY2025 10-K discusses AI only as generic competitive-landscape language. There is no unit count, no capex line, no margin attribution. Everything the market believes about FreshAI’s scale comes from press releases, blog posts, earnings-call color, and Google’s marketing — none of it audited.
  • There is little proprietary Wendy’s technology inside FreshAI. The language model, speech recognition, speech synthesis, and dialogue orchestration are Google’s. Google has since productized the capability as a “Food Ordering AI Agent” any competitor can license. Wendy’s real assets are menu logic, POS integration, and a two-to-three-year operational head start. Those are valuable. They are not a technology moat.
  • The most underpriced risk is biometric privacy law, not technology risk. Illinois treats a voiceprint as a biometric identifier, with statutory damages of $1,000 to $5,000 per violation, per person, and a private right of action. A drive-thru voice AI case against McDonald’s already survived a motion to dismiss on its core consent claim. Wendy’s has disclosed nothing about FreshAI voice data practices.

The Thesis: The Technology Works. The Pricing Is the Problem.

Most FreshAI coverage argues about whether the AI is any good. That is the wrong argument, and it has consumed three years of commentary that should have been spent on arithmetic.

By the standards of its own industry, FreshAI works. It has survived a graveyard that swallowed McDonald’s, humiliated Taco Bell, and produced a securities-fraud action against Presto. In this author’s own repeated use across the drive-thru — detailed in our companion piece, FreshAI and Wendy’s Stock — the customer experience has been consistently strong. That is a qualitative field observation, not a statistical study, and we label it as such. But it points in the same direction as the survival evidence: this is a competently engineered system.

The problem is not the technology. The problem is that Wendy’s priced it at $20,004 a year and asked franchisees fighting a 7.8% same-restaurant sales decline to absorb it.

That single design decision explains the missed rollout target, the franchise hesitancy, and the strategic drift better than any theory about AI capability. And it points to a fix that is entirely within Wendy’s control — which is why this story is more interesting than the “AI hype versus AI backlash” frame it usually gets stuffed into.

Key Numbers

Figure What it is Source
$1,667 / month
($20,004 / year)		 What FreshAI costs a franchisee, per restaurant, on top of the existing technology fee 2025 Franchise Disclosure Document
0.95% of sales What that fee equals at the FDD’s $2,108,454 average franchised restaurant BuyWendys calculation from FDD
~95–100 bps Incremental restaurant profit FreshAI must generate for the average franchisee to break even BuyWendys calculation
1.33% vs. 0.67% The same flat fee as a share of sales at a $1.5M restaurant vs. a $3M restaurant — the fee is regressive BuyWendys calculation
423 Company-operated U.S. restaurants, out of 5,969 total — which is why the 500–600 target was unreachable without franchisees FY2025 Form 10-K
86% / ~99% Orders completed with no employee intervention / “success rate” including human-assisted orders — both from a four-restaurant 2023 pilot, never updated Wendy’s Square Deal blog, Dec 2023
Zero FreshAI figures quantified in any Wendy’s SEC filing FY2025 10-K; Q1 2026 10-Q

What This Guide Covers

How Much Does Wendy’s FreshAI Cost?

Wendy’s FreshAI costs franchisees $1,667 per restaurant per month — $20,004 per year — according to the Quality Is Our Recipe, LLC 2025 Franchise Disclosure Document. The fee is invoiced quarterly in arrears and charged on top of the restaurant’s existing technology fee. Against the FDD’s average franchised restaurant sales of $2,108,454, that works out to roughly 0.95% of annual sales — meaning FreshAI must generate approximately 95 to 100 basis points of incremental restaurant profit before the average franchisee breaks even.

That is the number the entire FreshAI debate should have been organized around, and almost no coverage has reported it. The figures below come from a primary document most FreshAI commentary has never opened.

Why a franchisee might pay it anyway. Wendy’s has stated that roughly 75% to 80% of its customers use the drive-thru, and that the menu presents “billions of possible order combinations” once customizations, substitutions, sizes, combos, and limited-time offers are accounted for. That combination — enormous channel concentration plus combinatorial complexity — is the entire strategic justification for FreshAI, and it is why rules-based voice systems never worked and large language models suddenly could. It is also why McDonald’s, which started in 2021 on an earlier generation of technology, failed. Sourcing note: the 75–80% figure is management’s characterization from Wendy’s May 2023 press release, not an audited disclosure, and may have shifted with digital and delivery growth.

The disclosed fee

  • $1,667 per restaurant per month, invoiced quarterly in arrears.
  • $20,004 per restaurant per year.
  • This sits on top of the restaurant’s existing technology fee.
  • The FDD states FreshAI may become mandatory for new restaurant builds in the future.

What that fee represents as a percentage of sales

The same FDD reports average 2024 annual gross sales of $2,108,454 across the 5,325 U.S. franchised restaurants in its full-year sales table, with a median of $1,984,382. Run the division:

Annual restaurant sales FreshAI annual fee Fee as % of sales
$1,500,000 $20,004 1.33%
$1,984,382 (FDD median) $20,004 1.01%
$2,108,454 (FDD average) $20,004 0.95%
$3,000,000 $20,004 0.67%

The conclusion this forces

FreshAI must create roughly 95 to 100 basis points of incremental restaurant-level profit before the average or median franchisee breaks even on the subscription alone. That is before hardware, installation disruption, crew training, and ongoing support.

Put plainly: the fee is not a rounding error on a franchisee’s P&L. It is economically equivalent to nearly a full percentage point of sales, layered onto a charge stack that already includes a 4% royalty, 3.5% national advertising, 0.5% local advertising, and a base technology fee. Our full breakdown of that stack — and the value-sharing thresholds Wendy’s would need to clear before extracting any more — is in Wendy’s FreshAI Economics: Franchise Margin & Fee Analysis.

And here is the design flaw

The fee is flat, so it is regressive. A $3 million restaurant pays 0.67% of sales. A $1.5 million restaurant pays 1.33% — double the effective rate — for the same product.

That is precisely backwards relative to where the value lands. High-volume restaurants have more transactions to spread the fixed cost across and more peak-hour congestion for the technology to relieve. Low-volume restaurants pay a higher effective rate and have fewer transactions from which to generate offsetting labor savings or upsell.

A single national price therefore produces attractive returns at high-volume stores and destroys value at low-volume ones. If Wendy’s wanted a pricing structure engineered to stall adoption across the bottom half of its system, it could not have designed a better one.

The fee is flat, so it is regressive. A $1.5 million restaurant pays double the effective rate of a $3 million restaurant — for the same product, with fewer transactions to earn it back.

Is FreshAI worth it for a franchisee?

It depends almost entirely on restaurant volume, and Wendy’s has published no data that would let an operator answer the question with confidence. Here is the decision framework, stated as plainly as the evidence permits:

  • At $3 million in sales: the fee costs 0.67% of sales. FreshAI needs to clear roughly 70 basis points of restaurant-level profit to pay for itself. Peak-hour congestion is likely severe enough that throughput gains alone could justify it. Plausibly worth it.
  • At the $2.1 million FDD average: the fee costs 0.95%. Break-even requires ~95–100 bps. A genuine coin flip, and unanswerable without controlled data Wendy’s has not released.
  • At $1.5 million: the fee costs 1.33% of sales, and the restaurant has fewer transactions across which to generate offsetting labor savings or upsell. Likely value-destroying at the current price.

The honest answer to “is it worth it” is therefore: nobody knows, including Wendy’s — or if Wendy’s knows, it has not said. No matched-control study comparing FreshAI restaurants to similar non-FreshAI restaurants has been published. No labor-hours reconciliation exists. No incremental contribution-profit figure has been disclosed. A franchisee is being asked to underwrite roughly 100 basis points of cost against an entirely unquantified benefit.

That is not a technology problem. It is a disclosure problem, and it is the reason a rational operator waits.

Why the 500–600 Target Was Never Going to Happen

In February 2025, on the Q4 2024 earnings call, then-CEO Kirk Tanner said FreshAI was live at roughly 100 restaurants and targeted 500–600 by year-end 2025.

Two facts make that target very difficult to reach.

Fact one: the arithmetic. Per the FY2025 10-K, as of December 28, 2025 the U.S. system had 5,969 restaurants — of which only 423 are company-operated (about 7.1%), with 5,546 franchised across 203 franchisees.

Wendy’s rolled FreshAI out to company-operated restaurants first. That is the standard playbook: the company controls those P&Ls and needs nobody’s permission. But there are only 423 company-operated U.S. restaurants in total. Even at 100% penetration of every company store in America, Wendy’s would still fall short of the 500-unit low end of its own target.

Even at 100% penetration of every company-operated restaurant in America, Wendy’s would still fall short of the low end of its own target. The 500–600 goal was mathematically unreachable without franchisees.

The 500–600 goal was mathematically unreachable without substantial franchise adoption.

Fact two: the pricing. Franchise adoption required operators to voluntarily add $20,004 per restaurant per year — a fee costing low-volume stores 1.33% of sales — during a period when U.S. same-restaurant sales were falling. Wendy’s reported U.S. comps down 7.8% in Q1 2026.

Put those together and franchise hesitancy stops looking like resistance to change. It looks like competent operators running the numbers and reaching a defensible answer. A rational franchisee facing negative comps does not add a near-1%-of-sales fee on a promise of unquantified benefit. They wait for someone else to prove it.

The rollout did not stall because the AI is bad. It stalled because the fee is high, flat, and regressive, and the benefit is unproven at the unit level. That is a solvable problem — and solving it is the most valuable thing new CEO Bob Wright could do with FreshAI.

A necessary label on that claim. Wendy’s has never stated why franchise adoption lagged, and no franchisee survey on FreshAI pricing exists in public. The pricing explanation is a BuyWendys.com inference — the most parsimonious one available from the evidence, since it requires only three verifiable facts (a $20,004 flat fee, a regressive incidence curve, and a 7.8% comp decline) and no assumptions about operator psychology. But it is an inference, and we label it as one. The alternative explanations — corporate distraction through three CEO transitions, deliberate strategic de-prioritization, or technical deployment constraints we cannot see — are not excluded by anything in the public record. What would distinguish them is disclosure: a franchise adoption count, or a change in the pricing structure. Watch for both.

The 86% Number: What It Actually Measures

These figures have been repeated so widely, for so long, that they have hardened into assumed fact. Here is what Wendy’s actually disclosed in its December 2023 Square Deal blog update, from a four-restaurant pilot in Columbus, Ohio:

  • 86% of orders completed without team member intervention.
  • Approximately 99% “success rate” — where success means the order was started by the chatbot and successfully submitted to the point-of-sale system, even if a human employee joined the conversation to fix it.
  • 22 seconds faster than the Columbus market average. Note the comparison base: this measures four heavily-supported pilot restaurants against the average of an entire local market — not against matched control restaurants. It is a favorable comparison by construction, and it deserves the same skepticism we apply to the 86% figure below.

Read the second bullet again.

The 99% figure is not an accuracy metric. It is a completion metric with the human safety net explicitly included. An order where the AI misheard the customer, entered the wrong items, and a crew member rebuilt it from scratch still counts as a success, provided it eventually reached the POS. That is a legitimate operational measure — it tells you the system rarely hard-fails and strands a customer. It tells you nothing about how good the AI is. It is closer to a measure of how good the fallback is.

The 86% figure is the one that maps to money. If 14% of orders need a human, a human must remain within earshot of the drive-thru at all times. You have not eliminated a position. You have changed what that position does — which may still be valuable, but it is a different economic claim, and it is a much harder one to convert into a franchisee’s P&L.

Three reasons to discount both numbers in 2026

1. They are pilot metrics from four restaurants in Columbus — Wendy’s headquarters market — with presumably heavy corporate attention and dedicated support. Pilot performance beats scaled performance in essentially every operational technology deployment ever measured.

2. They have never been refreshed. It has been more than two and a half years. Wendy’s has had every incentive to publish a better number if it had one. Google’s marketing cites a “95% success rate” at larger scale — a different definition again, from the vendor rather than the customer. Wendy’s own May 2025 FreshAI blog update cited daily order volume instead of accuracy. Volume is a much easier number to make impressive.

3. Soft definitions are now a regulatory red flag, not a footnote. That requires its own section.

The Presto Precedent: Why “Success Rate” Is Now a Loaded Term

On January 14, 2025, the SEC brought what it described as its first AI-washing enforcement action against a publicly traded company. The target was Presto Automation, supplier of drive-thru voice AI to Checkers, Rally’s, and Del Taco.

The findings:

  • Presto’s system relied on offshore human agents, primarily in the Philippines and India, to complete orders.
  • In the advanced pilot version, from June through December 2023, a human agent entered the orders approximately 70% of the time.
  • Presto publicly disclosed a system-wide “non-intervention” rate of 85% — counting only restaurant staff intervention. The offshore agents doing the actual work were not counted as intervention at all.

Presto’s headline number — 85% non-intervention — is almost identical to Wendy’s 86%. The entire difference lives in the definition of “intervention.”

To be explicit and fair: there is no evidence that Wendy’s has done anything remotely like this. No filing, no reporting, and no source we examined suggests offshore human agents or any comparable arrangement in FreshAI. Wendy’s disclosure is on its face more honest than Presto’s — Wendy’s told you its 99% figure included human-assisted orders. Presto concealed the equivalent fact.

But the Presto action changes what a serious analyst is obligated to ask. “Success rate” and “non-intervention rate” are not standardized terms; they mean whatever the issuer defines them to mean. The SEC has now demonstrated that the gap between the marketing number and the operational truth can be enormous — 85% claimed versus 70% human-entered — and that the gap is actionable securities fraud.

The correct posture: accept no QSR voice-AI performance figure without its definition attached. For FreshAI, the outstanding question is simple, specific, and answerable:

What percentage of FreshAI orders, across all deployed restaurants including franchise units, are completed today with zero human involvement of any kind?

Wendy’s has not answered that since 2023. Until it does, the labor-savings case — and therefore the franchisee ROI case — is unproven.

What Wendy’s Has Never Told the SEC

Here is a fact that reframes the entire disclosure question, and it is one we verified directly.

Wendy’s FY2025 Form 10-K contains no FreshAI quantification whatsoever. The filing discusses artificial intelligence only in generic competitive-landscape and risk-factor terms. There is no unit count. No capital expenditure line. No margin attribution. No performance metric. The Q1 2026 10-Q is the same.

This matters for two reasons, and they cut in opposite directions.

It softens one accusation. Wendy’s did not “go quiet” on FreshAI unit counts after missing its target. Wendy’s never disclosed FreshAI in its SEC filings at all. The 100-restaurant figure and the 500–600 target came from an earnings call. The “nationwide” and “tens of thousands of orders daily” claims came from a company blog. The 24-states and 95%-success figures came from Google’s marketing. Every single thing the market believes about FreshAI’s scale and performance sits outside the audited financial statements.

It sharpens a different one. Companies measure what they intend to monetize, and they disclose what they intend to be judged on. A program material enough to charge franchisees $20,004 a year — and potentially to mandate on new builds — is material enough to quantify for shareholders. The absence is a choice.

For investors, the practical implication is a discipline: tier every FreshAI claim by source before using it. SEC-verified, company communication, partner marketing, or trade press. Almost all of them fall into the last three buckets. We apply that tiering throughout our FreshAI coverage, including the demand-side analysis in FreshAI’s Hidden Demographic Edge.

How FreshAI Actually Works: An End-to-End Teardown

Methodological disclosure: neither Wendy’s nor Google has published an architecture diagram, reference architecture, or technical specification for FreshAI. We found no patents from The Wendy’s Company or Quality Is Our Recipe, LLC covering this system. What follows is a reconstruction, with every element labeled by evidence class. Where we infer from Google’s general product stack rather than a Wendy’s-specific confirmation, we say so.

Part 1: Capturing the voice (hops 1–4)

The path from a car pulling up to a machine-readable transcript. This is the least documented and, per Wendy’s own CIO, the hardest part of the system — engine noise, wind, car stereos, and back-seat passengers all competing with the person actually ordering.

# Stage What happens Evidence class
1 Vehicle detection A presence sensor or inductive loop at the menu board detects an arriving car and triggers the session. Inferred. Standard drive-thru infrastructure; not confirmed for FreshAI.
2 Audio capture Existing drive-thru microphone and outdoor speaker capture the customer’s voice. Inferred. Wendy’s confirmed “integration with restaurant hardware” but specified no microphone-array upgrade.
3 Signal conditioning Noise suppression, echo cancellation, speaker isolation. The hardest problem in the chain: engine noise, road noise, wind, car stereos, and back-seat passengers all compete with the speaker. Challenge confirmed, method inferred. Wendy’s CIO Kevin Vasconi publicly discussed ambient noise, dialect, and extreme customization as the core difficulties.
4 Speech-to-text (ASR) Audio is transcribed to text. Confirmed that ASR is used — Vasconi: “speech to text in our particular case.” The specific Google engine is inferred, not confirmed.

Part 2: Understanding and deciding (hops 5–9)

Everything above happens to sound. Everything below happens to meaning — and this is where Wendy’s proprietary contribution actually lives. Note the pattern in the evidence column: the reasoning layers are Google’s and confirmed only in general terms, while the layer Wendy’s owns (menu grounding) is the one described most specifically in the public record.

# Stage What happens Evidence class
5 NLU and dialogue management Intent extraction plus an order state machine tracking additions, removals, and modifications across the conversation. Confirmed at platform level. Google lists FreshAI as powered by its “Food Ordering AI Agent.” Whether the implementation is Dialogflow CX or Vertex AI Agent Builder is not disclosed.
6 LLM reasoning Handles combinatorial complexity — parsing “no pickles, extra onion, make the drink a Frosty instead” into structured modifications. Confirmed generically. Wendy’s said “Google’s foundational LLMs.” PaLM 2-era at 2023 launch; the platform has since moved to Gemini. The production model variant is not disclosed. Any article naming a specific variant is inferring.
7 Menu and inventory grounding The model is constrained to real SKUs, current pricing, active LTOs, and item availability — including slang mapping (“shake” → Frosty; “JBC” → Junior Bacon Cheeseburger). Confirmed. Wendy’s cited “the data from Wendy’s menu”; Google’s blog used the Frosty example directly. This is the heart of Wendy’s actual contribution.
8 Guardrails Business rules preventing impossible orders and keeping the conversation on-task. The defense against the “18,000 cups of water” trolling that embarrassed Taco Bell. Confirmed generically. Wendy’s cited “business rules and logic for conversation guardrails.”
9 Upsell logic Contextual suggestive selling — the primary claimed mechanism for check lift. An AI never skips the upsell during a rush. Claimed benefit. Magnitude never quantified in any primary disclosure.

Part 3: Executing the order (hops 10–13)

Where the conversation becomes a transaction. Hop 13 is the one that matters commercially: the human handoff is what separates the 86% figure from the ~99% figure, and it is why an employee must remain within earshot of the drive-thru regardless of how good the AI gets.

# Stage What happens Evidence class
10 Order confirmation Order displayed on a digital menu board synchronized to the voice conversation. Confirmed. Note: digital menu board installs are frequently conflated with FreshAI installs in press coverage. They are not the same thing.
11 POS injection Completed order submitted into the point-of-sale system. Confirmed generically. POS vendor not disclosed — a meaningful gap, since POS integration is where most QSR technology deployments actually die.
12 Kitchen display Order routes to the kitchen for made-to-order prep. Inferred.
13 Human handoff Crew member joins to correct, complete, or rescue the order. This path defines the gap between 86% and 99%. Confirmed. Explicitly described in Wendy’s own disclosure.

Edge versus cloud: an unresolved question with real cost consequences

Wendy’s and Google consistently describe FreshAI as running on Google Cloud’s foundational models. No source confirms any in-store edge inference. Three consequences:

  • Latency. Cloud round-trips add network transit to every conversational turn. Humans perceive delays above roughly 200–300 milliseconds as awkward. Reported complaints about the system cutting customers off after brief pauses are consistent with aggressive endpointing — often a symptom of a system compensating for latency spent elsewhere.
  • Connectivity dependency. If inference is cloud-side, the restaurant’s broadband circuit becomes a hard dependency for its highest-revenue channel. Wendy’s has published no network architecture, bandwidth requirement, failover design, or SD-WAN profile. What happens to the drive-thru when the circuit degrades is, publicly, unknown.
  • Cost structure. Cloud inference is a per-order variable cost. Edge inference is capital amortized across orders. Which one Wendy’s is on materially changes whether that $20,004 fee is a high-margin software stream or largely a pass-through — a question we examine in the FreshAI economics analysis, and one Wendy’s has never answered.

Contrast: CKE’s vendor path (Valyant AI, later acquired by ConverseNow) explicitly used on-premises hardware. Different architecture, different cost curve, different failure modes. Wendy’s cloud-native choice is a real architectural bet that has never been publicly examined.

Build Versus Buy: Where Does the IP Actually Live?

Layer Owner
Large language model Google
Speech-to-text Google
Text-to-speech Google
Dialogue orchestration (“Food Ordering AI Agent”) Google
Cloud infrastructure Google
Menu catalog, business rules, LTO logic, slang-to-SKU mapping Wendy’s
POS and restaurant hardware integration Wendy’s
Rollout governance, crew training, operational learning Wendy’s
The “Wendy” persona and brand voice Wendy’s

Wendy’s contributions are real. Menu grounding — teaching a model that a customer asking for a “shake” wants a Frosty — is exactly the unglamorous, domain-specific work that separates a system that functions from a demo that impresses. The FDD itself describes FreshAI as a combination of a third-party food-ordering AI agent plus Wendy’s proprietary menu logic and recommendation engine. That proprietary layer is real, and it is the thing Wendy’s is actually charging for.

But it is not defensible technology, and the structural problem is this: Google has productized the underlying capability. The “Food Ordering AI Agent” is a listed vertical AI agent on Google Cloud, and Google markets Wendy’s as the reference customer proving it works. Google’s commercial incentive is to sell that same capability to every other chain with a drive-thru and a budget.

Wendy’s paid to be the beta customer, generated the proof points, and helped Google build a product it will now sell to Wendy’s competitors.

The honest characterization: a two-to-three-year head start in operational learning, not a proprietary technology moat. The head start is genuine — competitors licensing the agent in 2026 still must do their own menu grounding, POS integration, crew training, and failure discovery, which is a year or more of work. But it is a decaying advantage, not a compounding one. The clock is running.

The Competitive Landscape: A Field of Wreckage

Chain Vendor Status What happened
McDonald’s IBM (via Apprente) Killed The essential comparison. Running in 100+ restaurants since a 2021 partnership. A June 2024 franchisee memo confirmed shutdown in all test restaurants no later than July 26, 2024. Accent and dialect failures were widely reported; the viral failure videos were brutal. The largest QSR operator on earth, with effectively unlimited resources, could not make this work.
Taco Bell / Yum! Nvidia (2025) → Omilia Zig-zag, now the largest Deployed with Nvidia in early 2025, reaching a reported ~500 restaurants before Yum slowed the rollout in August 2025 after viral trolling (the “18,000 cups of water” episode). Yum then pivoted. A July 2026 announcement reportedly places Taco Bell voice AI in 890+ U.S. restaurants across 38 states with Omilia. If accurate, that would make Taco Bell — not Wendy’s — the industry’s largest deployment by unit count. Single-source and recent; verify before relying on it.
White Castle SoundHound (“Julia”) Durable The other genuine success. SoundHound reports ~90% order completion and ~60-second orders, and describes Julia as complete end-to-end AI rather than human-assisted. White Castle notes it handles combinatorial complexity (72 ways to order the #1 combo).
CKE (Carl’s Jr., Hardee’s) Valyant AI → ConverseNow Consolidated ConverseNow acquired Valyant in 2024. Notably used on-premises/edge hardware — a different architectural bet than Wendy’s.
Checkers/Rally’s, Del Taco Presto Automation Collapsed SEC AI-washing enforcement action, January 2025. The cautionary tale that redefined how these claims must be read.
Panera, Popeyes, Arby’s OpenCity (“Tori”) Pilots Limited public performance data.
Domino’s, Wingstop ConverseNow Deployed (phone-heavy) Cloud-based, with menu-management tooling for LTOs and out-of-stock handling.

What the wreckage teaches

Accent and dialect handling. This killed McDonald’s. It is the hardest technical problem in the stack and a fairness problem as much as an engineering one — a system that works well for some speakers and poorly for others is discriminatory in effect. Wendy’s added Spanish-language capability, which is a genuine step. Whether FreshAI’s accuracy is uniform across regional American accents is not publicly documented, and it should be.

Adversarial customers. Taco Bell demonstrated that a public voice AI wired to a real order system is an attack surface. Guardrails are the difference between a working system and a viral humiliation. Wendy’s built them from day one. That looks prescient.

Dishonest measurement. Presto did not fail because the technology was impossible. It failed because it misrepresented how much of the work humans were doing, and the SEC noticed. Technology risk and disclosure risk are separate risks — and the second one is the one that ends companies.

The Four Value Pools — and Why Only One Is Measurable

FreshAI can theoretically create restaurant-level value through four channels. The problem is that none of them has been isolated in any public disclosure.

  1. Labor productivity. Not necessarily headcount elimination — more often redeployment from headset to production. But redeployment is not automatically a financial saving. It only counts if it reduces paid labor hours or produces measurable incremental output. With a 14% intervention rate, someone must stay near the drive-thru regardless.
  2. Order accuracy and waste reduction. Fewer remakes, refunds, and complaints — but only if the AI beats the human baseline under real conditions, which Wendy’s has never demonstrated with matched controls.
  3. Throughput and capacity. Distinct from labor savings. A restaurant with identical staffing can be materially more profitable if it processes more cars at peak. The 22-second figure gestures at this but is not the same as measured cars-per-hour gain.
  4. Average check. Consistent suggestive selling. But check lift is not profit lift. The correct measure is incremental contribution profit after food, packaging, discounts, royalties, and any throughput penalty from over-upselling. An AI that raises ticket but slows the line can reduce total restaurant profit.

There is also a fifth channel that almost nobody models, and it may be the most interesting: demand. A structurally bilingual drive-thru lane is an accessibility guarantee no staffing model can match — every order, every shift, in the customer’s first language. Wendy’s heaviest state footprints overlap substantially with the highest Hispanic-population states. That is a traffic argument, not a cost argument, and cost levers cap out while demand levers compound. We develop it fully in FreshAI’s Hidden Demographic Edge. Wendy’s has never quantified it, never marketed it, and the market prices it at zero.

Failure Modes, Criticism, and Real Customer Experience

The field observation

In this author’s own repeated use of FreshAI across ten separate drive-thru orders, the experience was consistently strong — fast, accurate, and natural enough that the technology receded into the background, which is the only real test. We label this exactly what it is: a qualitative field observation, not a statistically significant study. Ten orders prove nothing about system-wide accuracy, labor productivity, or franchisee ROI. But they do rebut the lazy assumption that FreshAI is a gimmick that annoys customers. Many restaurant technologies demo well and feel terrible in real use. This one does not.

Documented friction

  • Premature cut-off. The most consistent public complaint: the system interrupting customers who pause while deciding. A barge-in and endpointing tuning problem, and a genuinely hard tradeoff — too patient feels sluggish, too aggressive talks over people.
  • Complex customizations remain a weak point, which is ironic given that combinatorial complexity was the stated reason to use an LLM.
  • Social friction. Some customers report uncertainty about how to speak to the system. A real adoption cost that appears in no accuracy metric.
  • No-fallback abandonment. Reports of customers driving away after errors when no human intervened. Every one is a lost transaction the “99% success rate” would likely never capture.

The biometric privacy exposure

This is the section institutional investors should read twice.

The Illinois Biometric Information Privacy Act (740 ILCS 14) classifies a voiceprint as a biometric identifier. BIPA is unusually dangerous for three reasons:

  1. Private right of action. Individuals sue directly; no regulator need act.
  2. Statutory damages per violation, per person: $1,000 negligent, $5,000 intentional or reckless.
  3. Five-year limitations period.

Apply that to a drive-thru. A restaurant serving several hundred cars a day, over a multi-year deployment, across dozens of Illinois locations, generates an enormous class. The exposure is not linear — it is combinatorial.

This is not hypothetical. In Carpenter v. McDonald’s Corp. (N.D. Ill., filed April 2021), the plaintiff alleged McDonald’s drive-thru AI extracted voiceprints analyzing pitch, volume, age, gender, accent, and national origin. In January 2022 the court allowed the core Section 15(b) claim — failure to obtain informed written consent — to proceed. The broader environment has reportedly escalated further: a November 2025 ruling is reported to have certified an Illinois voiceprint class against Amazon numbering roughly 1.2 million members. We have not independently verified that ruling; the exposure argument here rests on Carpenter and the statute itself, both of which are directly verifiable.

What we could not find: any Wendy’s disclosure on FreshAI voice data. Specifically:

  • Does FreshAI create, store, or derive a voiceprint — or is audio transcribed and discarded?
  • What consent signage exists at Illinois FreshAI drive-thrus?
  • What is the audio and transcript retention policy?
  • Is there a written biometric retention and destruction schedule, as BIPA requires?

These have clean, verifiable answers. Wendy’s simply has not published them. If the answer is “no voiceprint is ever created; audio is transcribed and discarded,” the exposure is materially lower and the company should say so loudly. The silence is the problem. An asymmetric legal risk with a cheap disclosure remedy, left undisclosed, is exactly the kind of thing that surprises shareholders.

The Bull Case

  • Survivorship is signal. McDonald’s failed. Presto committed securities fraud. Yum zig-zagged. Wendy’s built a system in 2023 that still runs in 2026 and remains Google’s flagship proof point. That is real execution competence.
  • The customer experience is genuinely good — supported by our own repeated field use and by the conspicuous absence of the viral failure tape that destroyed McDonald’s. Sustained absence of fail-content is itself evidence of engineering quality.
  • The head start is real. Menu grounding, POS integration, crew training, and failure discovery are a year-plus of work no competitor can skip by writing Google a check.
  • Drive-thru throughput economics are structurally favorable. Real order-time compression creates capacity at zero incremental real estate cost — the highest-return improvement available to a QSR operator.
  • Guardrails were built in from day one, which is why Wendy’s avoided the trolling humiliations that forced Taco Bell to slow down.
  • Unpriced demand optionality: a structurally bilingual lane network overlapping the youngest, fastest-growing U.S. consumer cohort. Unmeasured, unmarketed, and priced at zero.
  • The pricing problem is fixable. Tiered or performance-based pricing is entirely within Wendy’s control and would likely unlock franchise adoption immediately. This is the cheapest available lever on the entire program.

The Bear Case and Material Risks

  • The fee consumes the value. At $20,004 — ~95 bps of average franchised sales — FreshAI must clear roughly 100 bps just to break even. If real benefit lands near 100 bps, the fee captures essentially all of it and the franchisee nets close to nothing. The worst outcome is successful technology paired with unsuccessful economics.
  • The flat fee is regressive and likely value-destroying at low-volume restaurants — precisely the ones Wendy’s most needs to strengthen.
  • The 500–600 target appears missed, and it was arithmetically unreachable without franchise adoption that Wendy’s has never quantified.
  • No technology moat. Google owns the stack and is actively selling it to competitors. Wendy’s financed the proof of concept for a product now sold against it.
  • Headline metrics are stale, pilot-scale, and definitionally soft — and post-Presto, soft definitions carry regulatory as well as analytical risk.
  • FreshAI cannot fix the actual problem. U.S. comps fell 7.8% in Q1 2026; company-operated margin compressed from 14.8% to 11.4%. A drive-thru efficiency tool does not solve a traffic and value-perception crisis. Technology cannot replace demand.
  • Franchise relationship risk. Making FreshAI mandatory on new builds — which the FDD contemplates — before proving operator returns would convert a competitive advantage into another imposed cost during a period of required capital investment.
  • Leadership churn. Three CEOs in roughly twelve months (Tanner → Cook → Wright) plus a new CFO/CSO. Multi-year technology programs do not thrive under that turnover.
  • Undisclosed biometric privacy exposure.
  • Cloud dependency creates a network single point of failure for the highest-revenue channel, with no published failover architecture.

The BuyWendys Conclusion

FreshAI is a well-executed deployment of someone else’s technology, correctly aimed at a real constraint, and then priced in a way that guaranteed it would stall.

Wendy’s got the engineering right. It built guardrails before it needed them. It grounded the model in real menu data instead of shipping a generic assistant. It piloted in company stores where it could absorb failure. It disclosed — honestly, if unhelpfully — that its 99% success rate included human-assisted orders, at the exact moment a competitor was hiding that same fact from the SEC. And in practical use, it delivers a genuinely good customer experience, which is more than McDonald’s, Taco Bell, or Presto could ever claim. Judged as engineering, FreshAI is the best drive-thru AI program in the industry, and it is not close.

Judged as commercial design, it is a self-inflicted wound.

Wendy’s took a working system and attached a flat $20,004 annual fee to it — a fee equal to roughly 95 basis points of average franchised sales, and a punishing 1.33% at a $1.5 million restaurant. Then it offered that deal to franchisees during a period of falling traffic, without ever publishing the controlled unit-level data that would prove the benefit exceeds the cost. Franchisees did the arithmetic and largely declined. The 500–600 target, which was mathematically unreachable without them, went unmet. And the company, distracted by three CEO transitions and a genuine turnaround crisis, stopped talking about it.

FreshAI is not failing. It is mispriced. And mispricing is the most fixable problem on this list.

The strategic read is that FreshAI is not failing. It is mispriced. And mispricing is the most fixable problem on this list — cheaper than new engineering, faster than a new model, and entirely within the control of a management team that has not yet turned its attention to it. Tiered pricing by restaurant volume, a performance-linked structure, or a period in which early adopters retain most of the value they create would likely unlock adoption within a quarter.

That is the opportunity hiding in this story, and it is why FreshAI deserves more scrutiny than the “is AI good or bad” debate has given it.

For investors, the operating framework:

  • Do not underwrite a technology premium. There is no defensible IP. Wendy’s is not an AI stock; it is a restaurant company using AI to improve operations. That is a better and more credible frame, and it caps the multiple.
  • Model the fee, not the hype. $20,004 per restaurant, ~95 bps of average sales, roughly 100 bps of benefit required just to break even. That is the equation that determines whether FreshAI ever scales.
  • Discount the 86% and 99% figures until Wendy’s publishes a current, system-wide, franchise-inclusive zero-human-involvement rate. Not a “success rate.” An intervention rate.
  • Watch franchisee renewal above all else. Mandatory adoption creates deployment scale. Voluntary renewal proves economic value. It is the single cleanest signal available.
  • Price the BIPA exposure as a real tail risk until Wendy’s discloses its voice data practices.
  • Judge Wendy’s on Project Fresh, not FreshAI. The turnaround is the thesis. The AI is useful optionality inside it — no more, and no less.

What Would Change Our View

Trigger Direction Why it matters
Wendy’s introduces tiered or performance-based FreshAI pricing Strongly positive Directly addresses the regressive-fee problem. The cheapest, fastest available unlock for franchise adoption.
Disclosure of ≥500 live FreshAI restaurants with a franchise/company split Positive Proves franchisees are voluntarily paying $20,004 — the only real evidence the unit economics work.
Strong voluntary franchisee renewal after a full contract period Strongly positive The cleanest possible proof of economic value. Operators do not renew a fee that destroys their margin.
A current, system-wide zero-intervention rate above 90% Strongly positive First genuine evidence of a labor-structure change rather than labor reallocation.
Matched-control study isolating FreshAI’s restaurant-margin contribution Positive Converts a hand-wave into a modelable line. Would establish whether benefit clears the ~100 bps break-even.
Cohort disclosure of Spanish-language order mix or bilingual-lane comps Positive Would reclassify FreshAI from a COGS lever to a demand driver — and growth-column technology re-rates restaurant stocks.
Clear statement that no voiceprint is created and audio is discarded Positive Largely defuses the BIPA tail risk at near-zero cost.
FreshAI made mandatory on new builds before ROI is proven Negative Converts a competitive advantage into an imposed franchisee cost. Franchise-relationship risk.
Continued silence on unit counts and economics through FY2026 Negative A company with good numbers publishes them.
Major competitors licensing the same Google Food Ordering AI Agent Negative Compresses the head start; confirms the absence of a moat.
A BIPA class action filed against Wendy’s over FreshAI Strongly negative Converts a tail risk into a balance-sheet event.

Gaps in the Public Record

What we could not verify, and what verification would require:

  • No architecture diagram exists. The teardown above is a reconstruction. Verification would require a Wendy’s or Google engineering disclosure or a published reference architecture.
  • No FreshAI patents found from The Wendy’s Company or Quality Is Our Recipe, LLC. This supports the “no proprietary core IP” conclusion, though absence of patents is not absence of trade secrets.
  • Model variant, ASR engine, TTS engine, and orchestration SKU are not confirmed for Wendy’s. Any article naming a specific model variant is inferring. So are we, and we label it.
  • The POS vendor is not disclosed.
  • Wendy’s net economics on the $20,004 fee are unknown. Gross billing is not profit. Google Cloud costs, usage-based compute, implementation, support, QA, menu configuration, and security must all be deducted. Whether FreshAI is a high-margin software stream or a near-pass-through service is undetermined.
  • No FreshAI quantification appears in any SEC filing. Every scale and performance figure in circulation is company communication, partner marketing, or trade press.
  • Current unit count is unknown. Third-party counts conflict; digital menu board installs are routinely conflated with FreshAI installs. Treat all unit counts as unreliable.
  • Google’s marketing figures vary by capture date (24 states; 50,000/day; 60,000/day; 95% success) and use definitions distinct from Wendy’s own. Vendor marketing is not company disclosure.
  • No cohort-level demand data exists — no Spanish-order mix, no FreshAI-lane versus control-lane comps.

Frequently Asked Questions

What is Wendy’s FreshAI?

FreshAI is Wendy’s conversational AI system for taking drive-thru orders by voice, announced May 9, 2023 and built on Google Cloud’s foundational large language models. It first piloted in June 2023 at a company-operated restaurant near Columbus, Ohio. Internally, Wendy’s calls the assistant “Wendy.” The Franchise Disclosure Document describes it as a third-party food-ordering AI agent combined with Wendy’s proprietary menu logic and recommendation engine.

How much does FreshAI cost a Wendy’s franchisee?

$1,667 per restaurant per month — $20,004 per year, per the 2025 Franchise Disclosure Document, invoiced quarterly in arrears and charged on top of the existing technology fee. Against the FDD’s $2,108,454 average franchised restaurant sales, that equals about 0.95% of sales. FreshAI must generate roughly 95 to 100 basis points of incremental restaurant profit for the average franchisee simply to break even.

Does Wendy’s FreshAI actually work?

Better than its competitors, yes. In its 2023 pilot Wendy’s reported 86% of orders completed without employee intervention and service 22 seconds faster than the local market average. It has survived while McDonald’s shut down its IBM system and Presto Automation collapsed under an SEC enforcement action. Our own repeated field use found the customer experience consistently strong — though we treat that as a qualitative observation, not a study. The caution: Wendy’s pilot metrics come from four restaurants and have not been updated since December 2023.

How many Wendy’s locations have FreshAI?

Wendy’s has not disclosed a current figure, and it has never quantified FreshAI in an SEC filing. In February 2025 management cited roughly 100 locations and targeted 500–600 by year-end 2025. That target was arithmetically unreachable without franchise adoption, since Wendy’s operates only 423 company-owned U.S. restaurants out of 5,969. Treat any specific number you encounter as unverified.

Why did the FreshAI rollout stall?

Most likely because of pricing, not technology. The flat $20,004 annual fee is regressive — costing a $1.5 million restaurant 1.33% of sales versus 0.67% at a $3 million restaurant — and Wendy’s asked franchisees to absorb it while U.S. same-restaurant sales were falling 7.8%. Without published controlled data proving the benefit exceeds the cost, a rational operator waits. This is a fixable pricing-design problem.

Is FreshAI replacing Wendy’s employees?

There is no evidence it has eliminated positions. Because roughly 14% of orders required human intervention even under favorable pilot conditions, an employee must remain available to the drive-thru. The realistic effect is labor reallocation — freeing crew for food preparation and service — rather than headcount reduction. Reallocation is only a financial saving if it reduces paid hours or produces measurable incremental output, and Wendy’s has never published a labor-hours impact figure.

Why did McDonald’s cancel its AI drive-thru and Wendy’s did not?

McDonald’s ran an IBM-powered automated order taker in 100+ restaurants beginning in 2021 and confirmed in a June 2024 franchisee memo that it would be shut off in all test restaurants by July 26, 2024. Accent and dialect failures were the most reported problem. Wendy’s launched two years later on a materially more capable generation of language models, and built menu grounding and conversation guardrails in from the start. Timing and architecture both favored Wendy’s.

Does FreshAI record my voice?

Wendy’s has not publicly disclosed its FreshAI voice data practices. We found no statement addressing whether voiceprints are created, what consent signage exists, or what the retention policy is. This is a genuine gap. Illinois law treats voiceprints as biometric identifiers with statutory damages of $1,000 to $5,000 per violation, and a similar drive-thru case against McDonald’s survived a motion to dismiss on its core consent claim.

Does FreshAI give Wendy’s a competitive advantage?

A temporary one, not a durable one. The language model, speech recognition, speech synthesis, and dialogue orchestration are all Google’s, and Google has packaged that capability as a “Food Ordering AI Agent” any competitor can license — using Wendy’s as the reference customer. Wendy’s real advantage is two to three years of operational learning: menu grounding, POS integration, crew training, failure discovery. That is a decaying head start, not a moat.

Will FreshAI turn around Wendy’s stock?

No, and management appears to agree. Wendy’s reported U.S. same-restaurant sales down 7.8% in Q1 2026 and company-operated margin compression from 14.8% to 11.4%. FreshAI is a drive-thru efficiency tool; it cannot solve a traffic and value-perception problem. Technology cannot replace demand. The turnaround thesis rests on Project Fresh — value, menu, and footprint rationalization. FreshAI is useful optionality inside that thesis, not the thesis itself.

What language does FreshAI speak?

English and Spanish. Wendy’s announced Spanish-language ordering in its May 2025 FreshAI update; customers can switch by saying “Spanish” or “Español” at the start of the order. This may be the most undervalued feature of the entire program — a structurally bilingual lane is an accessibility guarantee no staffing model can match, and Wendy’s largest state footprints overlap heavily with the highest Hispanic-population states.

Sources

Primary — Franchise Disclosure Document

  • Quality Is Our Recipe, LLC, 2025 Franchise Disclosure Document, issued March 28, 2025: Items 6, 11, and 19 and Exhibit T. Source for the FreshAI fee ($1,667/month), the charge stack, restaurant counts, and average and median franchised restaurant sales.

Primary — Wendy’s corporate and SEC

Primary — Google

Regulatory and legal

  • U.S. Securities and Exchange Commission, administrative order regarding Presto Automation, January 14, 2025 (File No. 3-22413).
  • Carpenter v. McDonald’s Corp., No. 1:21-cv-02906 (N.D. Ill.), filed April 26, 2021; motion-to-dismiss ruling January 13, 2022.
  • Illinois Biometric Information Privacy Act, 740 ILCS 14.

Trade and business press

Related BuyWendys.com research

Methodology: all calculations are BuyWendys.com estimates derived from disclosed figures. Systemwide illustrations assume FDD average sales and should not be read as company forecasts. Company-operated restaurant economics do not represent actual franchisee profits, which vary by labor, occupancy, management, geography, financing, and volume.

Author and Disclosure

Ben Edmond is the founder of BuyWendys.com, an independent research publication covering The Wendy’s Company (NASDAQ: WEN), Wendy’s restaurants, franchise economics, AI initiatives, dividends, valuation, short interest, and store-level location intelligence. Learn more about Ben Edmond →

Position disclosure. Assume the author and/or operators of BuyWendys.com hold a long position in The Wendy’s Company (NASDAQ: WEN) and stand to benefit if the share price rises. We may buy or sell at any time without notice. Weigh everything in this article with that conflict of interest in mind.

Field observation disclosure. The author has personally used Wendy’s FreshAI across ten separate drive-thru orders and found the customer experience strong. This is qualitative field evidence, not a statistically significant performance study, and it is labeled as such wherever it appears. It does not establish system-wide accuracy, labor productivity, or franchisee return on investment.

Not investment advice. BuyWendys.com publishes independent opinion and commentary for informational purposes only. Nothing here is investment, financial, trading, tax, legal, or franchise advice, and nothing here is a recommendation, solicitation, or offer to buy or sell any security. We are not a registered investment adviser or broker-dealer. All investing involves risk, including total loss of capital.

Not affiliated with The Wendy’s Company. BuyWendys.com is independent and is not affiliated with, endorsed by, or sponsored by The Wendy’s Company or Google. “Wendy’s,” “FreshAI,” and related marks belong to their owners and are referenced here for commentary and identification.

Accuracy. This article distinguishes throughout between verified fact, management guidance, partner marketing, third-party estimate, and BuyWendys inference. Where the public record is silent, we say so rather than filling the gap with assumption. Verify all figures against primary sources before acting on anything here. See our Disclaimer.