The Voice Agent Maturity Curve

I had two operator conversations the same week last month. They argued opposite sides of the same question without realising they were arguing about different products.

The first was a multi-unit casual operator who had lost his front-of-house manager and was watching his hosts miss roughly one call in three during Friday and Saturday service. He wanted a voice agent deployed by month-end. His logic was simple: a robot picking up the phone was better than a phone ringing into a void.

The second was the owner of a single-room tasting-menu concept who had been pitched a voice agent at a conference and spent the flight home composing a furious email — my guests call my restaurant to talk to my hostess, that is the entire point. He wanted me to validate his refusal.

I told both of them they were partially right. The first was right about the call-screening problem; he was wrong about the product he thought would solve it. The second was right about the brand cost at his concept; he was wrong about what the category contains. They were using the same two words — voice agent — to mean five entirely different products at five entirely different stages of maturity. The vendor marketing collapses the five into one. Most operator confusion in this category lives in that collapse.

This is the second Mise framework essay. The first, the Four Margins of a Restaurant, is the lens this one sits inside. Read it first if you haven’t.

The thesis

Voice AI in hospitality is at five distinct stages of maturity. Each stage anchors to real products in market today. Each stage carries a different operator decision. Confusing the stages — buying at one and deploying at another — is the most common and most expensive mistake in this category.

I call the framework the Voice Agent Maturity Curve because the products move along it. A vendor at Stage 2 today is genuinely trying to be at Stage 4 within eighteen months. Operator decisions should be made against today’s stage, not next year’s marketing slide.

The curve, in one diagram:

Stage 1 ──── Stage 2 ──── Stage 3 ──── Stage 4 ──── Stage 5
 spam        reservation   concierge    brand        operating
 shield      taker                      voice        co-pilot
 [cost]      [revenue]     [revenue]    [brand]      [infra]

The bracketed label below each stage is the operator margin the stage moves. Stage 1 takes labor cost out. Stages 2 and 3 capture revenue that was leaking. Stage 4 reshapes the brand-perception margin. Stage 5 — still ahead of us — rebuilds the back-office labor stack.

Five stages, five decisions. In order.

Stage 1: Spam shield

What it is. The agent answers the phone, screens robocalls, takes a message for callers who clearly need a human, and routes nothing to the host stand the host stand could not have handled in five seconds. Barely a conversation. Closer to an intelligent answering machine than to a hostess.

Real products. Yelp Receptionist, launched October 21, 2025 alongside the more capable Yelp Host, sits at this stage for non-restaurant categories; for restaurants the spam-shield piece is bundled into Yelp Host. PolyAI includes spam identification in its baseline call-handling stack. Horizontal voice-AI vendors will sell a spam-shield package as a gateway product.

Who should buy. Restaurants that pay a human to answer a phone full of robocalls. Spam is typically 20–40% of inbound; eliminating it returns measurable host attention to the floor. Cost of getting Stage 1 wrong is essentially zero — the alternative was voicemail.

Who shouldn’t buy. Owner-operated independents who pick up between courses. Operators whose phone only rings at the start of the pushes. Stage 1 earns its place on call volume, not call value.

Stage 2: Reservation taker

What it is. The agent handles inbound reservation requests during the service hours when the dining room can’t keep up. Scripted, narrow, confident inside its lane. Books, confirms, makes simple modifications, routes everything else to a human. No improvisation; no personality beyond the voice you picked from a dropdown.

Real products. Yelp Host, launched October 21, 2025 at $149/month ($99 for Guest Manager customers), handles reservations, takeout, and FAQs in eight languages (English, Spanish, French, German, Italian, Portuguese, Dutch, Turkish). Slang AI, which raised a $36M Series B in February 2026 led by USVP and has built a proprietary dataset from 25 million guest interactions, is the strongest pure-play vendor at this stage. SoundHound for Restaurants, via its Toast partnership, sits between Stage 2 and Stage 3 by handling voice ordering as well as reservations.

Who should buy. Mid-volume casual and mid-range concepts where the host stand is genuinely missing calls. Multi-unit operators with tight host-coverage economics. Properties above roughly 30 calls a day. Yelp’s $149 price was set carefully: materially below the fully-loaded cost of one host-hour per day.

Who shouldn’t buy. Fine-dining concepts where the phone is part of the brand. Properties whose call volume can’t clear the contract overhead. Operators whose existing host is sitting at the stand with nothing to do — Stage 2 converts missed calls into captured bookings; you have to be missing calls.

Stage 3: Concierge

What it is. Multi-turn conversation. The agent handles modifications, FAQs, wait times, takeout orders with substitutions, and the long tail of can you do gluten-free? and what’s your corkage? Integrates with the reservation system, the POS, and increasingly the loyalty CRM. Not scripted — following a graph of intents and routing to the right system on each turn.

Real products. The VOICEplug AI × OpenTable global integration, announced November 24, 2025 across 20 countries (US, Canada, UK, Australia, Mexico, France, Germany, Italy, Spain, Netherlands, Ireland, Portugal, Singapore, Japan, Hong Kong, Taiwan, Thailand, UAE, Saudi Arabia, Brazil), handles reservations, group bookings, private-booking inquiries, cancellations, and waitlists in conversational voice, syncing availability to OpenTable in real time. PolyAI’s restaurant deployment, also distributed via OpenTable (support docs here), sits at Stage 3 for enterprise customers. Toast IQ’s conversational AI assistant, announced October 29, 2025 across roughly 148,000 customer locations, is an internal-facing concierge for the operator rather than the guest — same engineering family.

Who should buy. Multi-unit groups, hotel F&B, and high-volume independents whose calls are no longer all reservations. If your phone calls average longer than two minutes, your operation has crossed into Stage 3 territory whether you deployed the agent or not.

Who shouldn’t buy. Operators whose front-of-house standards depend on the agent handling things it can’t credibly handle yet. The variance on quality is still meaningful, and a confidently-wrong concierge confirmation is the worst failure mode in the category. The fix is a 30-day diligence period; the trap is skipping it because the vendor pitch was good. (See our Toast AI desk review and our Yelp coverage for what diligence on a Stage 2/3 vendor actually looks like.)

Stage 4: Brand voice

What it is. The voice agent is the brand voice. Custom persona, custom tone, multilingual, integrated with the marketing stack and the guest CRM. A returning guest hears something that recognises her. A first-time caller hears the same voice from the property’s video ads. The agent stops being a utility and becomes part of the product.

Real products. PolyAI is the clearest current example. Fogo de Chão selected PolyAI to power “Selma,” a custom-persona agent answering every call across all 88 U.S. locations. PolyAI reports 100+ enterprise customers, 2,000+ live deployments across 45 languages and 25+ countries, including Marriott and Caesars outside hospitality. Newo.ai, which raised a $25M Series A with 15,000+ agents reportedly in use, sits between Stage 3 and Stage 4 for SMB operators who want a custom-persona host without an enterprise contract.

Who should buy. Multi-unit groups whose brand is part of their pricing power. Hotels and clubs where the voice on the phone is the first impression. Concepts where the marketing budget is large enough that the voice agent is a brand asset, not an expense line. Fogo de Chão is the canonical example — a brand investing in service consistency across geography, choosing a vendor whose product is precisely that consistency.

Who shouldn’t buy. Single-unit operators whose brand voice is, in fact, their hostess. Any agent in front of those guests dilutes the brand they are paying to protect. The fine-dining concept I opened with belongs here; he was right to refuse, and should keep refusing for at least two more years.

Stage 5: Operating co-pilot

What it is. (Forward-looking. Prediction, not measurement.) The voice layer becomes infrastructure. The agent handles vendor calls, supplier outreach, scheduling, internal coordination across units, and the back-office labor nobody enjoys doing. The guest-facing voice agent and the operations-facing voice agent are two surfaces of the same platform. Voice stops being a channel and starts being a substrate.

Real products (today, in adjacent categories). Bland, which raised a $40M Series B in January 2025 and supports up to 1M concurrent calls with sub-400ms latency, is the clearest infrastructure-tier outbound platform — operators in adjacent categories are already using it for supplier outreach and appointment confirmation. Vapi, by one industry tracker’s count, processes 62M monthly calls as an orchestration layer. Retell positions itself as the production-grade receptionist for service businesses including dental, real estate, and increasingly hospitality. Toast’s Drive-Thru product, launched April 2026 with voice-ordering through partners including Incept AI, is the QSR-side proof-point that voice is becoming the throughput layer, not a feature layered on top. See our McDonald’s drive-thru desk review for the cautionary tale.

Who should (start preparing to) buy. Multi-unit groups where back-office labor is a meaningful operating-margin drag — chain coordination, supplier confirmations, vendor invoicing follow-up, scheduling. The math becomes obvious north of roughly 30 units. Single-unit operators don’t have the volume to justify Stage 5 today and won’t for the rest of this decade.

Who shouldn’t buy. Anyone who reads the previous paragraph and concludes Stage 5 means replace your manager with a phone bot. It does not. Stage 5 gives your manager a phone-call agent the way Stage 4 gives your host stand one. The manager remains; the labor profile shifts toward judgment and away from coordination. Prediction: credibly deployable for multi-unit operators in 2028–2029. Single-unit operators get it bundled into their POS by 2031.

How to figure out which stage you should buy at

Four heuristics. They compose; you check all four, and the lowest stage any heuristic produces is your ceiling.

Heuristic one: call volume. Under 8 calls a day, stay at Stage 1 until volume changes. Between 8 and 30, Stage 2 is your ceiling. Above 30, Stage 2 is obvious and Stage 3 is reasonable. Above roughly 100 a day across a multi-unit footprint, Stage 3 is obvious and Stage 4 is worth a pilot.

Heuristic two: brand sensitivity to AI failure. The most consistently underweighted factor. If your guests notice they’re talking to a machine, your deployable stage is one stage lower than your call volume suggests. Fine dining caps at Stage 1 today, full stop. The asymmetry — one bad interaction tells five people — doesn’t pencil at higher stages until technology variance comes down meaningfully.

Heuristic three: operating-margin pressure. The most consistently overweighted factor. Voice agents save real money at Stages 2 and 3, brand money at Stage 4, and infrastructure money at Stage 5. They do not save much money at Stage 1 — that’s a customer-experience and spam-defense play. Operators who buy Stage 1 expecting labor savings are disappointed and then evangelise against the category. Don’t.

Heuristic four: data infrastructure. Stage 4 requires the guest-CRM data to make recognition useful. If your reservation system and guest history aren’t joined, you can’t deploy Stage 4 even if the vendor swears the integration is one-click. The OpenTable AI strategy (covered in our desk review) is essentially a long bet that OpenTable’s guest-history data is the moat for Stage 4 economics. They may be right.

The fifteen-year arc

Where this is going. All predictions, flagged as such.

By end of 2026, Stage 2 will be table-stakes across multi-unit casual operators. The Yelp/Slang/SoundHound/VOICEplug competitive set forces adoption through price compression alone.

By 2028, Stage 3 will be widely deployed. Voice-agent integration with reservation/POS systems will deepen; operators who built CRM and guest-data infrastructure in 2026–27 will benefit, and operators who skipped that investment will pay premium prices for integration catch-up.

By 2029–2030, Stage 4 is technologically available to most operators and economically viable for those whose brand justifies the cost. The winners are a smaller set than Stage 2’s winners. The category bifurcates into operators who use the agent as a utility (Stages 1–3) and operators who use the agent as a brand asset (Stage 4).

By 2032–2035, Stage 5 is the substrate of multi-unit operations. The back-office labor profile of running a chain restaurant looks different in 2035 than in 2026, in roughly the same way front-of-house labor changed when handheld POS terminals went mainstream in the 2010s. Voice becomes the connective tissue, not a customer-facing feature.

Predictions about fifteen-year arcs are wrong in specifics and — if the framework is good — right in direction. I’ll re-stamp this essay each year with the new anchor products and the corrected timeline. The five-stage structure I expect to survive. The vendors named here I expect to mostly survive; one or two won’t. Voice is a real platform shift. The operator decision is which stage you belong at today, given who you are today. Not next year’s marketing slide. Not your competitor’s headline deployment.

The operator who wanted a voice agent yesterday belonged at Stage 2 and bought it that month. The operator who refused belonged at Stage 1 only for the next two years; he kept his hostess. Both deployed correctly against their own stage. That is the entire point of the framework.

— Eitan writes Mise. The Voice Agent Maturity Curve will be re-stamped each year as the products move along it. Tips: [email protected].