Voice AI That Answers 92% of Calls and Adds $680K a Year

The Company

Saffron Lane Hospitality operates 34 casual-dining locations across the Southwest, a group built on generous plates, loyal regulars and a takeout business that has quietly grown to 28% of revenue. Its restaurants sit in the unforgiving middle of the market: not counter-service enough to skip the phone, not fine-dining enough to staff a dedicated phone line, and busy enough that the phone rings hardest precisely when the dining room is fullest. The group's economics depend on two clocks — the table turn and the takeout ticket — and both suffer when nobody can reach the host stand.

Labor is the other half of the story. Hourly restaurant labor has been scarce for years, and the National Restaurant Association has reported that most operators find recruiting and retaining staff harder than before the pandemic. Saffron Lane's managers were hiring constantly, training constantly and still covering phones with people they needed on the floor. Leadership wanted automation, but with a constraint that shaped everything that followed: whatever answered the phone had to feel natural to loyal guests who had been calling the same location for years, not robotic, and it had to make the operation better rather than merely cheaper.

Where Things Stood

Before Emerson Technology engaged, the phones rang off the hook exactly when the restaurants were busiest. During the dinner rush, host stands let calls ring out, and the group estimated 30% of inbound calls went unanswered — each abandoned takeout order worth about $34. Managers pulled floor staff to answer phones, which protected the takeout revenue at the direct expense of table service, while weekend hold times pushed callers to competitors who picked up faster. The group was, in effect, taxing its best hours twice: once in lost takeout orders and once in degraded dining-room experience.

Order taking was inconsistent in ways that compounded the staffing problem. Upsells happened only when a distracted employee remembered them, which meant the same caller could get offered a drink, a dessert or nothing depending on who answered and how chaotic the moment was. Phone orders were keyed into the point-of-sale system afterward, producing entry errors that surfaced as wrong food, refunds and apologetic phone calls of the very kind nobody had time to answer. Industry data showed the group was not unusual — annual drive-thru studies such as Intouch Insight's QSR report place average order accuracy around 85-87%, demonstrating how much routine order-taking still fails even in chains built entirely around it.

The math leadership finally confronted was simple. A 30% abandoned-call rate against a $34 average ticket, multiplied across 34 locations at peak hours every week, was a seven-figure annual leak — and the labor spent failing to answer those calls was itself a cost. The question was not whether to fix the phones; it was whether a fix could be trusted with the group's guests.

The Challenge

The challenge was to automate the highest-volume, lowest-judgment call flows without damaging the guest relationships the brand depends on, and to do it in environments — noisy dining rooms, heavy accents, menus that change daily — that break naive voice systems. The technical bar was specific: a voice agent that can take a full order conversationally, handle modifications, honor 86'd items, and write the result directly into the POS without human re-keying. Anything less simply relocated the error-prone manual step from a person to a process.

The design bar was just as high. Guests forgive a lot, but they do not forgive a dead end — an automated system that cannot route a catering inquiry or a complaint to a human, with context attached, creates exactly the experience that turns callers against automation permanently. The group also wanted the system's discipline to be an improvement, not just its availability: the same three upsells offered every time, orders entered once and entered right, and Spanish-speaking callers handled natively rather than transferred. Finally, leadership insisted on proof before chainwide rollout — the agent had to survive the group's loudest, hardest stores before it earned the easy ones.

Discovery and Diagnosis

Emerson Technology began with a two-week diagnostic across a sample of locations, counting calls at the host stand during peak hours and reconstructing what an unanswered call actually cost. The audit confirmed the estimate that roughly 30% of inbound calls went unanswered during the dinner rush and priced the leak at approximately $34 per abandoned takeout order — a number that immediately funded the project in leadership's eyes and set the success bar for everything after. The team also recorded and categorized call types, finding that the overwhelming majority were simple transactions: order ahead, reservation check, hours question. Only a small share — catering, large parties, complaints — genuinely required human judgment.

The technical diagnosis focused on where voice systems fail. The team reviewed the group's menus, hours and 86'd-item practices across all 34 locations and found enough variance that a centrally scripted assistant would be wrong somewhere every day — which settled the architecture decision that the agent had to be constrained to each location's live menu data rather than to a script. Wendy's and Google Cloud had publicly piloted generative-AI drive-thru ordering in 2023, part of a broader wave of voice automation across major quick-service brands, and the team studied the public lessons from that wave: latency matters, constraint matters, and the failure mode that kills adoption is a confident wrong answer. The diagnostic concluded with a pilot design: five locations, the loudest and highest-volume in the portfolio, eight weeks, with every transcript graded against kitchen tickets.

The Solution

Emerson Technology built a conversational voice agent on cloud telephony and an AWS-hosted LLM constrained to each location's live menu, hours and 86'd items. The architecture has three layers.

The constrained agent

The agent answers every call in under a second, takes orders conversationally and offers the same three upsells every time — consistency that no distracted human shift can match — and writes finished orders directly into the POS with no re-keying, eliminating the dual-entry step where most phone-order errors were born. Constraint is the core safety mechanism: the model is fed each location's live menu, prices, hours and 86'd items through an API, so an open-ended language model's tendency to invent items is designed away — a constrained agent can sell only what exists. The agent handles Spanish natively, which mattered both for guests and for the group's standing in its communities.

Latency and the guest experience

Response latency stayed below 800 milliseconds by streaming speech rather than turn-taking, tuning prompts aggressively and prefetching menu data before questions are fully asked — because pauses of more than a second read as rudeness on a phone call, and a pause followed by a wrong answer reads as worse. The team treated the guest experience as an engineering spec: natural pacing, confirmation of complex items, and graceful handling of interruption and background noise from a dining room at full volume. Mystery-guest testing during the pilot measured the only perception metric that matters — whether callers realized they were not speaking with a person.

Human handoffs and the pilot discipline

Calls involving catering, large parties or complaints transfer to staff with a spoken summary attached, so the employee who picks up already knows the guest's name, party size or grievance — guests forgive a transfer, but they do not forgive starting over. The pilot ran for eight weeks at five loud, high-volume locations, deliberately the hardest in the portfolio, and the team reviewed every transcript against recorded kitchen orders before rollout, publishing weekly accuracy reports per store. That review cadence caught model drift, menu-data gaps and accent-handling weaknesses while they were cheap to fix — and it produced the per-store accuracy evidence that made the chainwide rollout a business decision rather than a leap of faith.

The First 90 Days

Weeks one and two were spent on the missed-call baseline and the menu data pipeline: peak-hour call counts at the pilot stores, average ticket verification, and each location's live menu, hours and 86'd logic connected through the API so the agent could never sell what the kitchen could not make. Weeks three through six deployed the agent at the five pilot locations in a supervised mode — taking real orders while store managers listened on a live monitor — with daily transcript reviews in the first two weeks, then weekly. The team graded transcripts against kitchen tickets and published accuracy by store every Friday, a discipline that surfaced and fixed a handful of failure modes early: a modifier phrasing the agent misheard, one store's 86'd list updating late in the morning, and a background-noise threshold that needed retuning at the loudest location.

Weeks seven through ten hardened the handoff rules. Catering, large parties and complaints were routed to staff with spoken summaries, and the team discovered during the pilot that the summary was doing as much work as the routing — staff who picked up knowing the context resolved those calls in half the time. Latency was tuned below 800 milliseconds across the pilot estate, and Spanish-language handling was validated with native-speaking staff at two locations. Weeks eleven and twelve extended coverage to the remaining pilot scenarios — reservation changes, hours questions — and assembled the rollout case: containment rates, order accuracy against kitchen tickets, mystery-guest scores and the revenue recovered from calls that previously would have rung out.

The transition rule for staff was set plainly by the group's operations lead: the agent handles transactions, humans handle judgment, and nobody is replaced — they are reassigned to the work that needs a person. Managers received a one-page playbook and a named contact during the first month, and the stores that adopted the fix-list mentality fastest were, tellingly, the busiest ones, because their managers had the most hours to reclaim. By day 90 the pilot stores were asking to expand the agent's scope, and the chainwide rollout was scheduled on the strength of evidence rather than enthusiasm.

The Results

A year after the chainwide rollout, the P&L and the phone log agreed. The voice agent now handles 92% of inbound calls without staff involvement, while abandoned calls fell from 30% to 4% — a recovery of demand that had been walking out the door every Friday night. Consistent upselling lifted average order value 9%, and recovered takeout volume added an estimated $680K in annual revenue. Managers reclaimed roughly 25 staff-hours per location each week for floor service, and mystery-guest scores show guests rarely realize they spoke with an AI.

The second-order effects compounded the first. Order accuracy improved because the manual re-keying step disappeared, and the refunds and apology calls that came with it largely did too. The consistency of the upsell — the same three offers, every order — turned out to be worth as much as the availability, and it exposed how much variance had been hiding in human order-taking. Managers describe the dinner rush as quieter in the ways that matter: floor staff stay on the floor, and the phone is simply no longer a crisis. The group is now extending the same agent to reservation changes and catering quotes, reusing the containment discipline that made the first deployment measurable.

What Your Business Should Do

Saffron Lane's playbook is repeatable for any multi-unit operator, and it starts with arithmetic rather than technology. Quantify your missed-call revenue first: have someone count unanswered calls during peak hours for one week and multiply by average ticket — that number funds the project and sets the success bar. Pilot at your loudest locations, testing where background noise, accents and menu complexity are worst rather than easiest, because if the agent survives your hardest store, the chainwide rollout becomes comparatively straightforward. Constrain the AI to your menu, feeding the agent live menu, prices, hours and 86'd items through an API — open-ended language models invent items, while constrained ones sell only what exists. Write orders straight into the POS so they appear on kitchen screens without human re-keying, because manual dual entry reintroduces the very errors the automation was meant to remove.

Then engineer the parts that make callers trust the system. Define human handoff rules precisely, routing catering, complaints and large parties to staff with a spoken summary always attached — guests forgive a transfer, but they do not forgive a dead end. Tune for sub-second response latency by streaming audio, prefetching menu data and trimming prompts until replies feel immediate, because pauses of more than a second read as rudeness on a phone call. Review transcripts weekly by store, grading a sample against actual kitchen tickets and publishing accuracy by location, since that review catches model drift before guests ever notice it. Track containment, not call volume — measure the share of calls completed without staff help, plus revenue per answered call, because volume statistics flatter the system while containment and revenue prove the business case. The industry context is blunt: major brands including Wendy's have piloted generative-AI ordering with Google Cloud, drive-thru studies place routine order accuracy at only 85-87%, and the National Restaurant Association reports that most operators find staffing harder than before the pandemic — the phone is the single highest-leverage place to apply automation in a restaurant.

Before you choose a vendor, sit with the questions Saffron Lane's leadership had to answer honestly. How much takeout revenue walked away last Friday night because nobody could reach the phone? Which employees answer phones today, and what higher-value work would they do with those hours back? Do your best servers deliver the same upsells on every phone order, or only on their best days? Would your regulars forgive an AI order-taker if their food arrived faster and more accurately? What percentage of your inbound calls are simple transactions that a constrained system could finish end to end? If you piloted at your hardest location, what would you learn that a friendly store would never teach you? Use the action checklist in this practice book to work these through with your team, and revisit them each quarter as your menu, your call mix and the voice technology itself evolve.