Key takeaways
- Measure a two-week baseline of missed calls, after-hours contacts and staff phone time before switching anything on.
- Count value in four streams: recovered bookings, recovered takeout, catering leads and staff hours moved back to guests.
- Use contribution margin, not revenue, and subtract setup, review time and vendor fees.
- Only 20% of operators are running small AI pilots and 3.2% are implementing AI companywide[1], so a clean measurement plan is still rare.
Why most AI ROI claims do not survive a P&L review
Vendors tend to report activity: calls answered, chats handled, minutes saved. Owners need contribution: extra covers and orders that would not have happened, and labor hours that actually changed. The gap between the two is where most AI business cases fall apart.
The stakes are real. Food and labor each take roughly 33 cents of every sales dollar[2], and 33% of operators said their restaurant was not profitable in the first half of 2026[3]. A tool that adds work for managers or produces bookings that would have come anyway is a cost, not an investment. Operators know this: cost and business fit are the two most cited obstacles to adopting technology[4].
Step 1: take a baseline before you switch anything on
Two weeks is enough for most restaurants, as long as it covers a normal weekend and no holiday. Pull the data from your phone provider, POS and reservation book rather than estimating.
- Inbound calls by hour, answered versus missed or abandoned, including calls that reached voicemail.
- Voicemails left and how many were returned within the same service.
- Website chat or contact form messages, and how long they waited for a reply.
- Phone reservations and phone takeout orders, from your book and POS order source fields.
- Catering and event inquiries, with the date each one was first answered.
- Staff time on the phone: ask hosts and counter staff to tally calls for three busy shifts, or sample call duration from the phone report.
Write the baseline down with dates and sources. If you cannot reproduce a number later, you cannot use it to judge the tool.
Step 2: count value in four streams
| Value stream | What to count | Where the data lives | Common trap |
|---|---|---|---|
| Recovered reservations | Bookings from calls or chats that previously went unanswered (after hours, rush overflow) | Call log plus reservation source notes | Counting every AI-handled booking, including ones staff would have answered |
| Recovered takeout | Orders placed after the assistant answered or texted the ordering link | Online ordering reports with a tagged link | Ignoring orders that would have come through the app anyway |
| Catering and events | Qualified leads captured and the share that turned into deposits | Lead log or CRM | Valuing leads at full event revenue on day one |
| Staff time | Phone minutes moved off the host stand and counter | Phone reports and shift tallies | Counting minutes saved when no schedule or task changed |
Recovered reservations are usually the largest stream for full-service restaurants because booking calls cluster when nobody can answer. OpenTable estimated that a quarter of reservations are made between 10 pm and 10 am[5], which is exactly when the phone is off. For counter-service and takeout-heavy places, speed of answer matters most: 94% of consumers say speed is critical when ordering off-premises[6].
Step 3: price each stream at margin
Use contribution margin, the money left after the food and direct labor that each extra cover or order consumes. If you do not track it, start from your own P&L: sales minus cost of goods minus hourly labor that varies with volume. The worked Example below assumes 35 percent for extra dine-in covers and 25 percent for takeout; replace both with your own figures, and go lower for delivery orders that carry platform fees.
Staff time is only value if something changes. Moving 10 hours a week of phone duty off the host stand is worth the loaded hourly wage only if you cut a shift, avoid adding one, or redeploy those hours to work that earns money, such as table turns or upselling events. The median host earns $15.00 an hour before payroll taxes and benefits[7]; use your own loaded rate.
Step 4: put every cost on the other side
- Vendor fees, including per-use charges, phone numbers and SMS registration.
- Setup time: writing the fact sheet, building handoff rules, testing calls. Value manager hours at their real cost.
- Ongoing review: reading transcripts, updating holiday hours, fixing wrong answers. Budget an hour or two a week at first.
- Integration work with your POS, reservation book or ordering platform.
- Error cost: refunds, comps or lost guests from wrong answers. Track every incident during the trial.
Tools that log every call with a summary and transcript make this review faster; Ava, for example, sends each call to an inbox with both, which doubles as your audit trail for the value count. Whatever you use, make sure you can export call records, because you will need them for step 5.
Step 5: a worked example
Example baseline
In this example, a fictional 110-seat neighborhood restaurant misses 140 calls a month, mostly 6 to 8 pm and after closing. Sampling shows about 40% of those callers wanted a table, 25% wanted takeout and the rest asked questions.
Example recovered reservations
In this example trial month, 56 missed booking calls were answered. Comparing with the reservation book, 30 became bookings that did not exist in the baseline period, averaging 3 covers at a 62 dollar check, so 90 covers and 5,580 dollars in sales. At a 35% contribution margin that is 1,953 dollars.
Example recovered takeout
In this example, of 35 takeout callers who received the ordering link by text, 21 ordered, averaging 38 dollars: 798 dollars in sales, 199 dollars at a 25% margin.
Example staff time
In this example, host phone time fell by about 12 hours a month, and the manager cut one early-evening shift of 4 hours a week, worth roughly 330 dollars a month at a loaded 19 dollars an hour.
Example net result
The example ends with a total monthly contribution of about 2,480 dollars, minus tool fees, 6 hours of manager review time and one comped meal from a wrong answer. If the net is positive for two consecutive months, the case holds.
The Example deliberately leaves catering out. Event leads take weeks to convert, so track them separately and add them to the case only once deposits land.
Chat tools need a different yardstick
Website chat rarely produces a booking on its own; it removes friction before one. Measure it by what it deflects and what it hands on. Deflection is the share of chats that end with the question answered and no follow-up call or email. Handoff quality is how many chats that needed a person reached one with the guest's details attached, so nobody had to ask twice. Conversion is the share of chats that end with a reservation request, an ordering link opened, or a catering lead logged.
Tag every link the chat sends with a source code, such as ?src=chat, so your ordering and booking reports can separate chat-driven traffic from everything else. Without tags, chat value disappears into the general website numbers and the tool will look either useless or magical, depending on who is presenting.
Attribution mistakes to avoid
- Comparing a trial month with a different season. Restaurant sales swing month to month, so compare like with like or adjust for the calendar.
- Counting chats and calls handled, not outcomes. A chat that answers parking questions is useful but not revenue.
- Letting the vendor run the numbers. Use your own POS and reservation data and keep the vendor's dashboard as a cross-check.
- Ignoring guests who hung up on the assistant. Track abandoned calls during the trial as carefully as in the baseline.
- Forgetting the change in staff behavior. If hosts stop answering because the assistant will, missed-call recovery is not real.
A 90-day measurement calendar
- Days 1 to 14: baseline, no changes to staffing or marketing.
- Days 15 to 21: setup and internal test calls; do not count these weeks.
- Days 22 to 51: first measured month, after-hours and overflow only.
- Days 52 to 81: second measured month, all hours with handoff to staff.
- Days 82 to 90: compare both months with the baseline, write a one-page result and decide.
Decision thresholds to agree before day 1: net contribution positive in both measured months, missed calls down by at least half, no double-booked tables, and no unresolved guest complaints caused by the tool.
Keep the one-page result simple enough for an owner or partner to read in two minutes: the baseline, the two measured months, the net contribution in each, the incidents and what was fixed, and a clear recommendation. Attach the raw exports so anyone can check the arithmetic.
For what the assistant should and should not do during those months, see what to expect from AI reservation assistants. Before signing a longer contract, run the vendor privacy questions and check whether your state has AI disclosure rules for customer calls.
Frequently asked questions
How long should I measure before deciding on an AI phone tool?
Two weeks of baseline and two full months of use is a sensible minimum. One month can be distorted by weather, holidays or a local event.
Should I count every booking the AI takes as new revenue?
No. Count only bookings from calls or chats that would previously have gone unanswered, such as after-hours and rush overflow calls. Bookings staff would have answered anyway are not incremental.
What margin should I use for recovered covers?
Use your own contribution margin after food and variable labor. Food and labor each take roughly a third of sales on average[2], so incremental margin is far below the check total.
Is staff time saved real money?
Only when a schedule or task changes: a shift cut, a hire avoided, or hours redeployed to revenue-earning work. Otherwise record it as capacity, not savings.
References
National Restaurant Association (2026). Research Insight: The Future of Restaurant Technology (press release, November 5, 2026) (opens in a new tab) ↑
National Restaurant Association (2026). Restaurants remain resilient despite challenging business conditions (opens in a new tab) ↑
National Restaurant Association (2026). Restaurants remain resilient despite challenging business conditions (opens in a new tab) ↑
National Restaurant Association (2026). Research Insight: The Future of Restaurant Technology (press release, November 5, 2026) (opens in a new tab) ↑
Cornell University Center for Hospitality Research (Sheryl E. Kimes) (2009). How Restaurant Customers View Online Reservations (Cornell Hospitality Report Vol. 9, No. 5) (opens in a new tab) ↑
National Restaurant Association (2025). From Trend to Transformation: Off-Premises Dining Now Essential for Restaurant Consumers, Operators (opens in a new tab) ↑
US Bureau of Labor Statistics (2025). Occupational Employment and Wage Statistics, May 2025 national estimate: Hosts and hostesses (restaurant, lounge and coffee shop) median hourly wage (series OEUN000000000000035903108) (opens in a new tab) ↑