A8gent
All articles

Industry

Restaurant Catering Enquiry Automation: The $3,000 Email

A table for two books itself. A catering order or a private room enquiry arrives by email, web form, Instagram DM or voicemail into an inbox owned by nobody, and sits there through service. It is the highest-value inbound a restaurant receives and the worst-handled, for reasons that are structural rather than lazy.

A8gentAutomation practiceAugust 5, 202613 min read
Contents
  1. Why These Enquiries Get Lost
  2. Step One Is Not AI. It Is One Inbox.
  3. Automate the Acknowledgement, Not the Answer
  4. Qualification and Routing
  5. The Systems: What Connects and What Does Not
  6. A Worked Example, With the Arithmetic Shown
  7. Running It Without Embarrassing Yourself
  8. FAQ

Why These Enquiries Get Lost

It is worth being precise about the mechanism, because the fix follows directly from it and the wrong diagnosis produces the wrong build.

They arrive on the wrong channels. A reservation arrives through a reservation system, which is a queue with a state and an owner. A catering enquiry arrives at info@, or through a web form that emails info@, or as an Instagram DM, or as a voicemail, or as a note the host wrote on a pad. None of those is a queue. Several of them are inboxes that three people have the password to and nobody checks.

They arrive at the wrong time. Enquiries cluster in working hours, which for a restaurant is prep and lunch service. The person who can answer - an owner, a GM, an events manager - is on the floor. By the time they sit down it is 4pm, and by the time they are free properly it is midnight.

They are not urgent to anyone inside the building. A table waiting for food is urgent. An email about a party in six weeks is not, right now, and it will not be right now tomorrow either. This is the same absence-of-a-forcing-function problem that produces lapsed patients and unfollowed referrals elsewhere in this series.

And they are the most valuable thing in the inbox. A catering order or a private room booking is worth many multiples of a cover, is usually prepaid or deposited, has predictable labour, and frequently repeats annually. Losing one costs more than a slow night.

The competitive dynamic makes it worse. The person enquiring is almost never contacting only you. They are emailing four venues on a Tuesday afternoon, often on behalf of someone else who wants an answer, and they will have a shortlist by Wednesday. The venue that replies within the hour joins the conversation. The venue that replies on Thursday is replying to a decision. That is not a claim about human psychology - it is a claim about a process the enquirer is running, and you can verify it by asking the next three people who book you how many other places they contacted.

Our speed to lead and lead response time entries cover the general mechanics. For where this sits among a restaurant's other automation candidates, see AI automation for restaurants and our broader AI agents for restaurants piece.

Step One Is Not AI. It Is One Inbox.

Before any automation, the enquiries must land in one place with a state and an owner. This is unglamorous, it is most of the value, and skipping it is why sophisticated builds fail on top of a broken foundation.

Enumerate every channel. Actually enumerate them - walk the building and ask. The list is usually longer than management thinks: the website form, the generic mailbox, a personal mailbox belonging to whoever set the website up, the phone, voicemail, Instagram DMs, Facebook messages, WhatsApp, a third-party listing site, and the note pad by the host stand. Each one is a place an enquiry can die.

Funnel them into a single queue. Email forwarding, form webhooks, a shared inbox tool, a simple CRM - the specific tool matters far less than the property that every enquiry appears in one list. Social DMs are usually the hardest to automate and often need a human to copy them in; that is acceptable, provided it is somebody's explicit job.

Give every enquiry a state and an owner. New, acknowledged, qualifying, quoted, won, lost, no-response. One named person accountable. An enquiry with no owner is an enquiry with no next action.

Track age from first contact. Not from last activity - from arrival. Anything past a threshold escalates to a person by name. This is the mechanism that stops the 4pm problem.

If you build nothing else from this article, build that. A shared queue with states and ages will recover more enquiries than any AI layer bolted onto a chaotic intake, and it can be assembled in a day.

Automate the Acknowledgement, Not the Answer

The instinct is to have AI answer the enquiry. That is the wrong first move, and often the wrong move entirely.

What to automate immediately: an acknowledgement that goes out within minutes, on the channel the enquiry arrived on, from a real named person, that does three things - confirms receipt, sets an expectation for when a real answer arrives, and asks the two or three questions you will need anyway.

That message is genuinely valuable and carries almost no risk. It puts you in the conversation while the enquirer is still shortlisting, and it starts collecting information in parallel rather than after your first reply. It does not need a language model, though one can help draft variants by enquiry type.

The questions are the automation's real work. Date, time, headcount, format (drop-off catering, staffed event, private room, full buyout), budget range if they will give it, and any dietary or accessibility requirements. Ask few enough that people answer - three or four - and design them so the answers can be structured rather than prose.

What not to automate: the quote. A catering quote depends on whether you have staff that Saturday, whether the date collides with something else, whether the menu they want is feasible at that headcount, and what margin you need. Those are judgement calls made by a person who knows the operation. An automated quote is a public commitment, and an automated quote that is wrong is either a loss you honour or a conversation where you retract a price - both worse than a slower reply.

What not to automate: availability confirmation. Telling an enquirer the private room is free on the 14th, when it is provisionally held, is a specific and painful failure. Unless your system reads a single authoritative calendar that everyone actually uses - and in most restaurants no such thing exists - say you will confirm, and confirm.

The judgement rule: automate anything reversible and cheap to be wrong about; keep a human on anything that commits the business. Acknowledgement, question-asking, chasing, routing and reminding are all reversible. Prices, dates and capacity are not. Our what AI agents can do reliably piece covers the general boundary, and when not to use an AI agent the cases where the answer is simply no.

Qualification and Routing

Not every enquiry deserves the same response, and a system that treats a 12-person birthday the same as a 200-person corporate function wastes the manager's attention on the wrong one.

Route on value and complexity, using data you already have. Headcount and format are usually enough. A small drop-off order can often be handled entirely by a templated reply and an order link. A large staffed event needs the events manager, personally, quickly. A full buyout needs the owner. Getting this routing right is worth more than any improvement in copy.

Route on date proximity too. An enquiry for next Friday is a different urgency from one for a date in November. The near one may be un-fulfillable and should be answered fast either way, because a quick honest no is a good customer experience and frees everyone up. The far one has time for a proper conversation but will go cold if left.

Handle the ones you should decline, deliberately. Restaurants lose time on enquiries they were never going to take - dates already booked, headcounts they cannot serve, budgets far below what the format costs. A fast, warm, specific no, ideally with a recommendation of somewhere that can help, costs a minute and earns goodwill that returns as ordinary covers. Build the decline path; most systems only build the win path.

Chase, because they will not. A substantial share of enquiries go quiet after the first exchange - the organiser is waiting on their own boss, or comparing venues, or has simply got busy. Two or three spaced follow-ups on the original channel recover a meaningful number, and almost nobody does it consistently. This is the single most neglected step and it needs no intelligence at all - it needs a queue with ages and a rule. The equivalent build in the trades is our automate quote follow-up guide and the unsold estimate follow-up guide, and the mechanics transfer almost unchanged.

Cap the chasing and close the record. After the final follow-up, mark it lost with a reason. Those reasons, reviewed quarterly, are the most useful marketing data a restaurant can generate about its events business - and they only exist if the workflow forces the field.

The Systems: What Connects and What Does Not

A realistic account of the integration landscape, because assumptions here cause wasted scoping.

Toast publishes a developer guide documenting more than twenty APIs, including Analytics, Cash Management, Configuration, Credit Cards, Device Details, Gift Cards, Kitchen, Labor, Loyalty, Menus, Orders, Order Management Configuration, Packaging, Partners, Restaurant Availability, Restaurants, Stock and Tender. The Orders API "provides access to information about menu orders. You can use the orders API to get information about and create menu orders, checks, and payment," with GET, POST and PATCH supported for partner and custom integrations.

Two caveats that matter for scoping. Toast states that "Partner and custom integration access is granted based on specific implementation requirements. While the methods listed represent the full range of possible access, actual permissions are limited to only what is necessary to support each integration's functionality." So this is scoped, approved access - not a key you generate on a Tuesday afternoon. And access is "further classified by the integration type," with separate documentation for authentication and for refreshing tokens, which tells you a token refresh step is part of any build. See OAuth.

But note where the POS sits in this workflow. The enquiry arrives weeks before anything touches Toast. For most restaurants the POS integration is relevant at the end - turning a confirmed event into an order, a deposit or a report - and is not required at all for the part that is actually broken. Do not let a POS integration become the reason the intake queue does not ship. Build the queue first, connect the POS when the queue is producing confirmed events worth connecting.

Reservation platforms. We attempted to read OpenTable's developer documentation at docs.opentable.com and it rendered as an empty shell - a page title with no documentation content. OpenTable does publish an API partners page and a "become a partner" route plus a partner portal, which indicates access is gated behind a partner relationship. We therefore make no claims about its endpoints, authentication or capabilities; establish directly whether you can get access before designing around it. In any case, large-party and private-event enquiries frequently sit outside the standard reservation flow.

Social channels. Instagram and Facebook DMs are a real and growing share of event enquiries and the hardest to pull into a queue programmatically. Be honest in scoping: a human checking DMs twice a day and forwarding them into the queue is a legitimate design, and far better than an integration project that stalls while enquiries continue to die in an app.

The messaging layer. If you follow up by SMS over a US 10-digit long code, A2P 10DLC applies. Twilio describes it as "the standard that United States telecom carriers have put in place to ensure that SMS traffic to US end users through long code phone numbers is verified and consensual," requiring a registered Brand and Campaign, and notes unregistered senders "will receive additional carrier fees for sending unregistered traffic." See our A2P 10DLC guide and the glossary entry.

A Worked Example, With the Arithmetic Shown

Illustrative. We report no client results - substitute your own figures, which your inbox and booking records can supply.

Baseline. A restaurant receives 22 catering and private event enquiries a month across all channels. Average confirmed event value: $2,600 - your figure. Suppose the current conversion rate is 18 percent: 22 x 0.18 = 3.96 events a month, or $10,296 of monthly event revenue.

Where the losses are. Suppose a month's audit - genuinely worth doing before building - finds that of the 22:

  • 4 were never replied to at all (lost in a DM, a spam folder, or a voicemail nobody played)
  • 7 were replied to more than 24 hours later
  • 6 went quiet after one exchange and were never chased
  • 5 were handled promptly and properly

That audit is the entire business case and it costs an afternoon. The 4 never-replied and the 6 never-chased are 45 percent of all enquiries, and neither requires any intelligence to fix - they require a queue and a rule.

Modelling the improvement. Suppose the queue plus automatic acknowledgement plus a three-step chase converts the never-replied group at the same rate as the rest (18 percent) and lifts the never-chased group by half its base rate:

  • Never-replied recovered: 4 x 0.18 = 0.72 events/month
  • Never-chased uplift: 6 x 0.09 = 0.54 events/month

Total incremental: 1.26 events a month. At $2,600 that is 1.26 x $2,600 = $3,276 of monthly revenue. Restaurants should convert that to contribution, not revenue - if food and labour on catering run at, say, 62 percent, contribution is $3,276 x 0.38 = $1,245 a month, or about $14,900 a year.

Against cost. Build at $2,500 plus $60/month gives a first-year cost of $2,500 + $720 = $3,220 - your quotes; see automation build cost. Payback: $3,220 / $1,245 = 2.6 months.

Two honest observations. First, the entire modelled gain comes from replying and chasing, not from anything clever - which is a strong argument for building the cheap version first and evaluating AI afterwards. Second, the numbers are extremely sensitive to average event value: at $900 rather than $2,600 the same build takes over seven months to pay back, and a restaurant doing small drop-off orders should scope accordingly. Our ROI of AI agents piece covers the framing and payback period the arithmetic; for budget ranges see AI automation cost.

Running It Without Embarrassing Yourself

Measure four things. Median time to first response, measured from arrival rather than from when somebody noticed. Percentage of enquiries receiving any response at all - this should be 100 and initially will not be. Conversion by enquiry type and by response-time band, which is how you learn whether speed actually matters at your venue rather than assuming it. And lost reasons, captured as structured data.

The failure modes worth designing against:

  • Confirming availability the business does not have. The most expensive error here. Never let the automation assert a date is free.
  • The auto-acknowledgement that reads as a brush-off. A generic "we have received your enquiry" from noreply@ is worse than silence for a $3,000 event, because it signals you are not the kind of place that does this properly. Send it from a named person, reference their actual date and headcount, and make it sound like the first line of a conversation.
  • Chasing someone who already booked. Check state before every send. This is the same pre-send check discipline as every other build in this series - see idempotency.
  • Double-handling. Two staff replying separately because the queue has no owner. This is what states and owners prevent, and it looks unprofessional in a way customers remember.
  • The silent stop. A forwarding rule breaks, a form webhook expires, and the queue goes quiet - which looks exactly like a slow month for events. Alert on zero enquiries over a period, not only on errors. See observability and error handling.
  • Losing the history. Event enquirers repeat annually. Last year's thread - what they ordered, what went well, the allergy - is worth real money next year, and only if it was kept somewhere other than one person's mailbox.

Where a voice layer helps. A meaningful share of these enquiries start as a phone call during service, and a missed call is a lost enquiry with no trace at all. A missed-call text back is cheap, catches the call that would otherwise vanish, and feeds straight into the queue you have just built. That is usually a better second investment than making the email replies smarter.

If you want a second opinion on the sequencing, the free automation audit works through it, and automation services explains how we scope. The sibling problem - what happens to guests after the event, and the review that follows - is covered in restaurant review response automation.

FAQ

Should AI reply to catering enquiries automatically?
It should acknowledge them automatically and ask the qualifying questions. It should not quote a price or confirm a date. A catering quote depends on staffing that weekend, competing bookings, feasibility at that headcount and the margin you need - all judgement calls made by someone who knows the operation. An automated quote is a public commitment you either honour at a loss or retract awkwardly, and both are worse than replying an hour later with a correct number.
What is the single highest-value change?
One queue that every channel feeds into, with a state and an owner on every enquiry and an age tracked from arrival. It is not AI, it can be built in a day, and in most restaurants it recovers more revenue than any messaging sophistication layered on top of a chaotic intake. Enquiries do not usually get lost because the reply was poor; they get lost because nobody owned them.
Can I pull catering enquiries out of Toast?
That is not really where they live. Toast documents more than twenty APIs including Orders, which "provides access to information about menu orders" with GET, POST and PATCH for partner and custom integrations - but the enquiry arrives by email, form, DM or phone weeks before anything reaches the POS. Treat the POS as an end-of-process integration for confirmed events, and build the intake queue first. Note also that Toast states permissions "are limited to only what is necessary to support each integration's functionality," so access is scoped and approved rather than self-serve.
Does OpenTable have an API I can use for large party enquiries?
We could not verify what it offers. OpenTable's documentation site rendered as an empty shell to us - a page title with no content - though the company publishes an API partners page, a partner portal and a "become a partner" route, which indicates access is gated behind a partner relationship. We are not going to characterise endpoints we could not read. Establish directly whether access is available to you before designing around it, and bear in mind large-party and private-event enquiries often sit outside the standard reservation flow anyway.
How quickly should we respond?
Fast enough to still be in the conversation, which for most event enquiries means within the hour and certainly the same day. The enquirer is typically contacting several venues at once on a timetable set by someone else, so a reply on Thursday to a Tuesday enquiry is often responding to a decision rather than participating in one. If you want to verify this at your own venue rather than take it on faith, ask your next few confirmed bookings how many other places they contacted.
How many times should we follow up?
Two or three spaced touches on the original channel, then mark it lost with a reason and stop. Enquiries go quiet mostly because the organiser is waiting on somebody else, not because they decided against you, which is why chasing works. It is also the most neglected step in the entire workflow and it requires no intelligence at all - just a queue that knows how old each enquiry is and a rule about what to do at each threshold.
What about Instagram and Facebook DMs?
They are a real share of event enquiries and the hardest channel to pull into a queue programmatically. The honest design for most restaurants is a person checking DMs at fixed times and forwarding them into the queue - explicitly somebody's job, on a schedule. That is far better than an integration project that stalls for two months while enquiries keep dying in an app nobody is accountable for.
What should we do with enquiries we cannot take?
Decline them fast, warmly, and specifically - ideally suggesting somewhere that can help. Most systems only build the win path, so declines sit in the queue consuming attention and eventually get an awkward late reply or none. A quick honest no costs a minute, frees your events manager for the enquiries you can win, and generates goodwill that tends to come back as ordinary covers.

If this resonated

Module 02 is the long version of this post.

Eval set, refusal pattern, kill switch. Build it in a weekend.

All postsLast verified 2026-08-05

Done for you

Want this built for your business?

Tell us the workflow. We scope, build, and ship the agent with guardrails and the numbers to prove it worked. The scoping call is free.

Free scoping call. You own the code.

Request a scoping call