How AI actually drafts a catering proposal (and what it still can't do)

There is a version of this post that would tell you AI replaces your sales coordinator. That is not true, and operators who trust software vendors that promise it end up disappointed and re-hiring the coordinator six months later. Here is the honest version.

What AI drafting does well

When a prospect emails you "we're hosting 60 people for our annual retreat on May 14, looking for heavy appetizers and a dinner buffet, company is vegetarian-leaning, budget is around $70 a head," a well-tuned catering proposal model can do five things well.

Structure. It parses that email into fields: 60 guests, date May 14, service type heavy apps plus dinner buffet, dietary preference vegetarian-leaning, target per-head $70. It does this as reliably as a good coordinator on their third cup of coffee, and faster.

Menu selection from your catalog. Given a tagged menu where every item has allergen flags and dietary tags, the model picks options that fit the brief. For a 60-person vegetarian-leaning dinner it selects things like grain bowls as the hero, a protein option for the omnivores, two or three passed apps that work for the crowd, and sides that are visually varied on a buffet line.

Pricing at the margin floor. The model knows your contribution margin floor because it has your recipe-cost data and your labor model. It produces a quote that hits the $70 target and flags where it is getting there by cutting into labor versus food cost. It will not build a $70-a-head buffet that loses you money on paper. This is a meaningful upgrade from a brand-new coordinator who sometimes does.

Allergen and dietary propagation. Every menu item in the draft has the correct allergen flags, pulled automatically from the recipe cascade. The AI does not forget to note that the grain bowl is gluten-free because the recipe itself is tagged that way. A human coordinator forgets this about 4% of the time under stress.

First-pass tone matching. If your existing proposal library uses a specific voice (warm, specific, declarative), modern language models can match it within a few proposals of seeing examples. Good vendors fine-tune on your actual sent proposals; the copy in the draft reads like you wrote it.

For a 60-person corporate retreat, a good AI first-draft gets you to about 80% of a send-able proposal in under 90 seconds. Your coordinator then applies the last 20%: the judgment calls.

Where the ceilings are

This is the list your software vendor does not want to write. Do not trust them if they do not address it.

The model does not know your relationship with this client

Five years of history with a corporate client means your sales lead knows that their HR director is the real decision-maker, that their CFO always pushes back on 20% gratuity, that they are strict about invoice terms. The email you are parsing says nothing about any of this. The AI writes a generic proposal. Your coordinator pulls up the account, sees the pattern, and tweaks three things: moves gratuity off the top line, adds a net-30 line instead of due-on-receipt, and frames the vegetarian-leaning options as "per Karen's note last year."

None of that relationship context is in the email. The AI cannot draft it because the AI does not have it. A well-built catering CRM does, but the AI pulling from the CRM is still working on past data points; it is not feeling the pattern.

The model does not know which venues are difficult

"We're doing it at the Riverfront Hotel in the Magnolia Room" is a detail that a veteran coordinator reads as a red flag. The Magnolia Room has a service corridor that requires an extra 90 minutes of setup. Loading dock access after 4pm requires venue-ops approval. The hotel charges $200 for the use of their warming cabinets, which you can avoid by bringing your own. None of this is in your CRM. It lives in your ops lead's head.

The AI proposal will quote labor for a standard venue. Your human team corrects it before sending. If it does not get corrected, you learn the lesson on the event day when three of your FOH people are standing in a corridor waiting for access and your labor cost blows out.

The model cannot feel when to push for a tasting

Some proposals should not be a proposal at first. They should be an invitation to come in for a tasting. A good sales lead reads signal in the inquiry email (budget round, high-stakes corporate anniversary, new planner vendor for them) and intuits that the right first move is a tasting, not a quote. The AI does not feel that. It just produces the quote.

Long-run pattern recognition on booking probability

An AI model can in principle estimate "this inquiry has a 73% chance of booking." In practice, most catering operators do not have enough structured historical data to train a useful estimator for this. The ones who do are typically doing 2,000+ events a year. If you are a 400-event-a-year operator, an AI probability estimator on booking is pulling from too thin a pattern sample to be worth anything and will confidently tell you bad things.

Be careful with any vendor who sells you "AI conversion scoring" without being specific about how much of your historical data was in the training set.

Hallucinations on specific facts

Every AI system will occasionally produce a proposal with a menu item that does not exist in your catalog, or a price that is not your actual price, or a reference to an event feature you do not offer. A well-engineered drafting system keeps the hallucination rate under 1% and flags the draft for human review before send. A poorly-engineered one lets it through.

This is the single most important question to ask a software vendor claiming AI proposal drafting: what is the measured hallucination rate, and what is the review gate before send? If they do not have a number, the number is higher than you want it to be.

The right frame: AI as first-draft assistant

The frame that works is simple. The AI produces a draft that is 80% good. The human produces the last 20%. The last 20% is where the value lives. It is the part that accounts for client history, venue quirks, relationship intuition, and the judgment calls the AI cannot make.

This cuts proposal turnaround from a realistic 30 to 45 minutes down to a realistic 6 to 10 minutes, depending on event complexity. Operators using this correctly see their sales-coordinator capacity roughly triple. They do not lay off coordinators. They let each coordinator run more deals, win more, and spend more of their time on the high-value judgment parts of the job.

We built the Caterforia proposal drafting feature around exactly this frame. The AI never sends a proposal. The AI produces a draft. Your coordinator reviews, edits, and sends. The economics work because the drafting is fast and the review is meaningful.

What to look for in a vendor

Three questions, in order, when evaluating an AI catering drafting vendor.

1. What data is the draft pulled from? If the answer is "our general model," pass. If the answer is "your menu catalog, your recipe data, your customer history, fine-tuned on your last 500 sent proposals," keep talking.

2. What is the hallucination rate and what is the gate before send? If the answer is vague, pass. If the answer is specific (typically under 1% with a mandatory human review), keep talking.

3. Does it write proposals or does it assist writing proposals? If the pitch is "fully autonomous proposal sending," pass. The autonomous version will cost you a six-figure booking inside of 18 months when a hallucinated menu item shows up on a contract. If the pitch is "draft in 90 seconds, you send after review," you are talking to the right vendor.

Where this goes next

The honest forecast for the next 24 months in AI proposal drafting is not "fully autonomous." It is "review time drops from 8 minutes to 3 minutes" as the drafts get better. That is still a meaningful economic gain. An operator at 400 events a year saves roughly 40 hours of coordinator time a month at that delta. That is useful. It is not the post-human sales team your vendor may be pitching you.

When we say Caterforia's AI pulls real weight, this is what we mean. Not magic. Not autopilot. Real work on the drafting end, reviewed by a human who is still doing the hard parts.

Start at $1 a month for the operating system. Add AI when you are ready to integrate it into how your team works.