Seven techniques raise a server's average check without pressure: name one signature dish, offer a specific upgrade, give two options instead of a yes-or-no, pair a drink with the dish, preview dessert early, add a side with the main, and check back before the second round. AI drafts the table-side wording; the manager checks the facts.
Most lists of upselling tips stop at the technique. The part that gets used on a busy Friday is the sentence the server actually says, so each technique below comes with a script, a note on where it goes wrong and a way to rehearse it in a ten-minute huddle. The scripts are drafted with an AI assistant from your own menu, which keeps them specific: "the slow-braised short rib" beats "our chef's specials" every time.
Give the AI your menu and your margins, then ask for scripts
A general chat assistant (ChatGPT, Claude or Gemini; free tiers are enough) writes much better scripts if it has the real menu in front of it. Paste your food and drink menus as text, then add the information a server cannot see on the printed page: which dishes carry the best margin, which sell slowly and why you want to move them, and which drinks pair with what. A starting prompt:
You help me write table-side scripts for restaurant servers.
Menu (with prices): [paste]
Dishes I want to sell more of: [list, with a reason each]
Dishes with allergens flagged: [list]
Rules: each script max 25 words, spoken, no jargon, one clear offer,
never mention a price unless the guest asks, never answer allergen
questions (say "I'll check with the chef"), accept "no" in one line.
Write [number] scripts for [technique]. Mark anything you are not
sure is on the menu with [CHECK].Run this for each technique below. Keep the best scripts in one document you can print. If you want the menu itself to sell harder before a server speaks, writing menu descriptions with AI that still sound like you covers that side, and AI menu engineering helps you pick which dishes are worth the push.
What the assistant gives back, and what to fix
Ask for five scripts for Technique 1 and a typical result (illustrative) looks like this: "Tonight I'd start with the burrata, it's rich, fresh and perfect to share." "You can't go wrong with our legendary short rib." "Our chef absolutely loves the halibut, it's a total showstopper." Three problems show up at once. The second and third lines use hollow words ("legendary", "showstopper"), the third claims what the chef thinks, which nobody has told the model, and the halibut may not be on tonight. Fix each by replacing the adjective with a fact (braised since morning, comes with brown butter), deleting any claim about a person's opinion, and checking every dish against tonight's availability. What remains is usually two usable lines out of five, which is a good return for two minutes.
Technique 1: Recommend one named dish, not "our specials"
Guests who hear "anything to start?" say no. Guests who hear about one named dish with one reason often say yes, because the server has taken a decision off them. The rule: name the dish, give one sensory reason, and stop.
Script (illustrative): "The burrata with roasted tomatoes is what I'd start with tonight. It's creamy and sharp and it shares well between two."
Where it goes wrong: servers recommend the same dish to every table, which guests overhear at the next one. Ask the AI for three recommendations per category and rotate them by the day, or by what the kitchen has too much of. Also check the recommendation fits the table: a dish heavy with cream is a poor suggestion to a guest who has just asked for something light.
Technique 2: Offer a specific upgrade by name
"Would you like to upgrade?" is a closed question about an abstract thing. "Would you like the aged cheddar on that burger for an extra $2?" is a closed question about a real thing. Better still, build the upgrade into the order: "Do you want the burger with the aged cheddar and the bacon, or plain?" The guest picks between two products rather than between spend and no spend.
Script (illustrative): "Would you like that with the smoked cheddar, or the blue cheese? Both are $2."
A worked example with numbers (illustrative): a casual restaurant sells about 180 burgers a week at an average of $14. If servers offer the named upgrade to every table and 25% accept, that is 45 upgrades at $2, or $90 a week. Over a year that is roughly $4,680 before cost of goods, from one sentence. The point is not the sum, which depends on your food costs; it is that a small, specific offer repeated consistently beats a large, vague one made occasionally.
Technique 3: Offer two choices, never yes or no
"Would you like a side?" has a built-in no. "Would you like the fries or the greens with that?" assumes the guest wants one and asks which. This works best with sides, sauces and sharing plates, where both options are sensible and neither is a hard sell.
Script: "Would you like the roasted carrots or the crisp potatoes with that? People often pick both for the table."
Where it goes wrong: it can feel presumptuous when a guest has clearly finished ordering. The fix is timing, not a different script. Offer it as the guest settles on the main, not as they hand back the menu. Ask the AI to write a second version for guests who sound uncertain: "The carrots are lighter, the potatoes are richer. Which sounds better?"
A second illustration: the quiet midweek lunch
Not every service is a Friday. A 30-seat café on a quiet Tuesday might serve 40 covers at an average of $16. Suppose the manager picks two techniques only, Technique 3 (two choices on sides) and Technique 7 (check back for a second drink or a coffee), and trains for two weeks. If average spend per cover moves from $16.00 to $17.20, the gain is $1.20 across 40 covers, or $48 a day, which over a five-day lunch week is $240 (illustrative). It is small and it is real, and it costs nothing but a huddle, which is why quiet services are a better place to start than Friday night, when servers have no spare attention to practise a new line.
Technique 4: Pair a drink with the dish
Drinks are usually the highest-margin line on the bill, and guests who would not order a second drink unprompted will accept a pairing that is explained in one sentence. The trick is to pair by the food, not by the drink list.
Script: "The lamb works nicely with the Malbec by the glass, or the sparkling water with lime if you'd rather keep it light." Note that the script offers an alcohol-free option next to the alcoholic one. Not every guest drinks, and a server who pushes alcohol at a table that has declined it will lose goodwill. Responsible service rules apply wherever you operate; build any age checks or limits into your policy, not into a script.
How to produce the pairings: paste your drinks list into the assistant and ask for a pairing table, "dish, first drink option, alternative, one-sentence reason". Then ask your bar lead to delete anything wrong. The AI knows general pairings well but not what is currently in stock, and it will happily recommend a bottle you stopped buying in spring.
Technique 5: Mention dessert at ordering, not after clearing
Once plates are cleared, many guests are full and thinking about the bill. A short mention at the point of ordering plants the idea: "Save a little room, the sticky toffee pudding takes 15 minutes so I can put it in for you whenever you're ready." A dish that needs baking time gives the server a reason to bring it up early.
Script: "If you'd like dessert, the pudding takes about 15 minutes, so tell me when you're ready and I'll start it."
The after-clearing version should be a plain offer of two options: "Can I tempt you with the pudding or an espresso with something small?" Ask the AI to write both, plus a version for tables celebrating something, with a rule that celebrations are never assumed, only acknowledged if the guest mentions one.
Example: a short "no" script, so declines stay friendly
The most underrated part of any upselling script is what happens after a no. Ask the AI to write three one-line responses to a decline, then pick the one that sounds like your team: "No problem, I'll keep an eye on your drinks." "Of course, just shout if you change your mind." "Lovely, your starters will be a few minutes." None of these repeats the offer. A guest who hears one polite offer and a graceful exit is more likely to say yes next time than one who gets the line twice.
Technique 6: Pair the main with a side, and name the sharing dish
Many mains arrive looking plain on the plate, and a side makes the table look and feel generous. Naming a shared dish helps larger tables: "The cauliflower is big enough for four." Share-led recommendations lift the check without feeling like an upgrade because the whole table benefits.
Script: "The charred greens and the polenta chips both work for the table. We'd usually do one of each for four."
A realistic mistake: a server suggests three sides, overwhelming the guest, who orders none. Give the AI the limit explicitly: "offer at most two options at a time". Also watch portion sizes: if your sides are small, a suggestion for six might disappoint, so describe the size honestly.
Technique 7: Check back before the second round
The last technique is timing. The best moment to suggest a second drink is when the first is about three quarters finished, and the second best is just after the food lands. A check-back at the right time gives the server a natural opening: "How is the Malbec? I can bring another glass, or a different one to try with the cheese."
Script: "How's the drink? Can I get you the same again, or something different with the next course?"
This technique has no price in it, which is why it travels well to a quiet midweek lunch. In practice, the AI contributes less here than a floor manager who watches the room; use it for the huddle phrasing and let the manager coach the timing.
A pre-shift huddle sheet you can print
A script nobody rehearses is a script nobody uses. A ten-minute huddle before each service gives servers one technique, one dish and one rehearsal. Ask the AI to turn your week's focus into a one-page sheet with a consistent layout (illustrative example for a Friday):
FRIDAY HUDDLE
Focus technique: Name one dish (Technique 1)
Dish to push: Slow-braised short rib, $24. Reason: kitchen has extra, goes
with Malbec by the glass.
Say it like this: "The short rib is what I'd have tonight. It's been
braising since this morning and it falls apart on the fork."
Do not push: Halibut (supplier delay). Allergens: short rib contains
[kitchen to confirm]. Any allergy question goes to the chef.
Today's pairing: Short rib with Malbec by the glass, or sparkling water.
Rehearse: each server says the line once to the person on their left.
Target: 3 tables out of 10 order the short rib.Notice the line "kitchen to confirm". The AI should never fill in allergens; the kitchen does, from the recipe record. For a restaurant using AI tools more widely, twelve questions to ask before you sign up is a useful check on anything that touches guest data.
Choosing what to push from your own sales report
Do not pick the dish to promote by gut feel alone. Export a month of item sales from your point-of-sale system, paste the list into the assistant, and ask three plain questions: which five items sell least relative to the menu, which five drinks have the highest sales per serving, and which starters are most often ordered with which mains. The assistant can sort and summarise a pasted table well, but check any total against the report, since sums are where language models slip. The answers give you the huddle's focus: push the high-margin item that already has a loyal audience, not the one that never sold. If an item sells slowly because guests dislike it, no script will rescue it; take it off, or change the recipe.
Training a new server in one shift
New starters are where scripts pay for themselves, because an untrained server defaults to "anything else?". On their first shift give them three things only: the huddle sheet, the two scripts for the current focus, and a shadow shift with a colleague who says the lines naturally. Ask the AI for a one-page "first shift card" with the focus dish, the two scripts, and a short list of dishes they must never answer questions about without checking (allergens, preparation methods, anything off-menu). In a fast service, a card they can glance at in the pass beats a binder they will never open. At the end of the shift, ask what felt awkward; the awkward lines are the ones to rewrite.
A full Friday, walked through
An illustration of how the techniques sit together at one table of four: the server greets, then names one starter (Technique 1: "the burrata is what I'd start with"). When the mains are ordered she offers the two sides (Technique 3), mentions a Malbec by the glass or sparkling water with the lamb (Technique 4), and tells them the pudding takes 15 minutes (Technique 5). Part way through the mains she checks back about drinks (Technique 7). That is five light touches in a meal, none longer than a sentence, none repeated after a no. Added up, the table might order one extra starter, two sides, two glasses and one dessert; at illustrative prices of $11, $5 each, $9 each and $9 that is $48 over the baseline. A guest who ordered everything unprompted would not have felt pushed, and that is the standard to aim for.
How to tell whether average spend really moved
A change in takings is not a result. The way to tell:
- Pick two numbers from your point-of-sale reports: average spend per cover, and drinks as a share of sales. Total sales alone mislead, because a busy night always looks better.
- Record four weeks before you change anything, for the same weekday and service. Fridays against Fridays.
- Train on one technique at a time for two weeks, then add the next.
- After four weeks, compare the same weekday. An illustration: Friday average spend per cover rose from $38.00 to $40.50, which is $2.50 or about 6.6%. On 120 covers, that is about $300 a Friday. It is only believable if the guest count and the menu stayed roughly the same.
- Read reviews and comments alongside: using AI to analyse what diners are saying turns a month of reviews into themes. If "pushy" starts appearing, ease off.
Keep the scope honest: the AI writes words, the point-of-sale system gives you the figures, and a person decides what to change. If a number looks too good, check for a holiday or an event that week before you credit the scripts.
What the scripts should never do
- Never claim a dish is something it is not. A made-up provenance or an invented "chef's favourite" is a misrepresentation, and the AI will write one without hesitation if you ask it to be persuasive.
- Never answer allergy or dietary questions from a script. Every one goes to a person who has the recipe record.
- Never push drinks at a table that has declined them, or at anyone who appears to have had enough.
- Never repeat a declined offer. One try, one line, then move on.
- Never use a script to hide a price the guest has asked about. If they ask, say it.
If you run a very small team, start with Technique 2 and the huddle sheet only. Two changes consistently applied are worth more than seven handled loosely, and by the third week the servers will be adjusting the wording themselves, which is the sign the training has landed.
Questions managers ask about upselling scripts
Will scripts make servers sound robotic?
Only if they recite them word for word. Give each server a base line and ask them to put it in their own words in the huddle. The AI drafts the structure and the specifics (the dish name, the price, the reason); the server supplies the voice.
How long before average spend shows a change?
Allow four weeks of the same weekday and service to compare, because single nights swing with weather, events and party size. Compare Fridays with Fridays, and look at average spend per cover rather than total sales.
Should AI write the allergen answers?
No. Allergen and dietary answers must come from the kitchen's own recipe records and a named person. A script should send any allergy question straight to the manager or chef, never to a prepared line.
Is upselling pushy?
It is when it ignores the guest. A good script offers one specific option, accepts a no gracefully and never repeats the offer. Track complaints and guest comments next to spend; if comments turn sour, the scripts are too strong.
Further reads
- AI Food Insights Tool: Toast IQ, Square AI or a Spreadsheet? — Read item sales from your POS before picking a dish to push.
- How to Build a Kitchen Training Manual With AI — Put the scripts into a wider staff training manual.
- Should an AI Chatbot Answer Allergen Questions for Your Restaurant? — Why allergen answers stay with people, not AI.
- How Much Does AI Cost a Small Restaurant Each Month? — What AI tools cost a small restaurant each month.
- Trainual for Restaurants: Does Its AI Beat ChatGPT-Built SOPs? — Whether Trainual's AI beats ChatGPT-built training notes.
- How to Build a Restaurant Staff Rota With AI Demand Forecasts — Put your strongest servers on the busiest shifts.
- ChatGPT Prompts for Restaurants: 10 for Menus, Rotas and Replies — Ten prompts for menus, rotas and customer replies, with inputs, worked outputs and the checks that catch a wrong total, label or promise.
- Restaurant Loyalty Programme: Simple Rewards and AI-Written Offers — A no-app loyalty scheme: one reward rule, tracking in a sheet or till, consent done properly, three offer types written with AI, and repeat visits measured.
- Hotel Front Desk Scripts: Can AI Role-Play Difficult Guests? — How to turn your own guest disputes into AI role-play drills for front desk staff, with an in-character prompt, a scoring sheet and trainer notes.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: no vendor facts used beyond general ChatGPT, Claude and Gemini access; figures in worked examples are illustrative, not measured.