How Pubs and Bars Can Use AI for Events, Rotas and Stock

Start with stock, because that's where most bar money leaks: use bar inventory software, or a regular count plus AI, to compare what was poured with what was sold. Then feed 12 weeks of till data and your events calendar into an AI forecast to shape rotas, and use AI to draft event content such as quiz rounds and posts, checking every fact.

The order matters more than the tools. Events drive both staffing and stock: a match day, a band or a quiz night changes how many people you need behind the bar and how much you'll pour. So the events calendar is the input everything else hangs from. And no AI counts bottles for you. It spots patterns in counts you've already taken, which means the counting routine has to exist first.

Follow me on Instagram@sagnikteaches

Bar stock: a spreadsheet and AI, or bar inventory software?

There are three realistic set-ups for a single pub or bar. The right one depends mostly on how much you sell, because that decides what a percentage point of waste is worth.

Connect on LinkedInSagnik Bhattacharya

A break-even rule helps. Work out what one percentage point of drinks cost is worth to you each month: monthly drinks sales × 1%. At $20,000 a month in drinks sales, one point is $200, which is close to the cost of a dedicated tool. If you're selling well above that and suspect your pour cost (drinks cost as a percentage of drinks sales) is a few points higher than it should be, software is likely to pay back. Below it, start with the spreadsheet and move up only if the counts show a real leak.

Subscribe on YouTube@codingliquids

Counting speed matters as much as the software. Running a stocktake faster with AI and a phone scanner covers ways to get a full count done in a quiet morning.

Finding the leak: a monthly variance check

Variance is the gap between what you should have used, based on what the till says you sold, and what actually left the shelf. The calculation is simple and belongs in a spreadsheet, not a chat window:

1. Actual usage = opening stock + deliveries − closing stock, for each product, in bottles or litres.
2. Expected usage = units sold on the till × the measure or recipe for each drink.
3. Variance = actual − expected, in units and in cost.

One product worked through, with illustrative figures, shows the sums. House vodka opened the month at 8.0 bottles, 12 were delivered, and 5.2 were left at the closing count, so actual usage was 14.8 bottles. The till recorded 308 single measures of 30 ml, which is 9.24 litres, or 13.2 bottles of 70 cl. The variance is 1.6 bottles, about 37 measures, or a little more than one unrecorded measure a night. At $18 a bottle that's $29 for the month. The finished sheet, one row per product, looks like this:

Chat assistants are unreliable at arithmetic across dozens of rows, so let the spreadsheet do the sums (why AI is bad at maths explains why). Use AI for the two jobs it's good at: building the sheet and reading the results. Paste the finished variance table, with no staff names, and ask:

Here is last month's stock variance for my bar, by product:
product, actual usage, expected usage, variance in units, variance cost.

1. Rank the ten largest variances by cost.
2. For each, list the most likely causes for a bar, in order of
   likelihood (over-pouring, drinks not rung through, staff drinks or
   comps not recorded, line-cleaning waste, breakage, wrong recipe in
   the till, counting error, theft).
3. Say what I could check next week to tell those causes apart.
4. Flag any product where the variance looks like a counting or
   recipe-setup error rather than real loss.

Don't recalculate the numbers; treat my table as correct.

An illustrative reply, trimmed to the top three items, looks something like this:

1. House vodka: 1.6 bottles over expected ($29). Most likely: free-
   pouring on doubles, drinks not rung through, comps not recorded.
   Check: spot-measure pours on two busy nights; compare the comp
   notebook with the till's no-sale and void report.
2. Draught lager: 38 litres over ($61). Most likely: line-cleaning
   waste, fobbing when kegs change, pints topped up after settling.
   Check: record litres lost at each line clean for a month.
3. Aperitif liqueur: 4.2 bottles UNDER expected. Flag: more sold than
   used suggests a setup error, e.g. the till deducting a double
   measure for a spritz that uses a single.

What I'd do with it: item 3 is the most useful line in the reply, because it points at a till setting that also distorts your margin reports. Item 2 is a guess until you tell the assistant how often you clean lines and roughly what each clean loses, so add that and ask again. And the vodka figure is small on its own; the question is whether it repeats next month.

The fourth instruction in the prompt earns its place. A spirit showing a large variance month after month often turns out to be a till button set to the wrong measure, not a heavy-handed bartender. For draught, expect some loss from line cleaning and fobbing; what matters is whether it's stable from month to month.

Private hire creates a variance of its own. If a party's drinks package is rung through as one "Private hire: drinks package" button, the till records no individual glasses, so every bottle of prosecco poured that night shows as missing: two cases, twelve bottles, appear as loss. Before the monthly check, either ring package drinks through against a zero-price button per product, or note each package's agreed contents in the sheet as expected usage. Otherwise the AI will rank the prosecco top of the leak list, and it will be right about the numbers and wrong about the cause.

Rotas built from the till, not from habit

Most pub rotas are last month's rota with tweaks. A forecast based on your own sales does better, and a small bar can build one without special software:

1. Export hourly sales (or transactions) by day for the last 12 weeks from your till. Most systems have an hourly sales report.
2. Mark the unusual days: events, public holidays, big sports fixtures, heatwaves, the day the street was closed.
3. Set your staffing ratio, for example one person behind the bar per so many transactions an hour at your venue. Your shift leads know this number even if it has never been written down.
4. Ask AI to turn it into a staffing guide, as below.
5. Check the draft against people: availability, contracted hours, and the rest-break and working-time rules that apply to your staff. AI doesn't know those, and it doesn't know that one bartender is worth two on a Friday.

Attached: 12 weeks of hourly transactions by day for my pub, plus a
list of dates with events or unusual trading.

1. Excluding the unusual dates, give the typical transactions per hour
   for each day of the week, with a low and high range.
2. Separately, show how event nights differed from a normal night on
   the same weekday.
3. Using a ratio of one bar staff member per [X] transactions an hour
   (minimum [Y] on at any time), suggest how many bar staff I need
   each hour for a normal week, and the extra for a quiz night and a
   live music night.
4. Point out hours where I'm probably overstaffed or understaffed
   compared with these numbers.

An illustrative extract of the reply, using one bar staff member per 25 transactions an hour and a minimum of one:

Tuesday (normal)   5-7pm:   8-14 per hour    -> 1 bar staff
                   7-11pm:  30-45 per hour   -> 2
Tuesday (quiz)     8-9.30pm: 70-95 per hour  -> 4
                   9.30-11pm: 40-55 per hour -> 2
Overstaffed: Tuesday 5-7pm (2 rostered, 1 needed)
Understaffed: quiz night 8-9.30pm (2 rostered, 4 needed)

The shape is right and the owner acts on it, with two corrections. The bar has room for three people behind it, not four, so the fourth quiz-night hour becomes a floor person collecting glasses and running food. And one "normal" Tuesday in the data was a public holiday that nobody had marked, which pulled the early-evening range up. Mark it, rerun, and the 5-7pm figure drops further.

Rota tools can do this automatically once you outgrow the spreadsheet. Deputy, for instance, offers demand forecasting that pulls sales data from a connected till and builds schedules around it. AI shift scheduling versus a rota spreadsheet compares the two routes, and building a restaurant rota from AI demand forecasts goes further into the method. Keep staff names and personal details out of consumer AI accounts; do the name-level rota in your scheduling tool, or use a business plan that doesn't train on your data.

Events: quiz nights, live music and match days

Events are where AI saves the most visible time, and also where it can embarrass you in front of a full room. Four jobs it does well:

• Quiz rounds. Ask for questions by theme and difficulty. Then check every answer against two reliable sources, because assistants state invented trivia with total confidence. A disputed answer at 10pm is worse than a slightly easier round.
• Event specials. Pairing a drinks offer to a night, or a match-day sharing platter, costed from your own prices. Let AI suggest the idea and let the spreadsheet cost it. A quiz-night "team pitcher" of four pints for $20, against $24 bought separately, still works if a pint of lager costs you $1.30: $5.20 of cost on $20 of sales keeps the margin at 74%.
• Posts and listings. A week-ahead post, a day-of reminder, the event listing for your website and business profile. Writing weekly Google Business Profile posts with AI suits the regular events calendar well.
• The post-event review. Paste the night's hourly sales next to a normal night on the same weekday and ask what changed: when the rush came, which drinks sold, whether the extra staff were needed.

Event posts need the same editing as quiz questions. An illustrative first draft for a band night: "Get ready for an unforgettable night of epic live music! Grab your mates and join us for the party of the year!!" It tells nobody when, who, or whether it's free. The version that went out: "Live music this Friday from 9pm: a four-piece covers band playing 80s and 90s rock. Free entry, kitchen open until 9, back room open for tables of six or more. Reply here to reserve." Ask the assistant for the facts first (time, act, price, food, booking), and keep the exclamation marks for the band.

The review is where the calendar learns. After the band night in the fortnight below, the owner pasted the hourly figures beside a normal Friday and asked what differed. The illustrative answer: trade matched a normal Friday until 9pm, then ran at nearly double from 9.30 to 11; gin sold three times its usual Friday volume, while lager rose only by a third. Two changes followed for the next band night: the extra staff start at 9pm instead of 7pm, and the gin order doubles rather than the lager.

A quiz prompt that reduces bad questions:

Write a 10-question pub quiz round on [theme], for a mixed crowd.
Difficulty: 3 easy, 5 medium, 2 hard.
For each question give: the question, the answer, and one sentence on
where the fact can be checked.
Avoid questions whose answer has changed in the last five years, trick
questions, and anything with more than one defensible answer.

Here's the kind of mistake the last rule is there to catch. Asked for an astronomy round, an assistant produced "Which planet has the most moons? Answer: Jupiter." The answer has flipped between Jupiter and Saturn in recent years as astronomers confirmed new moons, and Saturn currently leads. A table that knows that will argue, and they'll be right. Drop questions like it rather than trying to word around them.

Before booking live music or extending hours for a big fixture, check your licence conditions. That's a question for your licence and local rules, not for an AI assistant.

The events calendar as the master input

Keep one sheet, updated weekly, that feeds both the rota and the stock order:

Date | Event (quiz / band / match / private hire / none)
Expected trade vs a normal night (-20% ... +80%)
Extra bar staff needed (hours) | Extra stock to order (by product)
Posted? (Y/N) | Actual sales vs normal (fill in afterwards)

Filled in for a typical fortnight, it might read:

Fri 10 | Band (covers act)      | +60% | +6 hrs | +2 kegs lager, +1 cider, +3 house gin | Y | +71%
Tue 14 | Quiz                   | +35% | +2 hrs (8-10pm) | +1 keg lager            | Y | +29%
Sat 18 | Match, evening kick-off | +45% | +4 hrs | +2 kegs lager, +1 keg stout   | Y | +52%
Sun 19 | Private hire (40 guests, back room) | +25% | +3 hrs | prosecco x 2 cases | Y | +18%

The last column is what makes the forecasts better over time. After three or four of each event type, you'll know that a quiz night adds a set amount to a Tuesday and a band adds much more to a Saturday, and you can order and staff to those numbers rather than to hope.

A month in a two-bar pub, with numbers

Say a pub with a main bar and a back room sells about $14,000 of drinks a week and runs a weekly quiz, a monthly band and the occasional big match.

Stock. The first monthly variance check shows spirits cost about 2 points higher than the owner expected. The AI's ranked list points first at over-pouring and unrecorded staff drinks. A week of watching and a simple staff-drinks tab on the till bring drinks cost down by about 1.5 points the following month: roughly $210 a week on $14,000 of sales. At that size, the owner decides a dedicated tool would pay back and trials one.

Rotas. The till forecast shows Tuesday before 7pm averages a handful of transactions an hour with two people on, and quiz night has a sharp rush from 8pm to 9.30pm that one bartender can't cover. The owner moves three hours a week from the early Tuesday shift to quiz night. Total hours stay the same; the quiz queue shortens.

Events. Drafting quiz rounds with AI and checking them takes about 40 minutes a week instead of two hours. The first week, two answers fail the check and are replaced.

Where pub AI plans go wrong

• Counting inconsistently. If one person weighs bottles and another eyeballs them, variance measures the counters, not the bar. It tends to show up as a product that swings: gin 2.5 bottles over one month and 2.3 under the next, with nothing changed behind the bar. Count part bottles in tenths, the same way, ideally by the same person, and fix the method before trusting the numbers. Why AI stock forecasts go wrong lists the other data problems that mislead forecasts.
• Letting one huge night skew the average. A cup final or a street party inflates a normal Saturday forecast. Mark it as an event and exclude it.
• Last year's fixture list. Sports calendars shift; a forecast built on last season's dates staffs the wrong nights.
• Unchecked quiz facts. Every answer, every week.
• Arithmetic in the chat. Let the spreadsheet or the software calculate; let AI explain.
• Treating a variance as theft. It usually isn't, and accusing staff on the strength of an AI's "likely causes" list is a quick way to lose a good team. Investigate first.

Further reads

• How to Forecast Covers and Cut Food Waste With AI — The same forecasting approach applied to a pub kitchen.
• How to Find Dead Stock and Slow Movers With AI — Spotting bottles and kegs that tie up cash on the shelf.
• How to Use AI to Set Reorder Points and Prevent Stockouts — Setting reorder points so busy nights don't run dry.
• AI Menu Engineering: Which Dishes to Promote, Reprice or Drop — Deciding which drinks and dishes to promote or reprice.
• How Much Does AI Cost a Small Restaurant Each Month? — What a hospitality AI stack costs each month in total.
• Best AI Social Media Tools for Small Businesses (2026) — Tools for scheduling the event posts you draft with AI.
• AI Inventory Forecasting for Small Businesses: How It Works — The four ingredients of a stock forecast, how lead time and safety stock turn it into a reorder point, the stockout trap, and five small-business examples.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: WISK pricing page; Deputy demand forecasting feature page; ChatGPT and Claude plan prices, checked September 2026.