Are Ghost Kitchens Still Worth It? A Cash-Flow Test You Run With AI

Sometimes: a ghost kitchen is worth it when what is left of each order after commission, food, packaging and refunds covers the kitchen's fixed costs at an orders-per-day figure you can reach. Divide monthly fixed costs by contribution per order to get break-even, then stress-test commission, volume and payout timing.

Published failure and profit figures for ghost kitchens come mostly from vendor blogs and unsourced reports, so I quote none of them. Your answer sits in your own commission plan, kitchen cost and order volume. In the worked sums below, the same menu needs about 48 orders a day from a rented unit and about 9 from a kitchen you already run.

Follow me on Instagram@sagnikteaches

One rule runs through the whole test: every number comes from a document you hold, never from an AI's memory. AI reads your payout statements, builds the sheet, writes the formulas and attacks your assumptions. It does not know your contract, your rent or your busiest hour.

Connect on LinkedInSagnik Bhattacharya

Rented unit, spare capacity or a new brand: three different tests

"Ghost kitchen" covers three setups, and the fixed-cost line differs in each. Mixing them up is a common reason a forecast looks better than it should.

Subscribe on YouTube@sagnikteaches

If your restaurant kitchen sits quiet for part of the day, test the first setup before anything else. It is the cheapest pilot you can run, and how an existing restaurant can test a ghost kitchen covers the practical side: menu, listing and kitchen flow. Here the question is money only. For costing each brand on its own menu, costing a brand after delivery commission goes line by line.

Ten inputs and the document each one comes from

The sheet needs ten numbers. Write the source beside each one, because a number with no source is a guess, and guesses survive into the final decision.

A wings-and-burgers brand in a rented unit, sum by sum

These figures are illustrative, not benchmarks. A restaurant owner is weighing a unit in a shared kitchen site for a delivery-only wings-and-burgers brand with an average subtotal of $24, on a 25% commission plan. Per order, the money runs out like this:

Contribution is what each order leaves to pay for the building and the people. The fixed costs for the month:

Over 30 open days that is $360 a day. Break-even is $360 divided by $7.45, or 48.3 orders a day, about 1,450 a month. The brand has to find 48 orders a day before the first dollar of profit appears.

The owner's forecast reaches 60 orders a day by month six, so the margin of safety is (60 - 48.3) divided by 60, or 19.5%. My own bar is a margin of at least 30%, which puts break-even at no more than 70% of the forecast, because nearly every shock in the sensitivity tables below pushes break-even up and almost none push it down. Here the bar means a forecast of 69 orders a day (48.3 divided by 0.7). The owner's forecast falls short of it.

From break-even to cash: the six-month ramp and the payout lag

Break-even says where the losses stop. Cash says how much must be in the bank to get there. Orders start low, so each month's result is that month's contribution minus the full $10,800 of fixed costs, with $4,000 of set-up costs entered on day one.

On this view the cash low point is about $20,800 in month four. Now add the payout lag. If the platform pays ten days after each sales day, the owner funds food, packaging and wages that far ahead of the money. Running the same ramp day by day with a ten-day lag moves the low point to about $28,800. A seven-day lag gives about $26,400 and a fourteen-day lag about $32,000. At 60 orders a day, a ten-day lag means the platform owes the brand about $9,400 at any moment (60 orders at a $15.60 payout, for ten days).

That is the verdict in two numbers. At month six the brand earns $2,610 a month and is still about $26,300 down on the ten-day view, so payback is roughly ten months further away even if volume holds at 60 a day. The unit passes break-even but fails a test built on cash. To lay the same logic out week by week with a bank-balance line, the 13-week cash flow forecast with AI help is the next step.

Sensitivity: which number moves break-even most

Break-even is a moving target, so test it two ways. The first grid keeps everything else at the base case and varies the two numbers an owner controls least: commission and average subtotal. Cells show break-even orders per day.

Read along the 25% row: a basket $4 smaller lifts break-even from 48 to about 60 orders a day. Read down the $24 column: moving from the 25% plan to the 30% plan adds about 9 orders a day of extra work for the same result. The second table changes one input at a time.

Commission, food cost and basket size are the big levers; packaging barely registers. Pricing is the lever you hold most directly. Raise app menu prices by 8% and a $24 subtotal becomes $25.92: commission, promotions and refunds rise with the price, but the food on the plate still costs $7.20, so contribution climbs to $8.70 and break-even falls from 48.3 to about 41.4 orders a day, provided customers keep ordering. Check how your sheet handles this. If food cost sits in a percentage cell, it rises with the price on paper and the sheet shows 44.3, understating the gain; for a price test, enter food cost in dollars per order. Whether takeaway menu prices should rise on delivery apps weighs that trade-off. The last row is not a forecast. It shows how thin the cushion becomes when several things slip together.

Pull your own rates out of a payout statement with AI

If your restaurant already trades on a delivery platform, your own statement beats any guess for commission, promotions and refunds. Export 30 days as a CSV, delete customer names, addresses and phone numbers, then paste the rows (or upload the file, if your assistant accepts CSV uploads) with this prompt.

Below is 30 days of order lines from my restaurant's delivery-app statement.
Customer names and addresses have been removed.
Columns: order_id, date, status, subtotal, merchant_funded_discount,
commission, marketing_fee, adjustment, payout.

Use only the numbers in the file. Do not assume any commission rate.
1. Count completed and cancelled orders separately.
2. For completed orders only, give the average subtotal, total commission
   and commission as a percentage of total subtotal.
3. Give merchant-funded discounts, marketing fees and adjustments as a
   percentage of total subtotal.
4. Give the average payout per completed order.
5. List rows that do not fit: zero subtotals, negative payouts, duplicate IDs.
Show the formula behind each figure so I can check it in a spreadsheet.

A realistic answer, with illustrative numbers, looks like this:

Completed orders: 412 (21 cancelled, excluded)
Average subtotal: $23.87 (total $9,834.44)
Commission: $2,305.51 = 23.4% of subtotal (25.0% of subtotal after discounts)
Merchant-funded discounts: $612.40 = 6.2% of subtotal
Marketing fees: $289.10 = 2.9% of subtotal
Adjustments: $198.35 = 2.0% of subtotal
Average payout per completed order: $15.60 (65.4% of subtotal)
Rows that do not fit: 9 rows with a $0.00 subtotal; 2 duplicate order IDs

Three checks before you trust it. First, the nine zero-subtotal rows sit inside the 412 and drag the average down. If they are test orders, exclude them (the average subtotal becomes $24.40); if they are real orders, ask the platform why they paid nothing. Second, the commission line: 23.4% of subtotal is the effective rate, because the plan's 25% is charged after $612.40 of merchant-funded discounts. Put 23.4% in the sheet's commission cell and 9.2% (discounts plus marketing fees) in the promotions cell, so the sheet matches the payout you really receive. Third, add up the platform's whole take: 32.6% of subtotal, close to the 33% (25% plus 8%) assumed in the worked example. With the measured split, break-even moves from 48.3 to 47.7 orders a day. The guess was fair, and now it has a source.

A prompt that does not mention the status column invites a classic error. Dividing $9,834.44 by all 433 rows, cancelled orders included, gives a $22.71 average instead of $23.87, and every figure built on it drifts. If the export is too long to paste, ask the assistant to write the spreadsheet formulas instead of calculating totals itself, and read how ChatGPT writes Excel formulas for the pattern that keeps them checkable.

Two prompts: build the sheet, then attack it

You can build the break-even sheet without AI, and it is worth seeing the cells once. Put the inputs in column B and the results below them:

A2  Average subtotal ($)             B2  24
A3  Commission (% of subtotal)       B3  25%
A4  Promotions + marketing (%)       B4  8%
A5  Food cost (%)                    B5  30%
A6  Packaging per order ($)          B6  0.95
A7  Refunds and remakes (%)          B7  2%
A8  Monthly fixed costs ($)          B8  10800
A9  Days open per month              B9  30
A11 Contribution per order           B11 =B2*(1-B3-B4-B5-B7)-B6
A12 Fixed cost per open day          B12 =B8/B9
A13 Break-even orders per day        B13 =B12/B11
A14 Break-even orders per month      B14 =B8/B11
A15 Forecast orders per day          B15 60
A16 Margin of safety                 B16 =(B15-B13)/B15

Then hand the layout to an assistant and ask it to extend it. The wording that matters most is the instruction that forbids invention:

Here is my break-even sheet layout and inputs [paste cells A2:B16].
Add: (1) a grid of break-even orders per day with commission 15%, 20%,
25%, 30% across and average subtotal $20, $24, $28, $32 down, using
formulas that reference my input cells; (2) a six-month ramp table with
orders per day [12, 24, 36, 46, 54, 60], monthly result and cumulative cash
after a $4,000 set-up cost; (3) a payout lag input in days that shifts
cash received. Use only my inputs. If a number is missing, ask me.

The second prompt is the one that earns its keep. Give the finished sheet to the assistant and ask it to play a sceptical accountant:

Act as a sceptical accountant reviewing a delivery-only food business.
Here are my inputs and results [paste]. List the five costs or assumptions
most likely to be missing or optimistic. For each, say which way it moves
break-even and which document would settle it. Do not invent figures;
where you need a number, ask me for it.

An illustrative reply, shortened:

1. Labour is a flat two cooks. If most orders arrive in a few evening hours,
   you may need a third person on busy nights. Raises fixed costs.
   Settle with a rota built from your busiest hour.
2. Food cost has no waste line. Small menus still bin prep.
   Settle with a week of waste logs.
3. No equipment allowance: tablet, label printer, cold storage.
   Settle with the unit agreement and quotes.
4. Refunds at 2% have no history for a new brand. Settle by using your
   restaurant's own statement as the floor.
5. The ramp is a straight line. A new listing has no reviews yet.
   Settle by tying promotion spend to the ramp.

Two of those five would change the numbers: the missing waste line and the thin equipment allowance. The rest are flags to watch. DoorDash's own tablet is listed at $6 a week after the trial, about $26 a month, which is small, but it shows how a missing line item sneaks in. An assistant can only find gaps; you decide which ones are real.

The same menu in a kitchen you already run: why the verdict flips

Run the identical menu through the owner's own restaurant kitchen and only the extras count. Suppose the brand needs about 20 hours a week of added prep ($1,600 a month with on-costs), $250 a month for equipment servicing and a second fryer lease, and $150 for packaging storage and software. Fixed costs are $2,000 a month, or $66.67 a day. Break-even is $66.67 divided by $7.45, which is 9 orders a day, about 269 a month.

A forecast of 15 orders a day gives a margin of safety of 40% and a monthly result of about $1,350 (15 orders a day over 30 days at $7.45 is $3,353, less $2,000). Set-up of $1,200 for photos, packaging stock and the listing is repaid in under a month at that volume. Orders will not start at 15, though. With a slower ramp (4, 8 and 12 orders a day, then 15) and a ten-day payout lag, the cash low point is about $4,200 and cumulative cash is back to roughly zero by month six. Even the five-things-go-wrong row from the sensitivity table, which lifted the rented unit to 95 orders a day, only lifts this version to 17.6 orders a day. Same menu, same commission, opposite verdict, because the rent and the cooks were already being paid for.

The test is only honest if the extras really are extra. Check four things before believing the 9:

• Capacity at your busiest hour. Delivery tickets and dine-in tickets compete for the same pass at 7pm. If dine-in slows, count that loss.
• Cannibalisation. If most delivery orders come from customers who used to order your main menu on the same app, the sale is not new. Compare your weekly app orders before and during the pilot.
• Staff goodwill. Added prep hours that land on your best cook cost more than the wage line suggests.
• Registration. Ask your food-safety adviser whether a new trading name or listing needs registering, and check your insurance covers the extra brand.

A smaller brand changes the sums again. Take a dessert-only brand from the pastry section of the same kitchen: subtotal $14, food cost 24%, packaging $0.70, commission 25%, promotions 8% and refunds 2%. That leaves $5.04 an order. Extra fixed costs of $1,100 a month need 7.3 orders a day. Small baskets need small overheads, because commission and packaging bite harder on a $14 order than on a $24 one.

Pickup and your own ordering page: the cheaper channels in the sheet

Two channels carry far less commission, and the sheet should show them separately. DoorDash lists pickup at 6%, so a $24 pickup order costs $1.44 in commission against $6.00 on the 25% plan. Orders on your own ordering page skip marketplace commission but pay card fees. Shopify's pricing page shows 2.9% plus 30 cents online on its Basic plan, which costs $39 a month or $29 on yearly billing; that is $1.00 on a $24 order. DoorDash describes its own online ordering as commission-free but subject to a payment processing fee.

Delivery on your own page is different, because you then pay the courier; compare it with the full app cost, not only the commission. Pickup is the cleaner win. If 4 orders a day moved from app delivery to collection from your own page, the commission saving is about $5 each, or $600 a month. In the sheet, add a channel column to the order forecast and give each channel its own commission and card-fee line, then ask the assistant to split your ramp by channel. What a takeaway actually pays on delivery apps compared with AI ordering goes further into the trade.

Where an AI-built forecast goes wrong

These are the failures to expect when an assistant builds or checks the sheet, and how each one looks on the page.

• It fills gaps from memory. Ask for a ghost-kitchen forecast without giving a commission and the assistant chooses a figure that sounds typical. It sits in the sheet looking like any other input. End every prompt with: use only the figures I give you; if one is missing, ask.
• It divides by the wrong days. Monthly break-even is 1,450 orders whatever the calendar, but "per day" depends on days open. If you open 26 days a month, the same 1,450 orders is 55.8 per open day, not 48.3. A draft that says 48 a day and then plans closed days sets a target the kitchen cannot hit.
• It averages away the peaks. Sixty orders a day sounds calm until three-quarters arrive in a four-hour evening window: 45 orders in 240 minutes is one every 5.3 minutes. Ask the assistant to build the day by hour from your till data and flag any hour where the cooks cannot keep up.
• It slips on long columns. Language models can miscalculate on big tables. Ask for formulas, not totals, recompute three orders by hand, and compare the assistant's total payout with the statement's own payout line. A gap of more than a few cents means a row was dropped or counted twice.
• It draws straight lines. Give it a month-one and a month-six number and it joins them. A new listing has no reviews, so the early weeks can be slow. Describe the shape you expect, then cap the curve at what your existing restaurant achieves on the app.

A go/no-go rule and a stop-loss, written before launch

Decide the rule before you have an emotional stake in the answer. Here it is filled in for the rented unit above:

Go only if all four hold:
1. Margin of safety is at least 30% (break-even at most 70% of forecast).
   Here: 60 a day against 69 needed.                       FAILS
2. Lowest cash point, including payout lag, is money I can lose.
   Here: about $28,800.                                    Owner's call
3. Payback after month six is under 12 months.
   Here: about 10, if volume holds at 60 a day.            Passes, barely
4. Break-even with commission 5 points higher (57.6 a day) is a volume
   two cooks can produce at the busiest hour.              Check the rota
Stop if: at week 8, orders a day are below 70% of the ramp
(under 17 a day against 24 planned), or cash falls below $X.

As drawn, the rented unit is a no-go, because it fails the first rule. The levers are in the sensitivity tables: a lower unit fee, one cook plus a part-timer, a cheaper commission plan, or a higher basket. The existing-kitchen version of the same brand passes every rule: 15 a day against 13 needed, a low cash point of about $4,200, and cumulative cash back to roughly zero by month six. That is the pilot I would run first.

Write the exit terms into the same document. A unit with a break clause and a platform agreement you can leave on short notice make the stop-loss real; a twelve-month lock-in turns a pilot into a lease. Once the brand is live, replace every forecast input with the figure from the latest statement each month, and let the sheet tell you when the picture changes.

Questions owners ask before they sign a kitchen agreement

Can I paste a payout statement into ChatGPT, Claude or Gemini?

Delete customer names, addresses and phone numbers first. The test only needs amounts, dates, statuses and order IDs. On a consumer plan, switch off the model-training setting in privacy settings; business plans do not train on your content by default. Treat the statement as commercial data, and check your platform agreement for a confidentiality clause before you share it with any tool.

How long should a ghost-kitchen pilot run before I decide?

Long enough to test the ramp in your sheet, which usually means eight to twelve weeks. A new listing starts with no reviews, so week two tells you little. Write the stop-loss before launch, such as orders per day below 70% of your ramp at week eight, and judge on contribution per order, not on ratings alone.

Does the test work for a catering or meal-prep delivery brand?

Yes, with different inputs. Use your average order value, replace marketplace commission with card fees if customers order direct, and count the days you actually cook. The formula stays the same: monthly fixed costs divided by days worked, divided by contribution per order, gives break-even orders per day.

What if the platform changes its commission after I launch?

Treat commission as a living input. Re-run the sheet when your plan changes and keep the old and new break-even side by side. Check your agreement for notice periods, whether you can switch plans, and what leaving costs, because the exit terms belong in your stop-loss rule.

Further reads

• Cloud Kitchen Inventory: Forecast Prep by Delivery Platform With AI — Forecast prep by platform so your ramp rests on evidence.
• How to Check Your Margins Product by Product With AI — Check food cost and margin item by item before you set prices.
• Break-Even Analysis With AI: A Worked Example for a New Product — The same break-even method on a different kind of product.
• AI for a Cloud Kitchen: Keep Every Brand's Menu Consistent on Apps — Keep every brand's menu and prices consistent once you launch.
• AI in Hospitality Management: Catch Supplier Price Creep on Invoices — Catch supplier price rises before they erode your food cost.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: DoorDash merchant pricing and promotions pages and Shopify pricing page (checked October 2026). All other figures are illustrative worked sums, not benchmarks.