Is an AI tool worth it for your hotel? Building the business case from your own numbers

AI ROI is what an AI tool returns against everything it costs, measured on your own task volumes. Build it from tasks: how often each happens, minutes before and after, what an error costs today, and every cost line, including the time staff spend checking output. A tool whose value you cannot measure is a decision, not a calculation.

A vendor deck arrives with a percentage saved and a payback in months. Neither number was calculated on your hotel: not your task volumes, not your wage costs, not your cost of a mistake. The case has to be rebuilt from your own figures, and it fits on one page. After reading, you will know which tasks to count, which cost lines to fill in, how to run the arithmetic, and how to judge a pilot honestly.

Why can't you copy another hotel's AI savings?#

Because a saving depends on four things that differ hotel by hotel: how often the task happens, how long it takes now, what the person's time costs, and what a mistake costs today. Change any one of the four and the same tool moves from obvious to pointless.

A percentage of hours saved is not money. It becomes money only when the hours are paid overtime you stop paying, work you would otherwise hire for, or time redeployed to something you were going to do anyway. Task mix differs too. A city hotel with heavy corporate correspondence and a resort with heavy in-person guest requests at the desk are not solving the same problem, even if both buy a writing tool.

Language and wage mix differ. A tool that drafts English replies well may be weaker in the languages your guests actually write in, and the review minutes rise exactly where the saving was supposed to be. Error cost differs as well: a wrong spa slot is recoverable in a phone call; a rate quoted in writing to an agency, or a fiscal document sent with the wrong amount, is not. Volume differs last. The same per-seat price is cheap at two hundred tasks a week and expensive at twenty.

The test to apply is simple. Ask any vendor for the task, the baseline and the measurement method behind a claim. If the deck shows only a percentage, it is marketing, not a case. For the questions worth asking before any purchase, see our guide to questions to ask an AI vendor before buying for your hotel.

Which tasks should you count first?#

Count the tasks that are frequent, produce a written output, and are already checked by a person: there the minutes are measurable and a mistake has a visible cost. This is also why back-office work usually comes before guest-facing work, as we explain in where to start with AI in a hotel: back office before guests.

Build the list from one ordinary week, not from a strategy document. Watch where the same work repeats: reservation replies, rate quotes to agencies, the morning report, night audit exceptions, repeated guest questions, review responses. For each task record four numbers: times per week, minutes per occurrence today, who does it and their loaded hourly cost (wage plus employer costs, divided by paid hours), and the cost of one error today in rework minutes, refunds or lost bookings.

Measure minutes on five real cases with timestamps or a stopwatch. Memory rounds to comfortable numbers and always in the direction of the tool. Prefer tasks with a written output, because you can compare before and after, and a reviewer can see what the tool changed. Skip tasks that happen twice a month: even a large percentage there is small money, and the setup cost dominates.

Count the review step in the after-minutes from the first day. An AI draft that takes two minutes and needs four minutes of correction is a loss, not a gain.

What does an AI tool really cost?#

The subscription is usually the smallest line. Setup, data preparation, training and the daily minutes spent checking output are the lines that decide the case.

The subscription or usage fee may be per seat, per task or per unit of use. Ask what the bill does when volume doubles, because the pilot volume is rarely the steady-state volume. Setup and integration covers connecting the tool to the property management system, the reservations mailbox or the channel manager; ask who pays the integrator's hours and who owns the connection afterwards. Data preparation, meaning clean rate codes, room types, policy text and standard answers, is staff time even when nobody invoices it, and it is usually larger than the subscription in month one.

Training and change needs more than one session. The team needs a written rule for what the tool may send on its own and what a person must approve before it leaves the building. Review time is minutes per output multiplied by outputs per week, at the loaded hourly cost, and it is the line most often missing from a vendor's arithmetic. Ownership means one named person who watches quality, keeps the standard answers current and decides when an output is not good enough to use. Exit means the contract term, the notice period, and what happens to your data and your configured content if you leave.

For how vendors structure these prices, see how AI tools are priced: seats, usage and what drives the bill.

What does the arithmetic look like on one real task?#

Take one task, measure it before and after, convert minutes to money with your loaded hourly cost, subtract every cost line, and see whether anything is left.

Before: 60 reservation emails a week at 6 minutes each is 360 minutes, or 6 hours a week. After: 2 minutes to prompt and read the draft, plus 1.5 minutes to check and correct it, is 3.5 minutes per email. That is 210 minutes, or 3.5 hours a week, so 2.5 hours are saved. Convert: at a made-up loaded cost of 100 per hour, the saving is 250 a week, about 1,080 a month. Subtract: a subscription of 300 a month leaves about 780 a month against a one-time setup of 1,200, which is recovered in roughly two months.

Then stress it. If the saving is half, monthly value falls to about 540 and net value to about 240. If the subscription doubles instead, net value is about 480. Both stay positive, but the margin is thin enough that one quiet month changes the picture.

Write the result as a sentence you can defend: "this task, at this volume, saves this many minutes a week, which is this much money after these costs." If the sentence needs a benefit you have not measured, the case is not finished.

How do you find the break-even point and the payback period?#

Payback is the one-time cost divided by the monthly net value. Break-even is the task volume at which monthly value equals monthly cost.

Monthly net value equals minutes saved per task, divided by 60, multiplied by tasks per month, multiplied by the loaded hourly cost, minus the monthly recurring cost. Payback in months equals the one-time cost divided by the monthly net value. If the net value is zero or negative, there is no payback, whatever the deck says. Break-even volume equals the monthly recurring cost divided by the money value of one task; below that volume the tool loses money every month it runs.

Run three versions of the same sheet: your measured numbers, half the saving, and double the cost. If the decision flips between them, the case rests on an assumption you have not tested. Do not add a productivity percentage you cannot see in the payroll or the roster. If nobody's hours changed and no hire was avoided, the saving is capacity, not cash. Finally, compare the payback period with the contract's notice period. A payback longer than the time it takes to exit is a warning, not a rounding error.

Which benefits are real but hard to count?#

Some gains are genuine and still do not belong in the payback number: list them separately and say what would have to be true for them to count.

Speed at peak is one. A reply in ten minutes instead of two hours can win a booking, but you cannot attribute that booking without a controlled comparison, so keep it out of the arithmetic. Consistency is another: the same answer to the same question on every shift, which matters for policy, for complaints and for training new staff. Coverage means a correct answer at two in the morning; count it only if the alternative today is a lost request, not a message nobody would have answered either way.

Staff retention and training time are real, but they need a before-and-after measure, not a feeling in the second week. Traceability, being able to see who changed what and when, has value in disputes and in staff reviews, and it is usually a by-product rather than a reason to buy. Keep a second column called unpriced. Decide explicitly whether the priced case stands without it, and revisit the list in three months.

How do you decide to continue, change or stop after a pilot?#

Set the thresholds before the pilot starts, measure the same task the same way at the end, and let the numbers choose between continuing, changing the scope and stopping.

Write the baseline, the target, the measurement method and the stop rule on one page before day one, and have the person who will do the measuring sign it. Measure the same five to ten real cases at the end as at the start, with the same person timing, and include the review minutes. A pilot that measures only the drafting time flatters the tool; the honest measure is the whole task, from request to approved output.

Three outcomes follow. Continue when value is at or above target. Change when value is positive but below target and one variable, usually volume or scope, explains the gap. Stop when value is negative at your real volume or the review time eats the gain. Check the contract before the pilot ends, because the notice period decides how cheaply you can act on a stop decision. Record the decision and the numbers next to it, so the second purchase starts from a real baseline instead of a deck. The full procedure is in our guide to running an AI pilot in a hotel: scope, baseline and a clear stop rule.

What to do next#

Pick one task this week, time five real cases, fill in the cost checklist, and run the three formulas on your own numbers. If the answer survives half the saving and double the cost, run a short pilot with the baseline and the stop rule written down in advance. If it does not survive, you have saved the price of the tool and the months spent discovering it.