AI project proposals tend to arrive with general promises like "save 50% of your team's time" or "double your productivity." Nobody knows where those numbers came from, and you have no way to tell whether they apply to your company. So one of two things happens. You turn down a project that would have helped, or you pay for one that saves almost nothing.
You can work out the expected return yourself before you start, using your own company's numbers. The calculation will not be exact, but it is good enough to decide.
How do you measure your current baseline?
Start with the task you want to automate and track two numbers for a week or two: how often it happens, and how many minutes it takes on average. Multiply them to get monthly hours, then multiply that by the hourly cost of the employee who does it.
Hourly cost includes salary, social security and benefits, divided by the hours actually worked in a month. Remember errors too. If the manual task causes mistakes that cost money, such as an invoice paid twice or a delayed order, record how often that happened and what it cost.
- Times per monthFrom a log or a week of tracking
- Minutes each timeA realistic average, interruptions included
- Hourly costSalary plus related costs
- Cost of errorsWhat manual mistakes really cost
- Current monthly costThe number you will compare against
What costs should you include?
The visible cost is building the solution. The costs people forget usually come after launch, and they can change the whole result.
| Cost type | Examples | When you pay |
|---|---|---|
| Build | Analysis, development, system integration, testing | Once |
| Running | Model usage by volume, hosting, tool subscriptions | Monthly |
| Human review | Employee time spent reviewing and correcting results | Monthly |
| Maintenance | Updating the solution when systems or documents change | Periodically |
| Internal preparation | Organizing data and documents, training the team | At the start |
Human review is the line item people overlook most. A solution that reads invoices but needs an employee to check a third of them only saves part of the time, so calculate it on that basis.
What is the simple ROI formula for AI?
Monthly savings equal the current monthly cost, minus the monthly running cost, minus the cost of human review. Then divide the build cost by the monthly savings to get the number of months until you earn back what you paid.
Here is a hypothetical example to show the math. A task happens 400 times a month and takes 6 minutes each time, which is 40 hours. If 10 hours of review remain after automation, you save 30 hours a month. Multiply that by your hourly cost, subtract the running cost, and compare the result with the build cost. Swap in your own numbers for these.
Run three scenarios: optimistic, expected and conservative, changing the share of cases that need review. If the project pays off in the conservative scenario, the decision is easy.
What benefits don't show up in the formula?
Some benefits are real but hard to put a number on. Customers get faster replies outside working hours. There are fewer errors in the data your decisions rest on. And work keeps going when the one employee who knows the task is away. Write these down next to your calculation, but do not let the decision rest on them alone. If a project only makes sense because of benefits you cannot measure, look at it again.
When is an AI project not worth it?
- When the task is rare, so the savings never add up to the build cost.
- When almost every result needs correcting, and the work turns into exhausting review.
- When the data or documents are a mess, so most of the effort goes into organizing them first.
- When nobody owns the project after launch and follows its results.
In the third case, the cleanup itself may be the first project. And sometimes a simple automation tool with no AI is enough, as in workflow automation with n8n, Make and Zapier.
How do you compare two or more AI proposals?
Ask every vendor to fill in the same table: build cost, expected monthly running cost at your current volume and at double it, the expected share of cases that need review, and who handles maintenance. The proposal that is cheapest to build can be the most expensive after a year if it costs a lot to run or needs more review. Ask about ownership as well. If you change vendors, do the instructions, data and code stay with you?
How do you test the calculation with a small pilot?
Run a limited pilot for a few weeks on part of the real volume, before the full build. During the pilot, track the same numbers you measured before you started, plus the share of cases that needed correcting. That share does more than any other figure to correct your first estimate. Then decide whether to continue, adjust the scope, or stop having spent very little.
We explain how to pick the first task in AI for your business: where to start. We also help companies work out feasibility and design the pilot before building, as part of our AI and automation solutions service. Send us a description of the task and its rough numbers and we will help you work out whether it is worth automating.
Questions our partners ask
How long does an AI project take to pay for itself?
It varies a lot with the size of the task, how often it happens and the build cost. The calculation in this article gives you an estimate for your own company, which is more accurate than any general average.
Should I count time saved as money saved?
Only if that time will go to useful work or replace a hire. If the employee ends up with no extra work, the real financial saving is smaller than the numbers suggest.
What if I cannot measure how things work today?
Measure one week roughly: how many times the task happens and how long each one takes. An estimate built on a real week is far better than a guess.
Written by the Visual Innovate team, the people who build, host and follow up systems, websites and apps for our clients in Jordan, Saudi Arabia, Oman and Egypt.
