The subscription is rarely the big cost. Here are the five real costs of automating a process with AI and how to check whether it pays for itself.
It is the first question almost every manager asks: how much does it cost? And the answer they usually find online is a 20 dollar subscription, which has nothing to do with the cost of making a process run for real inside a company.
You will not find a magic number here, because it depends on the process. You will find the costs that do exist, the ones people forget, and a calculation you can do yourself before asking for a quote.
Why is there no single price?
Because "automating with AI" can mean anything from an assistant set up with clear instructions to a system that reads documents, writes into your accounting software and answers customers on WhatsApp. The first is an afternoon of work; the second is a project with testing, permissions and maintenance.
So be wary of anyone who gives you a price without having seen the process. And be just as wary of anyone who promises it is all free because "AI already exists".
What are the real costs of an AI project?
Five costs show up in almost every small project. Most people only count the second one.
- Process design. Writing down the steps, the odd cases and what a person must decide. It is work for you and your team, and it is the most underestimated.
- Tool usage. Subscriptions, pay-as-you-go API usage (billed per token, meaning the amount of text going in and out) and, if you use WhatsApp, the messages Meta charges for.
- Integration. Connecting the AI to your files, billing software or CRM. This is where an afternoon pilot turns into weeks.
- Human review. Someone has to check what the AI produces, at least at first. That time counts as a cost, not as savings.
- Maintenance and data. Vendors change models and prices, your document formats change, and customer personal data needs care under Law 1581.
How much does the AI usage itself cost?
Less than most people expect for typical office processes, but it helps to understand how it is billed. On model APIs such as Claude, the price is per million tokens, with different rates for input and output, and it varies by model. Anthropic's documentation, for example, lists a 50% discount on input and output when you process non-urgent work in batches, and much lower prices for text that repeats across requests thanks to caching. Rates change often: check the provider's page the day you decide.
WhatsApp works differently. According to Meta's documentation, since 1 July 2025 you are charged per delivered message, with rates that depend on the message category and the recipient's country. Inside the 24-hour window that opens when a customer writes to you, you can reply with service messages, and as of October 2026 Meta states that each number includes 1,000 free service messages per month before charges start. Template messages outside that window do cost money. Check the current rate card before projecting your volume.
How do I work out whether automation pays for itself?
You need four numbers, and the first three you can measure this week:
- Hours the process takes today per month.
- Hours it would take afterwards, counting human review.
- Cost of one hour of that person, including salary and employer contributions. Ask your accountant for the figure instead of using salary alone.
- Monthly running cost of the system (tools, API, messages, maintenance) and the one-off cost to build it.
The math: (hours saved × cost per hour) minus monthly running cost gives the net monthly saving. The build cost divided by that net saving tells you how many months it takes to pay back.
An example with made-up numbers
Note: these figures are hypothetical and only show the calculation. Use your own.
Say a small company moves 400 invoices a month into its spreadsheet and each one takes 4 minutes by hand: about 27 hours. With automatic extraction and 1 minute of review per invoice, that drops to about 7 hours. That saves 20 hours a month. If an hour of that person costs 20,000 pesos, the gross saving is 400,000 pesos a month. If running the system costs 100,000 a month, the net saving is 300,000. And if building it cost 3,000,000, it pays back in ten months.
Look at what the example teaches: at this volume the project pays off, but not right away. At 100 invoices a month it would never pay back. At 2,000 invoices it would pay back in weeks. Volume decides more than technology does.
Which projects tend to get expensive?
- Processes nobody has written down. If everyone does it differently, it has to be standardized first.
- Lots of exceptions. If 40% of cases are unusual and need a person, the real saving is much smaller than expected.
- Integrations with closed software that offers no API or export.
- Trying to automate everything from day one. A pilot covering 70% of cases with human review beats a perfect system that never ships.
How do I avoid overspending?
- Measure before you buy. One week of logging times is worth more than any quote.
- Start with a small pilot and set in advance the number at which you will call it a success or a failure.
- Use what you already pay for (Google Workspace, Microsoft 365) before buying another platform.
- Ask whoever quotes to break out design, integration, operation and maintenance separately. If it all comes as one number, you cannot compare.
- Review your accounts every quarter. Models and rates change, and what used to be expensive sometimes costs less now.
This is general information. It does not replace your accountant for labor cost calculations, or a lawyer if the project touches personal data.