AI costs in your business: how to estimate them in advance and keep them under control

BOTrento Team7 October 2026 · Updated 13 October 20269 min read

What AI costs are made of

When a company says “AI costs us this much”, it usually mixes three different items. For both estimating and control, keep them apart.

The service is predictable and can be agreed as a flat fee. Model usage is variable but easy to estimate, because it grows with the volume of work you already know from your own inbox. People's time is often left out of budgets, yet it decides whether the rollout pays off.

Why it helps to pay the model directly

If you pay for model usage through your own API key directly at the provider, you see it with no markup and separately from the service. You know what the text processing itself costs and what the work around it costs. The principle is covered in more detail in Your own AI API keys (BYOK) in business.

How to estimate AI costs in advance: a 6-step method

The estimate does not need to be accurate to the euro. It only needs to be based on your own data and include a buffer.

1. Count sales cases per month. Take the last two or three months and add up the enquiries that were actually answered. Note seasonal peaks separately – event companies and equipment rental firms have pronounced ones.
2. Describe the steps per case. For example: sorting the enquiry, drafting a reply, drafting a quote, one or two follow-ups. Each step is a separate model call.
3. Estimate the length of texts. A short web-form enquiry and a long thread with attachments and a technical specification differ several times over. Split cases into “short” and “long” and estimate the ratio.
4. Choose the model by task. Sorting enquiries can be handled by a smaller, cheaper model; a quote draft with prices and terms may call for a more capable one. You will find the provider's price list in its admin console.
5. Run a trial measurement. Let the process run for a week on real data and check actual usage in the provider's console. Divide it by the number of cases to get the average cost per case.
6. Add a buffer and set a ceiling. Multiply the average by the expected number of cases, add a buffer for peaks and regenerated drafts, and set that amount as the monthly limit.

To give a sense of volume: in our internal pilot, 35 enquiries, 18 emails, 24 quotes and 15 actions went through the process in 7 days. Even one week of that kind of activity gives you enough data not to build your estimate on gut feeling.

A simple formula

Monthly model usage ≈ number of cases × average usage per case (from the trial week) + buffer
Total cost ≈ service + model usage + people's time spent on review

Limits and alerts: avoiding unpleasant surprises

An unexpected bill rarely comes from normal operation. It comes from a mistake: a loop that keeps regenerating, bulk processing of old emails, or a key used by several tools at once. A few simple measures help.

At the model provider

• One key per process. When enquiry handling has its own key, you know exactly what it costs, and if something goes wrong you can revoke it without affecting the rest of the company.
• Monthly limit or budget alert. Check in your provider's console which limits and alerts it offers, and set them based on the estimate from the previous section.
• Access for the responsible person only. One person creates and manages the key, not everyone who wants to experiment. How to create and enter keys is described in the guide to your own API keys.

On the service side

• Capacity in cases, not vague units. The number of sales cases per month is a figure that even an owner with no technical background understands.
• An alert before capacity runs out. You learn in good time that you are nearing the ceiling and decide for yourself.
• Increases only with your consent. No automatic switch to a higher plan and no surcharge you only discover on the invoice.

Monthly AI cost review template

One screen once a month is enough. Its purpose is not accounting but answering one question: are we paying for what we expected, and are we getting what we wanted?

The last line matters: costs alone say nothing until you put them next to what improved in the process. The easiest place to start is enquiry response time – how to measure it is shown in Enquiry response time: how to measure it and why to start there.

Hidden costs that get forgotten

• Tuning the instructions. In the first weeks, instructions, templates and rules get adjusted. If your own person does this by trial and error, you pay with their time and with repeated model calls.
• Reviewing outputs. A draft that has to be half rewritten saves little. Track the share of drafts approved without edits.
• Outdated price lists and source material. When the model works with old prices, every quote needs manual correction. A well-organised price list saves both time and usage – see A sales quote that doesn't come back with questions.
• Fragmented tools. Several individual subscriptions, each on a different account, are hard to add up and even harder to monitor.
• Processing unnecessary data. Sending whole thread histories and newsletters to the model costs money and adds nothing. Only the text needed for the task belongs in the model. For the data protection side, see the GDPR and AI checklist.

How BOTrento handles AI costs

BOTrento Chief of Staff is a managed process for enquiries, quotes and other sales steps. Costs in it are split exactly along the logic above:

• The plan is the service – setup, integration, supervision, escalation and reports. Capacity is expressed in sales cases per month; one case is one sales case actively analysed in the month, with at most five drafts. You can see the tiers on the pricing page.
• You pay for model usage directly through your own API keys (Anthropic Claude, OpenAI, Google Gemini, Groq, ElevenLabs), with no markup. Keys are encrypted, only the last four characters are visible, and you can remove them at any time.
• An alert at 80% of capacity and increases only with your consent, with no unexpected surcharge.
• Nothing is sent without human approval. Email and quote drafts are approved by your own person.
• We store data on servers in Germany (EU). The text needed for the task is processed by the model provider you choose with your own key; commercial API providers do not use the data for training. However, your own server in the EU that calls an external AI API is not a closed computation.

You start with a 14-day free trial, with no payment card and no API key of your own. The 14 days begin with the first successful upload of your data (an import or a mailbox connection), and you get 7 days to prepare. During the trial you get exactly the data you need for an estimate: the number of cases, actual model usage and agreed process KPIs. What such a process looks like in practice is shown in the process demo on enquiries.