BYOK for AI emails: what your own API key costs and when it pays off

BOTrento Team10 October 2026 · Updated 11 October 20268 min read

What BYOK means for AI emails and what you actually pay for

With a tool that helps you write replies to enquiries, follow-ups or cover emails for quotes, you pay for two quite different things:

• The service: process setup, connection to your mailbox and records, supervision, error handling and reports. You pay this to the vendor.
• Model usage: the text the language model reads (the enquiry, the thread history, supporting material) and generates (the draft email). It is billed in tokens.

Without BYOK, both parts sit in one price and you do not know how much of it is AI. With BYOK, you open your own account with a model provider (for example Anthropic, OpenAI or Google), generate an API key and enter it in the tool. You then pay usage directly to the provider, and the service vendor adds nothing on top.

We describe the principle in more detail in the article on your own AI API keys in business and how to manage them safely.

Method: how to estimate token costs per email

Nobody can promise you a specific price per token in advance. It varies by provider and model, and price lists change. The calculation method is always the same, though, and you can do it in an afternoon.

1. Pick one case type. For example, "reply to an equipment rental enquiry". Do not try to count all your emails at once.
2. Take a few real recent enquiries. Copy what the model would have to read: the enquiry text, the earlier thread if relevant, your price list or service description, and tone instructions.
3. Measure the input. Convert text length into tokens. Most providers offer a counter, or the usage overview shows the token count after a test call.
4. Estimate the output. How long should the draft be? A short confirmation is different from a reply with follow-up questions and a provisional price. If you want several draft variants, each is counted separately.
5. Multiply by the model's current price list. Input and output tokens are often priced differently, so calculate them separately.
6. Add rewrites. When you have a draft redone, the model processes the context again.
7. Multiply by monthly case volume. Use a typical month and your busy season.

Calculation template

You will usually find that cost is driven not by the number of emails but by the length of the context. A model that receives your whole price list and a long thread every time uses far more than one given only what it needs for the reply. For setting limits and a monthly review, see our article on estimating and controlling AI costs.

BYOK or a flat fee with AI included: when each pays off

Both options make sense. It depends on what matters more to you: simplicity, or visibility and control.

BYOK pays off when you

• handle a regular volume of enquiries and want AI costs separate from the service price,
• need to know which provider processes your texts, for instance because of internal rules or client requirements,
• want to be able to swap the model without changing vendor,
• already have an account with a provider.

A flat fee makes more sense when

• you use AI only occasionally and managing another account would cost more time than it saves,
• nobody in the company is responsible for billing and limits at the provider.

For the full picture of what a price is made of, including your team's time, see how much an AI assistant costs a business.

Control over data: what BYOK solves and what it does not

BYOK is mainly about who processes your texts and on whose account. It does not mean the data stays inside your company.

• The text needed for the task (enquiry, context, instructions) is sent through the API to the model provider you chose.
• Under their terms, commercial API providers do not use API data to train their models. Still, read the terms of the provider you choose.
• If a vendor stores data on an EU server and calls an external AI API, that is not a closed computation. The text to be processed goes to the model provider.

In practice: send the model only what it really needs for the reply. It does not need the client's whole history, internal notes or attachments unrelated to the answer. That also saves tokens. For specific situations involving personal data, consult a lawyer or your data protection officer.

Checklist: setting up your own API key safely for emails

• The provider account belongs to the company, not to one salesperson personally.
• The email process has its own key, separate from other projects, so usage is visible on its own.
• You have a monthly spending limit or alert set at the provider.
• You know who created the key, where it is stored and who may remove it.
• You never send the key by email or in chat. You enter it directly in the tool.
• The tool stores the key encrypted and never displays it in full.
• You can revoke the key at any time and issue a new one.
• Once a month, you compare usage with the number of cases processed.
• No draft email leaves without a person checking it.

For a step-by-step guide to entering a key, see the help page on API key setup.

How BOTrento handles it

BOTrento Chief of Staff is a managed process for enquiries, quotes and follow-up sales steps. You pay for AI through your own keys:

• BYOK with no markup. You enter your own keys (Anthropic Claude, OpenAI, Google Gemini, Groq, ElevenLabs) and pay usage directly to the provider. The plan covers the service: setup, integration, supervision, escalations and reports.
• Safe keys. Keys are encrypted, you see only the last 4 characters, and you can remove them at any time.
• A person approves. The process prepares draft emails and quotes, but nothing is sent without human approval.
• Data. We store it on servers in Germany (EU). The language model processes only the text needed for the task, through the provider you chose.
• Predictable capacity. The plan is counted in sales cases. You get a warning at 80% of capacity, and any increase happens only with your consent.

You start with a 14-day free trial: no payment card and no API key of your own. The 14 days begin with your first successful upload of customer data (import or mailbox connection), and you have 7 days to prepare. During the trial you see real token usage on your own data, not a spreadsheet estimate. In our internal pilot, the process handled 35 enquiries, 18 emails, 24 quotes and 15 actions over 7 days. The process demo shows what that looks like. Plans are on the pricing page.