AI integration
Scope
What is included
Every project has a different scope; the quote lists each included item one by one.
Last updated:
- Safe access for an AI model to your business data (MCP server, API)
- Per-tool permission limits and a read-only versus write distinction
- Logging: which tool was called, when and with which parameters
- Defining and enforcing the steps that need human approval
- Cost and usage tracking
Preparation
What we ask from you
- An example of the job you want to automate
- Access to the data sources to be used
- A decision on which step must always need human approval
Process
With your approval at every step.
- 01
Choosing the example job
We review one job you want to automate, together with how it is done today.
- 02
Access and permission design
We write which data and tools the model reaches with which permission, and where approval is needed.
- 03
Development and trial
We write the MCP server or API and try it with sample data and in real use.
- 04
Logging and monitoring
We turn on logging and cost tracking and hand over usage and change notes.
Details
About the service
The problem
When an AI model starts working with your business data, two questions come up: what can the model access, and what happens if it does something wrong? Often the answer is that the model was given a broad access key and nobody recorded who did what and when. We focus on setting this access up narrow, logged and reversible.
What MCP is
The Model Context Protocol (MCP) is an open protocol that lets AI clients call external tools. Each tool is defined by a name, an input schema and a description; the client can only call defined tools. This structure makes what a model can do visible in the tool list and makes it possible to grant each tool a separate permission.
What we set up
- An MCP server or API: An interface with clear limits through which the model reaches your data or tools.
- Permission limits: Reading and writing are handled separately; tools that write run with a separate permission and, where needed, extra approval.
- Human approval: Steps that are hard to undo or that touch the outside world (such as payments, publishing or sending email) do not run without human approval.
- Logging: Which tool was called, when, with which parameters and what the result was; the log is kept without copying secret data.
- Cost and usage tracking: The number of model calls and their approximate cost become visible.
We start with a small job
Instead of a large system that can do everything, we start with a small system that does one job safely. As the job works, new tools are added; for each new tool the permission and approval rule is written again. That way the boundary of the system always stays known and written down.
Where it makes sense
An AI model is usually useful for repeated jobs with clear boundaries: classifying incoming requests, summarizing a long document, searching a record system, or preparing a draft and submitting it for approval. We first write down how the job is done today with you, then decide at which step the model comes in and which step stays with you.
Limits
AI models can make mistakes; that is why tying critical steps to human approval is part of the system, not an optional extra. We do not promise a specific accuracy rate or savings. The terms of the provider whose model is used and where the data is sent are an important item of the scope; we put that in writing at the start.
What determines time and price
Price and time depend on the number of systems to connect and the quality of their documentation, the number of tools, how many steps need write permission, the need for logging and approval screens, the model provider to be used and the sensitivity of the data. The quote is prepared for a single example job and does not commit you.
If you want to turn the tool we write into a product others can install, see from prototype to product. Being found in search results and AI answers is a separate service: SEO and GEO.
Pricing
The price is settled in a written quote
Because it depends on the scope, we do not write this service's price on the page. Tell us what you want; we will send the scope and price line by line in writing. A quote does not commit you.
Related services
Often considered together
FAQ
Frequently asked questions
Which AI models do you work with?
We can work with models that have an MCP client and with providers that offer an API. The provider is chosen together while writing the scope, according to data sensitivity and budget.
How do you stop the model from doing something wrong?
We cannot stop it completely. We give tools narrow permissions, tie hard-to-undo steps to human approval and log every call.
Does my data go to the model provider?
We put in writing at the start which data is sent to which provider. For sensitive data we discuss locally run solutions or tool designs that narrow the data.
Do you promise savings or an accuracy rate?
No. We write a measurable goal into the scope but do not guarantee the result.
Can MCP support be added to existing software?
Usually yes. If the software has an API or a data layer, a limited MCP server can be written in front of it.
Let us send you the scope in writing.
A quote does not commit you; we settle the scope together and price it line by line.