PromptPartner DE

Your IT and legal teams are right to worry.

A hacked website shows you spam. A hijacked agent shows someone else your mailbox and your CRM. So access, guardrails and a log are in place before the first run, and your IT partner gets them in writing.

Four people with laptops talk in a glass meeting room
Access list: quote agentIllustration with sample data
  1. Reads: request mailboxread only
  2. Reads: price lists folderread only
  3. Writes: draft quotesdrafts folder
  4. Open: route into the ERPyour IT partner decides

Signed off by your IT partner

Our recommended enterprise architecture for AI

This is how we recommend setting it up: the agent reads your systems where they are, and you can change the model later without touching them. Which model sees which text depends on how sensitive the file is.

  1. Apps and agentsWhere people and agents work

    Can be swapped
    • Microsoft 365 Copilot
    • Claude
    • ChatGPT
    • Open-source agents
  2. AI modelReads and writes the text

    Can be swapped
    • Claude
    • OpenAI
    • Local open models
  3. Connecting layerOpen standards

    Stays
    • MCP gateway
    • AI gateway
    • Skills
    • Plugins
    • Company memory
  4. Your systems and dataWhere your work lives today

    Stays where it is
    • ERP
    • Mailbox
    • Files
    • CRM
    • Microsoft 365

Governance, across every layer

  • Sensitivity levelsFour levels, each with its own place to run.
  • GuardrailsWhat it may read and write, set in advance.
  • ApprovalA person approves what goes out.
  • LogEvery step written down.
  • Acceptance testChecked on fixed examples before it runs.
Only the top two layers get swapped. The rules sit in the connecting layer, so they hold whichever app or model sits above it.

These names describe software we work with. All marks are the property of their owners.

Our recommended data classification

We recommend four classes: public, internal, confidential and secret. Your labels often already say which is which.

Data classWhere it may run
PublicIn any tool.
InternalIn Microsoft 365 Copilot, Claude or ChatGPT, under an enterprise contract.
ConfidentialIn your tools, with your labels and your written approval for each vendor.
SecretOnly on a model that runs under your firm's name.

Every option processes the text on a server, in plain text. The question is whose server, and we settle that with you before the first file runs.

In Microsoft 365, a Purview DLP policy can stop Copilot from using files and emails with chosen sensitivity labels when it writes its answers (Microsoft Learn, checked 14 November 2026).

Who may open what is settled before the first run.

Your IT partner gets all four points below in writing, and where it matters, a person decides.

  • AccessThe agent sees the data of the one job we define together. Which folders, tables and mailboxes that covers is written in the IT brief.

  • GuardrailsAt first the agent only reads and suggests. It writes to the ERP only through the route your IT partner signs off, and only after one of your people has approved the result.

  • LogEvery step is written down, so each draft can be traced back to its sources.

  • Who sees whatWho in your firm sees which drafts is set by your IT and your access rights, as it is today.

What your IT partner gets

Two pages from us: the IT brief and the data-flow sheet. They say what access we need, what we read and where the data flows.

If you use Agent 365

If your firm licenses Microsoft Agent 365, your IT sees our agents there, in its own list of agents.

Agent 365 shows an organisation's agents in one view, including agents from other companies, and is licensed per user (Microsoft Learn, checked 14 November 2026).

Five things we put in writing

  • DataAt first we only read. Before the first file comes to us, we put in writing where it's stored, and what we measure belongs to you.

  • Plain answerIf your software can already do the job, or your data can't carry it, we tell you and stop there.

  • Fixed priceFrom the AI assessment on, each step has a fixed price and a fixed scope, and we bill no hours.

  • StopYou agree each step on its own, and you can stop after any of them.

  • AcceptanceThe monthly fee starts only once the agent passes acceptance on your own past cases.

If you stop, the agents, the rules and the data stay with you.

What we build for you sits in your own environment. After a trial run we delete your data and confirm it in writing, and the data map and every figure we measured belong to you.

When one cloud stops

On 12 November 2026, Outlook, Teams and Copilot in Microsoft 365 went down for more than a day. An agent that runs in one cloud only stops when that cloud stops. So we keep the job separate from the platform and the model, and it can move.

Source: The Register, 12 November 2026.

On your own server

Must the agents run on a server you control? We run open-source agents there, and you choose the model for each job: a local model, your own key or a cloud model.

Questions about data and security

What does our IT partner need to do?

Approve read access to the data of one job. What they need to know for that is in the IT brief, on one page.

Do you sign a data processing agreement?

Yes. We sign it before we process the first file.

Where do our files go when the agent works on them?

That depends on how sensitive they are. Public and internal files may run in your tools under an enterprise contract, and secret files go only to a model that runs under your firm's name. Before the first file runs, we put in writing where it is processed.

What happens to our data if we stop?

The agents, the rules and the data stay with you, because what we build sits in your own environment. Data from a trial run we delete and confirm to you in writing.

Still have a question about your data?

Ask it in the first call. You're welcome to bring your IT partner.

30 minutes with Lukas. You leave with at least one recommendation you can use, whether or not we ever work together.

How they work