Tamboer AIFree AI plan
Menu

Human control

The workflow proposes. Your employee decides.

Human control determines the data, permissions and stop rules. Even an AI agent never gets the last word.

Example with fictitious data

A quote request, from arrival to approval

Four statuses, always in this order. Between the third and the fourth there is a person.

  1. Request received

    The request comes in

    A customer asks for a quote through the website form. Nothing has been done with it yet.

    Who: the customer

  2. Draft prepared

    AI prepares a draft

    AI extracts measurements and address from the request, looks up the pricing rules and prepares a draft quote. Anything missing is listed as a question at the top.

    Who: the workflow

  3. Waiting for review

    The draft stands still

    Nothing goes to the customer by itself. The draft waits until an employee has checked price, discount and date.

    Who: you or your employee

  4. Approved

    You approve and send

    After your approval you send the quote yourself. What is recorded: who approved which version, and when.

    Who: you

Want to try it with a customer email instead of a quote? Open the reply demo. Nothing is sent there either.

The short answer

Every workflow gets five recorded controls: which sources may go in, what the workflow may prepare and may never do, who approves, when it stops on an exception, and what gets logged. The workflow sends nothing, books nothing and promises nothing. An employee decides and stays owner of the outcome. When customer data may go into an AI tool at all sits on the page about customer data in ChatGPT.

The five controls as design choices

01

Input

It is fixed in advance which data is needed and which data stays outside the workflow.

02

Authority

A model can read information and draft a proposal without sending, booking or making commitments itself.

03

Approval

The employee sees source, draft and deviation before giving approval.

04

Exception

Uncertain or deviating cases go to a human via an agreed escalation path.

Evidence

Record source, draft, check, decision and change per case, with a retention period that fits the process.

What the website itself processes

The public site measures visits without cookies using Cloudflare Web Analytics. GA4 and Microsoft Clarity measure by default; you can turn this off in Analytics preferences. Internal visits are excluded. No Google Ads measurement is active. The request form on /en/free-ai-plan stores your request at Cloudflare (D1) and sends a notification through SMTP2GO; WhatsApp and email keep working alongside it. The calculator and the process check stay in the browser. The factual description is on the privacy page.

No bare compliance promise

I do not simply call a workflow "GDPR-proof" or "AI Act-compliant". That judgement depends on the specific processing, the data, the purpose and the chosen technology.

The way of working is concrete, though. I work with a documented data flow, a risk classification and a human approval step, so that you can demonstrate how the system works.

Limits fixed in advance

Only the data the step needs goes into the workflow. The employee validates every source and owns the outcome. The workflow has version control and a documented escalation path. That is how I build within approved tooling and governance at one of the largest banks in the Netherlands, and that is how I work at Tamboer AI.

That is not a general guarantee for every new workflow. It is the reason to redesign data, permissions and exceptions for each process.

What happens with data in practice

For each process, before the build, it is recorded which data and sources are needed and which stay outside the workflow. If an implementation uses a third-party language model, we also record which text that service processes. Fixed business rules live in verifiable code wherever possible instead of in the language model, so that the same documented rule is applied to every case.

A formal data processing agreement, retention periods and the legal basis for the processing belong to the legal and privacy review before a form route or a data-processing implementation is activated. What happens with an outage, a model update or an API change of an underlying AI service is on the how it works page.

What the client must be able to verify

  • Which input was used.
  • Which rules were applied.
  • What the workflow proposed.
  • Who approved the outcome.
  • Which cases went to a human.

Submit a process

Stop and recovery procedure

On unexpected output, a broken integration or a supplier change, the affected step stops. The team uses the manual route, assesses the cause and only resumes after a passed regression test.

Control starts with your process

Send a short message: what you do, what you are struggling with and what you want to improve. I personally tell you where best to start.

Check my process