The principle: machine drafts, humans decide
Every AI step in our flows ends at a clear point. Either the result is safe enough for a rule, or it waits for your approval. Uncertain fields get a check note instead of a guessed answer.
- AI prepares: text, classification, summaries, categories
- Rules decide the clear cases: duplicates, thresholds, required fields
- People approve anything that goes outside: send, booking, payment, commitment, activation, access removal
- Uncertain fields are flagged: unchecked status plus a concrete note on what to check
Models in the EU, data in the EU
We prefer EU systems, including capable open-source models. Other models only if they are hosted in the EU. Operations and backups run on European infrastructure. That does not replace a processing agreement; one is signed per project.
- EU systems and open-source models first
- Other models only with hosting in the EU
- Credentials live in the instance, not in chats
- Access by person, revocation after the project
- No hidden training on your customer data
Four patterns from running flows
Abstracted from real work, no client names. Each pattern follows the same order: capture, check, then commit.
-
1. Capture
Audio enters the flow, split into chunks automatically.
-
2. Transcribe
Speech recognition writes the conversation with speaker separation.
-
3. Summarise
AI drafts an internal and a neutral external version.
-
4. Approve
A person checks and approves before anything is sent.
-
5. File
Record and document land in the system with an audit trail.
-
1. Note arrives
Message or file from day-to-day work.
-
2. AI reads along
The model pulls out people, companies and requests.
-
3. Check existing records
Matched against contacts and companies on file.
-
4. Create or enrich
New only what is missing. Otherwise enriched.
-
5. CRM is current
Contact, company and history sit in one place.
-
1. Document arrives
Invoice from inbox or chat; duplicates leave via checksum.
-
2. Read
Text recognition reads amounts, dates and numbers.
-
3. Classify
AI proposes category and tax; uncertain items are flagged.
-
4. Review and approve
Draft in accounting, a person approves.
-
5. Book
Only after approval does the document become binding.
-
1. Occasion is set
Onboarding or a ticket supplies the context.
-
2. Draft
AI writes the text from the case, with the open points.
-
3. Approval
A person checks tone and content.
-
4. Send
Only then does the message go out.
Where AI deliberately does not decide
Some steps stay manual with system support. Binding quotes, payments, cancellations and anything hard to undo need a person. AI delivers the draft; responsibility stays with you.
- No automatic send to customers without approval
- No booking or payment without review
- No commitments on dates or prices without a person
- No access removal without approval
- Every case stays traceable with time and status
Frequently asked questions
Which AI models do you use?
EU systems first, including capable open-source models. Other models only if they are hosted in the EU. Which model runs in your flow is written into the order.
Do you train models on our data?
No. There is no hidden training on your customer data. Processing only on instruction, with a processing agreement and deletion after the project.
What happens when the AI is unsure?
Then it does not guess. The field gets an unchecked status and a concrete check note, for example check invoice number or check due date.
Can the AI send without approval here?
Only where the risk allows it and you agreed in writing, for example purely internal notes. Anything that goes outside or moves money stays tied to approval.
Where do I see what the AI did?
Every case gets a log entry with time and status. Drafts, approvals and sends stay verifiable.