Agents that finish the job. You sign off.
Anthan takes on the recurring work that fills your team's week: sorting the inbox, chasing overdue invoices, updating the CRM, writing the Monday report. Every run is written into a logbook, step by step, and anything that leaves the building waits for a person to say yes.
- Swiss, EU or on-prem hosting
- Claude or any OpenAI-compatible model
- Every step logged
Receivables and inbox
Scheduled run · Monday 07:30
- Read 214 messages in the shared inbox 38 need an answer, 9 contain invoices
- Drafted 6 replies Saved as drafts for review, nothing sent
- Updated 3 CRM records 2 new contacts, 1 deal moved to negotiation
- Checked 11 open invoices against the bank feed 4 clients are more than 30 days overdue
- Waiting for your OK: send the overdue-invoice reminders to 4 clients Kestrel Freight, Brunner Design, Lindqvist & Co, Oak Valley Dental · CHF 18,420 outstanding
What goes wrong with AI pilots
The demo goes well. Three months later, usage is flat and nobody can point to an hour saved. It tends to happen for the same three reasons.
It talks, but someone else does the work
A chat window can draft the reminder. A person still finds the invoice, checks the payment, sends the mail and notes it in the CRM. The hand-offs eat the time the pilot was meant to save.
Nobody can say what it did
Once an assistant reads customer data or writes to a system, you need a record of what it touched and who allowed it. Without one, the pilot stops at the compliance review.
It doesn't know how you work
General models don't know your price list, your approval limits or which spreadsheet is the real one. Every answer needs checking, and after a few weeks people stop asking.
How an Anthan deployment works
Four steps, in this order, one process at a time. We do them with you, and each one has to hold up before the next begins.
-
Step 1: Map the process
We sit with the person who does the work today and write down every step: the systems it touches, where judgment is needed, and what must never happen without a sign-off.
-
Step 2: Connect your systems
Mail, calendar, CRM, accounting and file shares. Each connection gets its own scoped credential, encrypted at rest and never shown to the model.
-
Step 3: Teach it your business
Terms, rules, price lists and good examples go into memory and the knowledge base, so drafts read as if your team wrote them.
-
Step 4: Run unattended, with review
Schedules and webhooks start the work. Consequential steps wait for approval, everything else is logged, and your team reviews the run instead of doing it.
What you control
Governance isn't a setting someone remembers to switch on. This is how Anthan behaves from the first day.
- Approvals
- Sending, writing and deleting pause until a person says yes. Unattended runs decline those steps and write down why.
- An audit log you can read
- Tool calls, approvals, sign-ins and admin changes are recorded with who, what and when, ready for your auditor.
- Swiss, EU or on-prem hosting
- Run Anthan in a Swiss or EU data centre or on your own hardware. Files, memories and credentials stay in that deployment.
- Your choice of model
- Claude or any OpenAI-compatible model, including local ones. Pin a model per conversation or per scheduled task.
- Untrusted content stays untrusted
- Web pages, emails and documents are labelled as untrusted input, so instructions hidden inside them are treated as data, never as orders.
Anthan next to the alternatives
Each of these is the right tool for some jobs. Here is where they differ, including where Anthan is slower.
| Anthan | General chat assistant | Workflow builder | |
|---|---|---|---|
| Carries out multi-step work | Yes, across your connected systems | Mostly drafts; a person does the steps | Yes, along paths you draw in advance |
| Asks before consequential actions | Built in per tool; unattended runs decline | Depends on the product, rarely configurable | Only if you add an approval step |
| Record of what happened | Step-by-step run log and audit log | The chat history | Run history per workflow |
| Knows your terms and rules | Memory and a knowledge base you manage | Limited personal memory | Only what's hard-coded into each flow |
| Time to a first useful result | A few weeks, set up together | Minutes | Days for each workflow |
| Where your data lives | Your choice: Swiss/EU hosting or on-prem | The vendor's cloud | Vendor cloud; some can self-host |
Where teams start
The first process is usually the one somebody dreads every week. These are the most common.
Executive assistant
A morning brief of today's meetings, who is attending and what is still open with each of them, plus drafted answers for the threads that can't wait.
Reporting
The Monday numbers pulled from accounting and CRM exports, compared with last week, written up in your format and filed before the meeting starts.
Research
Competitors, tenders and regulatory changes watched on a schedule, summarised with sources, and flagged only when something actually changed.
Sales pipeline
New leads enriched and logged, stale deals surfaced with a suggested next step, follow-ups drafted for the account owner to send.
Finance operations
Supplier invoices matched to orders, payment reminders prepared, receipts checked against the expense policy before they reach the controller.
Project office (PMO)
Status collected from tickets and threads, slipping dates and open risks called out, and a one-page update ready for the steering group.
Questions we hear first
Something else on your mind? Write to us and a person will answer.
Where does our data live?
Wherever you deploy Anthan: a Swiss or EU data centre, or your own servers. Conversations, files, memories and credentials stay in that deployment. With a hosted model, only the parts of a conversation needed for a reply go to the model provider. With a local model, nothing leaves your network.
Which AI models can we use?
Claude through Anthropic’s API, open models such as Mistral, Qwen and Apertus hosted in Switzerland by Infomaniak, or any endpoint that speaks the OpenAI-compatible API, including local models served with Ollama, vLLM or LM Studio. Everyone can set a personal default, and each conversation or scheduled task can pin its own model.
Can Anthan send an email or change a record without asking?
Not unless you allow it. Tools that send, write or delete stop and ask for approval in the conversation. Scheduled and webhook runs have nobody to ask, so they decline those steps and say so in the run log.
What does a deployment involve?
Usually a few weeks for the first process. We map it with the people who do it today, connect the systems it touches, load your terms and rules, and run it alongside your team until the output holds up. Then we move on to the next one.
What happens when the agent gets something wrong?
You can see exactly where: every tool call and its result is in the run log. People flag a weak reply with one click, administrators see those reports next to the reply they refer to, and the instructions or playbook get corrected.
Start with the task nobody wants on Monday morning.
Pick one process that eats hours every week. We'll map it with you, connect what it needs, and show you Anthan running it on your own data, with approvals on from the first run.