Agentic platforms
We build platforms where AI agents take on part of the work across web, mobile, messaging and the back office, and people approve what matters. Then we deploy them and run them.
What you get
- Agents with a signed persona card and an autonomy rung your executive chose
- One platform behind every channel: web, mobile, messaging and back office
- A human approval queue for payments, bank details and anything irreversible
- A stop control on every agent, and every step logged
- A dashboard with cost per completed task, handover rate and drift
Who it is for
Your work runs across several systems and channels: invoices to match, counterparties to check, customer questions arriving on five apps. Some steps need judgement on documents that do not fit a form. You want agents to take that part, under rules your auditor can read and your accountable executive has signed.
What we build
- Back-office agents. For accounts payable: agents read invoices, match them to orders and goods receipts, draft supplier queries and flag fraud signals, and a person decides every exception.
- Agentic apps. Web and mobile apps where an agent does part of the work for your customers or staff.
- Agents on messaging. One reasoning and knowledge layer behind web chat, Telegram, Messenger, Instagram and WhatsApp.
- The platform under them. Tools, retrieval, routing across model providers, budgets per tenant, row-level security between tenants, and monitoring.
We use an agent only where the path cannot be written down in advance. Everything else becomes a workflow, which costs less to run and is easier to audit.
How every agent is governed
Each agent gets a persona card before we write code: its role, what it reads, what it may and may not do, when it hands over and what it logs. Your accountable owner signs it.
Its autonomy is a rung on a four-rung ladder: Assist, Suggest, Act with approval, Act and report. Your accountable executive chooses the rung. An agent moves up only when evaluation results support it, and moves down after an incident.
Guardrails apply on every rung:
- Tools carry the narrowest permission that does the job.
- Personal data is replaced with aliases before it reaches a model.
- Instructions found inside documents and emails are treated as data, never as commands.
- Every agent has a stop control that a named person can use, and each use is logged.
How an engagement runs
| Stage | What happens | Duration |
|---|---|---|
| 01 · Ground the case | Which steps need an agent; human and agent personas; a service blueprint with a lane for each | 3-5 weeks |
| 02 · Prove the value | An evaluation set frozen before the prototype; the prototype run against it | 3-6 weeks |
| 03 · Engineer the system | The platform for the pilot footprint: connectors, approval queue, guardrails, monitoring; persona cards signed | 8-16 weeks |
| 04 · Pilot in one place | One team works with the agents after a red-team session; every handover and override logged | 6-12 weeks |
| 05 · Scale what works | More teams, sites or channels, with integrations and budgets extended | 3-6 months |
| 06 · Run and evolve | Operation; every model change reruns the evaluation; quarterly value and rung review | Ongoing |
Each stage is contracted separately, and the price is quoted on request once its scope is fixed.
Built for Switzerland
The hosting option is chosen in stage 01, and personal data stays out of model calls wherever the design allows. Where an agent’s decision about a person is taken solely by automated means, revFADP Art. 21 gives that person the right to a human review, so we build the review path before go-live. If the platform’s output is used in the EU, the EU AI Act can apply, and agents that assess the creditworthiness of individuals are listed there as high-risk. Dates, articles and sources are on the regulatory reference.
Proof from the lab
| Build | Label | What it shows |
|---|---|---|
| Atrium | Working prototype | Five role-bound agents for accounts payable, a human decision queue with keyboard shortcuts, a duplicate-invoice check that uses past work as memory, and token spend per tenant |
| kredis | Beta | Counterparty risk files where agents draft and named people approve, with two people for key actions, a decision endpoint with no language model in its path and an append-only history |
| Telebimmer | Live | Proactive monitoring on the official BMW CarData service, delivered in Telegram, with a read-only token so owners can connect their own AI agent. Not affiliated with BMW AG |
| A multi-channel AI concierge | In production | Five channels on one layer, cost-aware routing across model providers with fallback, and a staff dashboard that estimates cost per channel |
| A tiered document intelligence pipeline | Working prototype | A stronger model only where the task needs it, with rate limits and budget caps per tenant |
Built on
Questions
When is an agent the right answer?
Only when the path through the task cannot be written down in advance. If the steps are known, a workflow of model calls and code does the job more cheaply and is easier to audit. In regulated processes the workflow is our default.
Can an agent release a payment or change master data?
Not on its own. Payments, bank details and anything that cannot be undone pass through a human approval queue at every rung of the autonomy ladder.
How does the platform connect to our systems?
Read-only first. Write access comes one action at a time, each with its own approval path, and every connector carries the narrowest permission that does the job.
Which models do you use?
The one the evaluation supports for each step, within the hosting option agreed for your data. The model sits behind a provider abstraction, so a successor can replace it once it passes the same evaluation set.
Can other companies’ agents use our platform?
Yes, if you want them to. An outside agent is a user too, so we design your tools for it: clear names, narrow scopes and short responses. Telebimmer already gives car owners a read-only token and a skill file for their own AI assistant.
Next step
Have a process an agent could take on?
The AI Opportunity Assessment tells you in 1-2 weeks whether an agent, a workflow or neither fits your process. The price is on request.
See the AI Opportunity Assessment