AI agent development for whole workflows
Receive, transcribe, match, extract, write back. We build agents that run a defined workflow through to the end, with approvals wherever a person should decide.
- AI agent
- An AI agent does not wait for the next instruction. It runs a defined workflow through to the end on its own: take the input, process it, write the result back into the systems. A chatbot answers questions. An agent completes steps. This is what people mean by agentic AI.
The meeting ends. The work does not.
The recording sits in a folder. Someone writes the minutes in the evening. What was decided arrives as accurately as that person remembers it.
Add several sites and you add time zones and languages. Just finding out who currently owns a topic can take until next week.
We build the part in between. Decisions stay with people, the handovers go to the machine. Wherever a check is needed, a person sees the result before it is written. That is how the meeting project for a client with sites in Germany and Japan has run for months.
Five stages, each on its own
Each stage runs independently. If one fails, the workflow resumes there, not at the start.
Intake
The recording and calendar data come in and are matched to participants and agenda.
Transcription
Speech becomes text. If the recording must not leave the building, speech recognition runs on your hardware.
Name matching
The same person appears as Müller, Mueller and M., plus whatever speech recognition misheard. Matching against the staff directory merges them. Skip this stage and the Owner column is worthless.
Extraction
Decisions, tasks, deadlines, owners. Before anything is committed, it all sits in one list that a person reviews.
Write back
Into the knowledge system, into Microsoft 365, into the line of business application. From then on the content is searchable.
What we put together
Speech recognition
Runs entirely on your side if you want it to. In the meeting project it runs on the client's hardware for exactly that reason.
Matching against master data
The staff directory is the truth. Spelling variants and misheard names are mapped onto it. If nothing matches, the agent asks instead of guessing.
Extraction rules
Decision, task, deadline, owner. A fixed pattern, so the results can be counted and searched later.
Human approval
A list to review before anything is committed. Corrections feed into the next runs.
Integration
Through APIs where they exist. Where they do not, through file handover or screen automation. We pick connection standards such as MCP only once we know which systems are involved.
Log
Which stage did what and when stays on record. Without it, nobody can trace a cause afterwards.
When an agent fits and when it does not
Lots of meetings alone are not a reason.
| Situation | Better choice | Why | What we recommend |
|---|---|---|---|
| Recurring meetings several times a week | Agent | Minutes and distribution cost hours | The typical case |
| A few meetings a month | Wait | Build and operation do not pay back | Standardise the template first |
| Clear decision rules | Rule based automation | A language model blurs decisions | Implement without a model |
| Recordings must not leave the building | Operation on your own hardware | Speech recognition runs in house too | Local configuration |
| Several sites and languages | Multilingual setup | Spellings and language switches lower accuracy | Matching as a stage of its own |
Scroll sideways to see all columns.
Four steps to parallel operation
We start with one type of meeting. That is where the assumptions prove right or wrong, and everything else depends on it.
Before that we measure the current state. Without a measurement there is no effect to show later. Technical documentation under the EU AI Act is part of the handover.
AI agent development: common questions
Related topics
Tell us which meeting costs the most follow up work.
Thirty minutes are enough to say whether an agent would carry there.