The agent's job description
The recurring work, written as a runbook with one owned outcome.
The org chart, redrawn.
AI colleagues in real roles: a job description, a human owner, work they are measured on. Not another tool. One role at a time.
The work keeps growing.
Your team doesn't.
AI tools made individual tasks faster. The process stayed slow, because work still moves from desk to desk and the context gets rebuilt at every handover. Give an agent one defined job, and the context moves with it.
"We have 30 people doing the work of 70. We can't hire. Could agents go in the org chart?"
15 years of enterprise work with 50+ clients, among them Bosch, Siemens, MANN+HUMMEL, Voith, Henkel, Zeiss, Stöckli, EnBW and TÜV SÜD.
The first agents are in production or in setup with several of these clients. Named results follow as each client signs them off.
It starts with one role.
Two steps, each at a fixed price. You approve the second before we build.
We built our first agents on our own work, and shut some down. The method comes from both.
Role-shaping workshop
Map the team's roles against the work an agent can own
Pick the best first role for an agent
Leave with the agent's job description
One agent, built and coached
Onboarded like a hire, with daily feedback
Measured against one agreed outcome
Autonomy raised only as it is earned
Handover
The agent, its job description and runbook
The skills, the evals, the access setup
Run it yourselves, or have us stay close
One job description in.
Two come back.
A real client output, anonymised: one customer-support role. Every responsibility comes back marked agent, together or human, so you see what an agent can own and what stays with people.
Agent
Order processing and system data entryRuns independently. Exceptions go to the human owner.
Together
Quotation and ticket preparationNeeds structured input from the service team.
Together
Tender process managementAgent drafts, the team approves.
Together
Complaint handlingAgent prepares, a human sends.
Human
Key account relationshipsRelationship dependent. Stays human.
What comes back
The person's job description, rewritten
What the role owns once the recurring work leaves it.
First roles we see again and again
Ticket triage
from a support JD
Owns
First-response time
Reads every inbound ticket, routes it and drafts the first reply for a human to send.
Account research
from a sales JD
Owns
Prep time per call
Builds the brief before every customer call, from history to open items.
Content QA
from a marketing JD
Owns
Review turnaround
Checks every draft against style, claims and legal rules before it goes out.
Invoice follow-up
from a finance JD
Owns
Days overdue
Watches open invoices and drafts each reminder at the right moment.
The part that matters
The second job description matters more. It shows what the person owns once the recurring work leaves the role.
From our role analysis, reviewed by our team before anything is returned.
Every agent starts supervised.
Some never graduate.
Autonomy is earned in steps and the audit trail never goes away. Roughly one in five agents we build is stopped or redesigned before it reaches production. We would rather stop the wrong job early than automate it well.
When an agent fails…
…this is what we do.
The data was never as clean as claimed
The agent didn't break the process. It exposed it.
Nobody could say what "good" looked like
A job with no definition of done can't be owned.
The human redid the work to check it
The most common failure, and it's not technical.
It answered questions but never learned
No feedback loop, no colleague. Just a tool.
Fix the input before the agent
The process was already broken. Now it's visible.
Redesign the job, or stop
We don't build a better agent for a bad job.
Clear checks first, then fewer checks
Promised, kept, and reviewed on a schedule.
Keep it as a tool, and say so
Useful is fine. Not everything becomes a colleague.
The questions IT, finance and
the works council will ask.
How is it priced?
Two fixed-price steps: a half-day role-shaping workshop, then one agent built and coached for six weeks. You approve the second fixed price before the build starts, so there is no open-ended programme.
Who has to run this internally?
One accountable owner per agent, a few hours a week during onboarding: reviewing drafts and giving feedback. No new team, no transformation office. We coach the owner along with the agent.
What about our systems and access?
We work with the access you can grant; a well-kept spreadsheet is a fine start. IT approves every access before the build, and the agent never holds credentials of its own. In production means real cases, on approved access.
What will the works council and compliance ask?
Every action is logged and reviewable, sign-offs stay with named people, and the agent takes structured routine, not jobs. The role documentation is written to be reviewed.
What do we own after the six weeks?
The agent, its job description and runbook, the skills, the evals, and the access setup. Run it yourselves, or have us stay close: both are normal.
Send us three
job descriptions.
- Send three job descriptions from the team with the biggest gap.
- The split comes back free within days: agent, together, human.
- Workshop and build only if you want it. No call, no pitch.
Matthias or Sebastian reads every submission personally.
Received. Almost there.
Matthias or Sebastian reads your job descriptions personally. The split comes back free, within days. No call, no pitch.
We've opened a pre-filled email to hello@minglabs.com. Paste your three job descriptions and hit send, and the split comes back within days.
Didn't open? Write us at hello@minglabs.com.
Redrawn one role at a time. Yours to keep.
