Company agents
The same work, owned by the business.
Your people have already built agents that save them hours. A company agent does the same job — written down, tested, on company credentials, logged — and keeps running when the person who built it moves on.
Shadow agent vs company agent
The workflow is the easy part. The seven layers around it are the product.
Anyone can build the workflow in an afternoon. What lets a business keep it, trust it and prove what it did is everything around it.
Layer
Lives on one laptop, unknown to the company
Documented, controlled, audited, in the vault
01Specification
ShadowLives in one person's head and a prompt
CompanyWritten spec, documented, handover-ready
02Ownership
ShadowOne login. Gone the day they leave
CompanyOwner, editors and users, with roles and SSO
03Connectors
ShadowPersonal tokens pasted into the prompt
CompanyApproved connectors, scoped credentials
04Guardrails
ShadowNone. It will email a vendor if asked
CompanyPolicies block, tag or log actions. Model and region restricted
05Testing
ShadowTried once, then trusted
CompanyDry runs and simulations before go-live
06Audit trail
ShadowNo record of what it did or touched
CompanyEvery run logged: tool calls, data, outcome
07Compliance
ShadowInvisible to the CISO, the auditor and the insurer
CompanyEvidence pack for ISO 42001, SOC 2 and cyber insurance
How they're organised
One agent per department, with a job for every sub-agent.
Each department gets a parent agent — the Finance agent, the Sales agent. People talk to it. Behind it sit sub-agents for specific jobs, each rebuilt from work someone was already doing.
- Access, connectors, memory and rules set on the parent apply to every sub-agent
- People ask "@Finance agent" — they don't need to know which sub-agent does what
- Adding a job means adding a sub-agent, not another tool to learn
Finance agent
finance · 3 sub-agents · 12 people
- Supplier reconciliationweekdays 09:00 · xero · outlookLive
- Invoice chaserevery 6 hours · xero · outlookLive
- Spend reviewwhen a bill is approved · xeroAwaiting approval
Example data
Why they're better
Worth switching to, not just safer.
A company agent has to beat the shadow version on the things the person cares about, or they'll keep using their own.
01 · Continuity
Keeps running when people leave.
It's owned by the company and runs on company accounts, so holidays, role changes and resignations don't break it.
02 · Quality
Tested before it goes live.
Dry runs on real data and a set of checks every version must pass. The owner approves a test run before anything runs on its own.
03 · Security
Company credentials, scoped tight.
Approved connectors with only the access the job needs. No personal tokens pasted into prompts.
04 · Evidence
Every run on the record.
What it did, what it touched and who approved it — the audit trail exists before anyone asks for it.
05 · Improvement
Gets better from real use.
How people actually use the agent feeds back into it, so it keeps up with the job instead of drifting from it.
06 · Reach
One agent, wherever people work.
The same agent answers in Slack, Teams, Claude and ChatGPT, under the same controls everywhere.
Where people use them
No new app to learn.
People ask the Finance agent the way they'd ask a colleague — in the tools they already have open. The agent, its access and its audit trail are the same on every surface.
FAQ
About company agents.
Where do company agents come from?+
From work your people already do with AI. Detection finds it, an interview with the person captures how it runs, and the agent is rebuilt from that spec and tested on real cases.
What happens to the person's own agent?+
Once the company agent does the job, they can stop running theirs by hand. It does the same work, with connectors and context they couldn't set up themselves.
Who decides when an agent goes live?+
Its owner. They watch a test run on real data and approve it. New versions go through the same approval, and the previous version keeps running until then.
Can an agent do something it shouldn't?+
Guardrails block, tag or log actions — emailing outside the company, for example — and every run is recorded, so you can see exactly what happened.
Keep reading
How they're found, and how they're run.
Book a demo
See the AI work already happening in your company.
Read-only access, nothing to install, first findings in days.