Digital Employees · Workforce model
One org chart, one task board, one policy engine — for people and AI
Humans and Digital Employees work side by side. Every Digital Employee has a named human owner, agents never manage people, and kill switches stop any agent in seconds.
A hybrid workforce, not a pile of bots
One3 doesn't sell "agents". It runs a hybrid workforce. Humans and Digital Employees sit in one org chart, receive work from one task system, are governed by one policy engine and are measured with one scorecard model. The only differences are the autonomy ceiling, the risk ceiling and who is accountable: every Digital Employee has a named human owner who is accountable for its work.
Digital Employee = Agent + Job (role, KPIs, supervisor) + Employment record
+ Compute salary (budget) + Identity (non-human) + Career (promotion, demotion, retirement)
Rules of the org chart
Four rules.
- 1
One accountable human
Every agent has exactly one accountable human owner and one operational supervisor. Its supervisor chain always ends at a human within three hops.
- 2
Agents never manage humans
Manager agents may propose assignments to people; people accept them. Agents can only assign work to other agents.
- 3
Dotted lines are allowed
A Content Writer agent can serve a Manager agent and a human lead. Conflicting instructions go to a conflict resolver.
- 4
The org chart is policy
Supervisor relationships drive approvals, escalations and data access.
Four classes of agent
| Class | Purpose | Typical autonomy ceiling | Max risk without a human |
|---|---|---|---|
| Executive | Cross-department synthesis, briefings, strategy options. They advise; they don't execute business Actions. | A2 | R1 |
| Manager | Turn goals into plans, allocate work to specialists, monitor KPIs, re-plan | A3 for R1–R2 orchestration | R2 |
| Specialist | Do the domain work through Actions | A1 → A4 per skill, earned | R3 after earned autonomy; R4 never without a human |
| Utility | Stateless helpers: extract, classify, translate, crawl, brand check | A4 within the caller's permissions | R0–R1 |
No privilege escalation
Utility agents inherit the caller's permissions capped by their own — the intersection, never the union. An executive agent can't use a utility to get around its own limits.
Team patterns
| Pattern | When to use |
|---|---|
| Hierarchical | Clear goals with separable sub-tasks, such as a campaign launch or month-end close |
| Pod | Ongoing programmes, such as an agency client account with content, SEO, ads and a human lead |
| Swarm | High-volume homogeneous work: tag tickets, enrich leads, extract invoices |
| Market / bidding | Many interchangeable agents with different costs (planned W4) |
| Debate / critic | High-stakes content or decisions where roughly 2× cost is justified; the judge is a human for R3+ |
| Pipeline | Deterministic stages: brief → draft → brand check → approval → publish |
| Escalation desk | Agents handle the default path; exceptions go to a human desk |
How people and agents work together
Seven collaboration patterns.
Copilot
Human leads; AI suggests inline; human executes. For judgement-heavy work.
Delegate
Human assigns a task with a definition of done; AI executes.
Supervisor
AI executes within policy; human reviews samples and exceptions.
Reviewer
AI drafts; human approves every output. For R3 work, new roles and regulated content.
Pair
Human and AI work in the same record in real time, each attributed.
Escalation desk
AI handles the standard path and hands off with a structured note.
Reverse delegation
The agent asks a human for an Action it isn't allowed to take: R4, physical tasks, calls.
Where you see it
Planned workforce surfaces.
Mixed task board
Human and AI assignees, AI badge, confidence and cost-so-far; reassign between them.
Approvals Inbox
Diff preview, rationale, evidence, risk tier and similar past decisions; approve, edit or reject with a reason.
Desk of an agent
Queue, in-flight runs, decisions, KPIs, budget burn, autonomy per skill; pause, demote or replay.
Workforce directory
People and agents by team, with skills and capacity.
Run replay
Step-by-step: context, model calls, tool calls, policy checks.
Weekly workforce report
Outcomes, cost, quality, incidents and promotions due.
Structured handoffs
When an agent hands work to a person, it sends a structured note, not a vague message: who it is from and to, the record, a reason code, a summary, what was done, what is needed, evidence links, customer sentiment, the SLA deadline and a confidence score. For example: "Customer says invoice double-charges the annual plan. Verified two payments. Needed: decide refund vs credit note (R4)." The human can accept it, ask the agent a question or return it.
Agent identity and controls
| Control | Design |
|---|---|
| Principal type | Agents are a distinct principal type, never a human account and never sharing a human's credentials |
| Credentials | No long-lived secrets; short-lived tokens per run; external connections held by the platform vault |
| Least privilege | Role template ∩ tenant grants ∩ owner's own permissions ∩ data classification. An agent can't have access its owner doesn't have |
| On-behalf-of | When acting for a user, both identities are carried and both must be allowed |
| Just-in-time elevation | Temporary scope for one object via the Approvals Inbox, expiring after use |
| Segregation of duties | The same principal can't both create and approve a payment, vendor, payroll change, credit note or user grant |
| Non-repudiation | Every Action record is signed and stored in a hash-chained audit log |
| Impersonation ban | Agents can't send as a person's mailbox or name unless delegated and marked AI-assisted. "Send as CEO" is forbidden |
| Tainted input | Untrusted content can't authorise R3 or R4 Actions |
Kill switches
| Level | Scope | Who can trigger | Target time *(estimate)* |
|---|---|---|---|
| K1 | One skill of one agent | Owner, supervisor, governance, detectors | < 5 s |
| K2 | One agent | Owner, governance admin, detectors | < 5 s |
| K3 | Tenant-wide AI workforce or one department | Tenant admin, One3 on-call with notice | < 30 s |
| K4 | Platform-wide model, provider or skill class | One3 incident commander | < 60 s |
Tested, not assumed
Kill switches are tested in quarterly drills, and drill results are audit evidence. K1/K2 are planned for W1, K3 for W3 and K4 drills for W4.
Questions about working with Digital Employees
Can an AI agent be my manager?
No. Agents never manage humans. Manager agents may propose assignments to people, who accept or decline.
Who is responsible when an agent makes a mistake?
Its named human owner is accountable, and the incident flow contains, reverses or compensates, and fixes the cause.
Is agent cost shown with oversight time?
Yes, always. An agent that needs review time per task isn't free labour.
Can an agent access data its owner can't?
No. An agent's permissions are capped by its owner's own permissions.
How do I stop an agent right now?
Use a kill switch from the agent's desk. K1 and K2 target under 5 seconds.
Try a hybrid team
Design partners run their first Digital Employees alongside their own people in shadow mode.
More in Digital Employees