Key takeaways
- Every outcome has a human owner, even when an agent does the work.
- Delegation is explicit: a person hands a specific piece of work to an agent.
- Approval comes before anything leaves the company or changes something hard to undo.
- Trust grows with a visible track record, and oversight means seeing what was done and why.
Why companies need an operating model for AI
Most companies have an operating model for people, even if nobody calls it that. There are roles, reporting lines, approval limits and review rhythms. Everyone roughly knows who can sign what, and who answers when something goes wrong.
AI agents arrive without that structure. They can draft, schedule, update records and send messages. The harder question is what they should be allowed to do, on whose behalf and with what checks. Without clear answers, companies swing between two failures: blocking agents so tightly that they add little, or letting them act so freely that nobody can say who approved what. Both are costly. Too cautious, and the company pays for AI it barely uses. Too loose, and one wrong message to a customer sets trust back for months.
The five parts of the model
| Part | The question it answers |
|---|---|
| Accountability | Who owns this outcome? |
| Delegation | Which work can be handed to an agent, and by whom? |
| Approval | What must a person approve before it happens? |
| Trust | How does an agent earn more responsibility? |
| Oversight | How do we see what was done and why? |
None of these parts is new. Companies already answer each question for their people. The model extends the same answers to agents, in writing, before the agents start work.
Accountability and delegation
Accountability stays with people. When an agent sends a reminder, the person who delegated it owns the result. This keeps responsibility clear and makes delegation a deliberate act.
Delegation works best when it is specific. "Handle the Bata account" is too broad for an agent. "Draft the quote email with the confirmed price for Harsh to review" is a clear handoff. Each handoff names the work, the person it came from and what done looks like. A useful habit is to keep a short list of work that may be delegated, owned by the team that does that work. The list grows as agents prove themselves on it.
Approval, trust and oversight
Approval is the safety catch. A simple default works for most companies: anything that leaves the company, commits money or is hard to undo needs a person to approve it first. Internal, reversible steps can move faster. The rule should be written down and easy to read, so nobody has to guess whether a step needs approval.
Trust is earned the way it is with a new hire. At the start, every action is reviewed. As the record builds, routine actions of a known kind can be approved more lightly. The record should be visible to the people who rely on it, so trust rests on evidence.
Oversight means anyone responsible can see what was proposed, what was approved, what was done and what happened next. When something goes wrong, the trail shows where. GeniOS follows this model. Each piece of intelligence ends with four answers: Delegate, Done, Later or Not relevant. Nothing is sent until a human approves, and the work closes only on proof.
Worked example: an approval chain at Vaultex
Vaultex needs its IS 15298 test reports ready for next month's Adidas vendor audit. Anushka owns the outcome. She delegates two pieces of work to agents: scheduling the remaining tests with the lab, and drafting a note to Adidas that lists the documents Vaultex will share.
The scheduling agent proposes dates. Anushka approves them, because they affect the production plan. The note to Adidas waits for Anisha's approval, because it goes to a major customer. Each step is visible: what was proposed, who approved it, and when the booking appeared in the calendar. Anushka stays accountable throughout. The agents carry the routine work, and the people make the calls. If anything goes wrong, Anushka can see exactly where.
A sample delegation and approval table
An operating model becomes real when it is written as a table people can check. Here is a starting point that many companies can adapt.
| Kind of work | May be delegated to an agent | Approval needed | Approved by |
|---|---|---|---|
| Internal reminders to colleagues | Yes | Yes at first, then in a batch once a track record exists | The owner |
| Proposing meeting times | Yes | Yes, before invites go out | The meeting owner |
| Routine replies to customers | Yes, as drafts | Yes, every time | The account owner |
| Messages to suppliers about dates | Yes, as drafts | Yes, every time | The purchasing owner |
| Prices, discounts and terms | No | Decided by a person | Finance or the account owner |
| Changes to orders or records | Only with a named owner | Yes, every time | The record owner |
The table should belong to the people who do the work, and it should change as agents prove themselves. The setting approval rules playbook walks through building your own.
How trust grows: three stages
- Suggest
The AI notices and suggests. People do all the work. This stage costs nothing in risk, and it shows quickly whether the AI understands the company.
- Draft and approve
Agents prepare the work and a person approves each piece before it happens. Most external work stays here for good, because a person should see what leaves the company.
- Lighter review
For internal, reversible work with a strong record, review becomes lighter: the owner approves routine items in a batch and checks samples. Anything that leaves the company still waits for its own approval.
Moving a kind of work from one stage to the next should be a deliberate decision, made by its owner, based on the visible record. Trust that grows by default is trust nobody granted.
Writing your operating model on one page
A human-AI operating model does not need a long policy document. One page that people actually read is better. It should cover:
- The principle that every outcome has a named human owner.
- The delegation table: what may be handed to agents, and by whom.
- The approval rule: anything that leaves the company, commits money or is hard to undo needs a person.
- How a kind of work moves between trust stages, and who decides.
- Where anyone can see what was proposed, approved, done and proved.
- What happens when something goes wrong: who is told, and how the rule is changed.
- When the page is reviewed. Quarterly suits most companies.
The enterprise AI governance article covers permissions and agent authority in more depth, and enterprise AI observability covers how to monitor agent decisions, actions and outcomes.
Common mistakes
- Writing the model after the agents arrive. Agents deployed without rules create habits that are hard to undo. Write the page first, even if it is short.
- Making approval slow. If approving takes longer than doing the work by hand, people bypass it. Each approval should show the draft, the reason and the context in one place.
- Sharing accountability. "The team owns it" means nobody does. Name one person for each outcome.
- Hiding the record. If only administrators can see what agents did, trust cannot grow across the company.
- Never revisiting the rules. A model written for the first agent will not fit the tenth. Review it on a schedule.
Questions about data, access and reliability usually come up at this point. The architecture, security and reliability FAQ answers the common ones, and how to implement AI agents in an enterprise covers rollout. To talk through an operating model for your own company, contact us.
A human-AI operating model is how a company keeps control while giving AI more of the work. It answers the questions every team will ask once agents arrive: who owns this, who approved it, what was done and how do we know. Companies that write it down early move faster, because people trust what they can see and check. Companies that skip it either hold AI back or lose track of it. Start small: one page, one team, one kind of delegated work, and a review every quarter. The model will grow with the company, and so will the share of work that AI can safely carry. The aim is a company where every outcome has an owner, every risky action has an approval and every result has proof.