The AI-native operating model

From tools to teammates

For most of the last three years, AI at work meant a tool in the hands of a person. Someone opened an assistant, asked a question, and used the answer. The person did the work. The tool made them faster.

Agents change that. An agent can take on a step of a process: read the invoice, match it to the order, draft the reply, propose the rota. The work moves from the person to the agent, and the person’s job moves to checking, deciding and handling what the agent cannot.

That shift is why the operating model matters more than the model itself. McKinsey’s 2026 survey found that nearly three-quarters of AI high performers have fundamentally redesigned their workflows, against about a quarter of other companies. BCG’s 2026 index points the same way: the leaders reshape workflows end to end, and more than 80% of AI spending now sits outside the IT budget, in the business units that own the work.

The direction of travel is clear. In BCG’s survey, 42% of companies expect to give agents real decision-making authority by 2030. Only 5% have the full set of controls in place today. The operating model is how a company closes that gap.

Three ways to share the work

Most processes can be split between people and agents in one of three ways. The right choice depends on the cost of an error and on how predictable the work is.

Assist is the natural starting point for anything that speaks for the company. At a UK water utility with about 6,000 employees, Twistag built an assistant that scores every draft against the utility’s own communication guidelines and suggests rewrites. Staff decide whether to accept each suggestion, and the assistant publishes nothing. Six weeks into the pilot it was scoring more than 3,000 communications a month, saving the communications team around 18 hours a week, with zero non-compliant language flags in regulatory bulletins.

Approve by exception suits high-volume back-office work. In the Aralab invoice agent described in Chapter 1, about 80% of invoices need no review. The rest land on a review screen, sorted by what the agent is unsure of. The finance team stopped keying data and started checking the cases that need a person.

Supervise is where the largest gains sit, and it should be earned. An agent moves to this mode once it has run in one of the first two modes long enough to show its error rate, and once the controls in Chapter 8 are in place.

A single process often uses more than one mode. An agent can draft customer replies in assist mode while it categorises incoming tickets under supervision. The mode belongs to each step, and each step can move up a level as trust grows.

Redesign the workflow, then place the agent

The common mistake is to add an agent to the process as it runs today. The process keeps its old hand-offs, approvals and queues, and the agent saves a few minutes inside one step. The gains that show up in the P&L come from changing the flow.

Redesign works best with the people who run the process, in five steps:

1. Map the work as it is. Every step, who does it, how long it takes, where it waits, and where errors come from.
2. Label each step. Rules a person could write down, judgement that needs context, or a relationship that needs a person.
3. Assign each step. Agent, person, or agent with approval. Rules go to the agent. Relationships stay with people. Judgement usually starts as agent with approval.
4. Set the hand-offs. Where the agent passes work to a person, what the person sees, and how fast it comes back. A review screen designed for speed matters as much as the agent.
5. Set the numbers. Cost per item, cycle time, error rate and share of items handled without review, each with a baseline.

The hotel group in Chapter 1 is a good example of this. The optimisation engine was only part of the change. Scheduling moved from each property alone to a shared view across the group, and general managers kept the decision on every proposed rotation. The process changed shape, and the agent made the new shape possible.

What changes for people

When agents take on routine steps, the work that remains for people changes in three ways.

• More judgement, less handling. People spend their time on exceptions, decisions and conversations. The finance team at Aralab moved to forecasting and supplier work.
• Ownership of outcomes. Someone has to answer for what the agents produce. McKinsey describes people moving from executing tasks to owning outcomes, with small human teams of two to five overseeing end-to-end processes run by many agents.
• New roles. Every agent needs an owner in the business, someone who tunes it, and someone who watches its numbers. In smaller companies one person often holds all three. BCG expects dedicated AI roles to grow from 7% of the workforce in 2026 to 22% by 2030.

Microsoft’s 2025 Work Trend Index, a survey of 31,000 people in 31 countries, suggests a useful measure for this shift: the ratio of agents to people for each role or team. The right ratio differs by process. The point is to track it on purpose.

Twistag runs its own engineering this way. In its delivery system, Pulse, five agents cover specification, architecture, code, quality checks and deployment. About 90% of the code Twistag ships is spec-driven and AI-generated. Engineers approve every change, and every pull request passes eight automated checks before a senior engineer signs it off. A five-person squad delivers what traditionally needs eight to ten engineers. The people did not disappear. Their work moved to architecture, review and time with the client.

Who owns what

An operating model also has to say who owns each part of the work. A split that works for most mid-market and enterprise companies:

• The business owns the outcome. Each agent has an executive owner whose budget or target it moves. That owner decides the mode, signs off the controls and answers for the result.
• A small central team owns the standards. Shared tools, security, model choices, evaluation and the register of every agent in use. It does not build every agent. It makes sure every agent is built and run the same way.
• IT and data own the foundations. Access to systems and data, identity, logging and cost tracking. Chapter 4 covers what the data must support, and Chapter 11 covers the platform once many agents run every day.

Without the central team, every department builds its own agents with its own tools and rules, and nobody can say how many exist or what they cost. Without business ownership, agents become an IT project that nobody adopts.

Measuring the new model

Adoption numbers, such as how many people used an AI tool this month, say little about value. Four numbers say more:

• Cost per item, such as an invoice, a claim, a reply or a shift, before and after.
• Cycle time from request to completion.
• Share handled without review, which shows how far each step has moved toward supervise mode.
• Override rate, how often people reject or change what the agent proposed. A falling rate builds the case to give the agent more room. A rising rate is an early warning.

Chapter 10 sets out how to report these to the board every quarter.

Common mistakes

• Adding an agent to an unchanged process. The old queues and approvals stay, and the saving stays small.
• No owner for the agent’s work. When something goes wrong, nobody answers for it, and trust in every agent drops.
• Moving the experts out too early. The people who know the exceptions are the ones who train, check and improve the agent. Keep them close for the first months.
• Jumping straight to supervise. Skipping the assist and approve stages removes the track record that justifies giving an agent more room.
• Every team for itself. Ten departments, ten tools, no register and no shared standards.

What to do this month

1. Pick one process from the short list in Chapter 1 and map it with the people who run it.
2. Label every step as rules, judgement or relationship, and choose a mode for each.
3. Name the business owner for the agent, and agree the four numbers with their baselines.
4. Set up a small central team, even two or three people, to own standards and the agent register.

Questions for the leadership team

1. In the processes on the short list, which steps follow rules, which need judgement, and which depend on a relationship with a customer or supplier?
2. For each agent planned or running today, who in the business answers for its results?
3. Where would the people who handle exceptions today sit once an agent handles the routine work?
4. Which numbers will show that the new way of working is better than the old one?
5. Who sets the shared standards for agents, and who decides which teams can build them?