AI use, explained for business leaders

Are you using AI? Wrong question.

Almost every organisation is using AI, so the answer is always yes. The useful question is how. Asking AI for an answer and letting AI make decisions in a live process are very different things, and they need very different levels of control.

AI as an adviser
It answers or drafts. A person reads it and decides.
AI with authority
It acts or decides, and the result takes effect without a person checking each one

Most AI use today is advice, and advice does not need a complex solution. Authority does.

1. The story

One insurer, four ways of using AI

The same organisation, the same AI models. Watch where the person sits in each case, and who ends up owning the outcome.

Choose an industry

Everyone says yes

Ask AI answers a person Claims handler AI assistant Claims system Not touched "What is our excess on flood claims?" Outcome owned by the person Assist AI drafts, a person approves Reviewed Claims handler Draft AI assistant Sent Claims system "Draft the settlement letter for this claim" Outcome owned by the person Act AI carries out what you asked Sarah, claims handler asks and authorises AI agent CRM and claims system Acting on Sarah's login "Update the claim file and book the follow ups" Sarah decided. The AI chose the steps. Decide AI makes the call, case by case Claim arrives No person reviews each case AI agent Appointed Evidenced Payment issued "Settle motor claims under $5,000" The AI decided. The business owns it. Using AI? Yes Using AI? Yes Using AI? Yes Using AI? Yes Four honest answers of yes. Four very different levels of risk, and of control needed. A person sees and owns every output The output takes effect with no person checking it Control needed: light Control needed: moderate Control needed: substantial Control needed: full Decide is where most of the value sits. Rules based automation stops at the exception. Exceptions wait in a queue for a person. AI can carry many of them all the way through, but only when every decision is bounded, attributable and reviewable. Today: exception, queue, person, decision With governed agents: exception, agent, decision plus evidence

2. Four ways to use AI

Same technology, very different responsibility

What changes from left to right is not how clever the AI is. It is whether a person still stands between the AI and the outcome.

Ask

AI answers questions, summarises and researches. A person reads the result and decides what to do.

Who owns the outcome
The person who asked
Proportionate control
A usage policy, rules on what data can be shared, and awareness that AI can be wrong
Typical examples
Policy questions, research, summarising documents

Assist

AI drafts work inside a process. A person reviews and approves each output before it goes anywhere.

Who owns the outcome
The person who approved it
Proportionate control
As above, plus quality checks and genuine review
Typical examples
Customer letters, reports, code that goes through review

Act

A person decides what needs doing and gives AI permission to do it. The AI chooses the steps and carries them out in your systems, often on that person's login. Nobody checks each step.

Who owns the outcome
The person owns the decision. Each step the AI took is often unattributed.
Proportionate control
Permissions limited to the task, its own identity, a record of each step, and ideally the ability to undo
The difference from Assist
Mistakes land first and are found later. A clear request makes them less likely, not visible before they land.
Typical examples
Updating records, booking meetings, raising tickets

Decide

AI decides what should happen in each case. Nobody asked for that particular outcome, and nobody reviews it.

Who owns the outcome
The business, through a named owner
Proportionate control
A formal appointment, an owner, certification, checks on every action and evidence for every decision
Typical examples
Settling claims, approving applications, resolving exceptions
Ask and Assist: a person decides, and checks every output before it takes effect.Act: a person decides, the AI carries it out unchecked. Decide: the AI makes the decision itself.
3. Where does yours sit?

Place a use case, see what it really needs

Pick an example. Each one is placed by a single test, and the answer shows what control is proportionate and, just as important, what is not needed.

Choose an industry
  • Use case
Ask
Assist
Act
Decide
YES

Does a person see and own every output before it takes effect?

What it needs

    What it does not need

      Illustrative examples. Where a real use case sits depends on how it is set up, not what it is called.

      4. The test

      One question decides how much control you need

      Does a person see and own every output before it takes effect?

      YES

      The AI is an adviser

      A person is accountable for every outcome, as they would be with any other tool or source of advice. Keep controls proportionate: clear usage rules, sensible data handling, and people who know AI can be wrong. A complex control platform is unlikely to be justified.

      NO

      The AI has authority

      Something other than a person is now shaping outcomes, so ask a second question: who makes the decision? If a person decides and the AI carries it out, that is Act: the judgement is still the person's, but you need confidence the AI will do what was asked, because a mistake lands before anyone sees it. Give it permissions limited to the task, its own identity, and a record of each step and who asked. If the AI makes the decision itself, that is Decide: it needs what any delegate needs, including a defined purpose, a named owner, approval before use, checks on each action and evidence for every decision.

      Ask is not risk freeSensitive data can still leak, and people can act on a confident wrong answer. These are policy and training issues, not reasons for heavy controls.
      Assist can quietly become DecideWhen someone approves hundreds of AI outputs a day without really reading them, the review is a formality and the AI is effectively deciding.
      Act is the common blind spotPermission covers the task, not each step the AI chooses, and an agent can be misled by what it reads along the way. On a member of staff's login, everything it does looks like that person in every system and every audit log. Most organisations have more of these than they realise.
      5. Why deciding is different

      Straight through processing, past the exception

      Rules based automation already handles the simple cases. It stops at the exception: the case the rules did not anticipate. Those go to a queue and wait for a person. That queue is where cost, delay and customer frustration build up.

      Today
      ExceptionthenQueuethenPersonthenDecision
      With governed AI agents
      ExceptionthenAgentthenDecision plus evidence

      AI can handle many of the cases rules cannot, which is why it can take straight through processing further than rules ever did. But this is the Decide end of the spectrum. The reason to keep a person on every exception only falls away once every agent decision is bounded, attributable and reviewable. Accountability is the precondition for autonomy.

      Exceptions straight throughShould rise as agents take on cases that used to queue
      Exceptions needing approvalShould fall as confidence in the controls grows
      Decisions overturnedEvery case where a person disagreed, tracked and learned from

      A governed agent can still be wrong. The difference is that it is bounded, attributable and reviewable, so a wrong decision is contained, traced and corrected.

      6. For business leaders

      Questions worth asking about your own AI use

      Using AI is not one thing. The control you need depends on whether AI is giving advice or exercising authority.

      1. For each way your teams use AI, does a person see and own every output before it takes effect?
      2. Where AI is doing things in your systems, whose login is it using, and does anyone know?
      3. Which processes would you most like AI to decide rather than assist, and what is stopping you today?
      4. Where people approve AI outputs, are they genuinely reviewing them, or approving out of habit?
      5. Are you spending on heavy controls for AI that only gives advice, or relying on light controls for AI that is really making decisions?