The Four Levels of Leverage

Assistance

Assistance is where almost everyone starts, and where most organisations still are. A person does the work and uses AI to get through it faster: a first draft of an email, a summary of a long policy document, a rewrite of technical wording into plain language, a starting point for research when the page is blank.

The person owns the output completely. They read it, correct it and decide whether it goes anywhere. That is exactly as it should be at this level, and it is also why Assistance changes nobody's cost base. The time saved is real but scattered across individuals. It rarely shows up anywhere a finance director would recognise.

Assistance is not a failure, and it isn't a stage to rush through. It is where people build the one capability every later level depends on: the judgement to evaluate an AI output rather than simply accept it. In our readiness diagnostics this is the capability most often missing. People can be enthusiastic, confident users and still believe things about the tools that aren't true, such as the idea that a workplace assistant can see every document in the organisation. A team that can't spot a weak or wrong output at Assistance will not be able to supervise AI at Delegation.

What it takes to do Assistance well is less about features and more about how people use them. The shift from typing search-style queries to writing proper briefs, which I covered in Two Boxes. Two Different Brains., is usually the single biggest improvement a team can make here.

Repeatability

Repeatability is the first level where AI starts to behave like an organisational asset rather than a personal productivity trick. The question changes from "how do I use this?" to "how do we do this, every time?"

At Repeatability, a team agrees a standard way of using AI for a recurring piece of work, and that standard produces consistent results regardless of who runs it. Examples include a shared prompt set for the monthly management report, so it reads the same whether the analyst or their manager prepares it, or a checklist the whole team runs over a contract before it goes to legal. The value here is not that the work gets faster, although it often does. The value is that the quality stops depending on who happened to be available.

Repeatability changes variance, and variance is expensive in ways that rarely appear on a report. Rework, inconsistent client communication and the quiet reliance on one or two people who "know how to do it properly" all come from variance.

Getting to Repeatability requires three things that have nothing to do with technology: an agreed definition of what good looks like, an owner for the standard, and somewhere the team keeps it so it can be improved over time. Many teams discover at this point that they never had an agreed standard for the work in the first place. AI simply made the inconsistency visible.

Delegation

Delegation is the level most leaders picture when they think about AI transforming their business, and it is where the economics genuinely change. At Delegation, AI owns a bounded outcome from start to finish, and people review by exception rather than checking every output.

An example is an assistant that triages inbound queries overnight, answers the routine ones within agreed limits and escalates anything unusual to a named person in the morning. Another is a documented reconciliation workflow that runs on schedule, with a person reviewing a sample of outputs and every exception. In both cases, hours move from doing the work to supervising it, and cycle time drops because the work no longer waits for someone to be free.

None of these conditions come with a licence, and none of them are primarily technical. They are process and governance questions, which is why Delegation projects so often stall when they are run as IT projects. For South African organisations, decision rights and the audit trail carry extra weight, because you need to be able to account for how personal information was used and who was responsible for decisions made with it.

This is also why the simplest preparation for Delegation is a piece of paper. Pick one process your team runs every week, and write down each step, who makes each decision, what data it draws on, what happens to that data by hand and where the exceptions go. If you can't finish that page, the process is not ready for AI to run, and you have just found the real first piece of work.

Transformation

Transformation is the level where the business does something it previously could not do, or could not do economically. It is not "Delegation, but more of it". Delegation makes an existing process faster and cheaper. Transformation changes what the organisation sells, who it can serve or how it makes money.

To illustrate: a professional services firm that could only afford to give detailed, personalised advice to its largest clients might, with well-governed Delegation already in place across its analysis work, offer a version of that service to thousands of smaller clients at a price point that was never viable before. The service is new. The customer base is new. The revenue line is new.

Transformation is rarely present in the organisations we assess, and that is not a criticism. It is a starting position. Transformation is built on top of reliable Delegation, which is built on top of consistent Repeatability, which in turn relies on people who have developed real judgement at Assistance. Organisations that announce Transformation strategies without those foundations tend to produce impressive pilots that never reach production.