Building AI skills in a transformation team: where to start

Most transformation teams already use AI in some form, whether or not anyone decided they should. Someone drafts a committee summary with a chatbot. Someone asks a tool to explain a paragraph of the Codes. Someone pastes a supplier list into a browser window to tidy the formatting.

Some of this is useful. Some of it is risky, and the supplier list is the example we worry about most. Almost none of it is planned. Building AI skills in a team is mostly a matter of turning scattered, individual use into something deliberate, and the order in which you do that matters. We would start with the rules, then understanding, then practice, and we would treat checking as the skill that holds the rest together.

Start with the rules, not the tools

Before anyone learns a new technique, the team needs to know what it may and may not do. Transformation teams handle sensitive personal information: race, gender and disability status, pay, and the personal details of bursars, learners and shareholders. Under POPIA, race and health information is special personal information, and disability information will usually be treated as health information.

Three rules cover most of the risk:

• No personal, client or confidential company information goes into an AI tool the company has not approved.
• AI output is a draft. Anything that leaves the team, or feeds a calculation, is checked by a person against its source.
• When AI helped produce something, say so to the person reviewing it.

Write these down, agree them with your information officer and IT team, and make them the first session of any training.

Build understanding before technique

People use AI well when they understand roughly how it works and where it fails. A transformation team does not need technical depth, but it does need a few working truths.

The first is that language models produce plausible text, not verified facts. They can state a wrong figure, an outdated rule or a section of an Act that does not exist, with complete confidence. The second is that they know nothing about your company's data unless you give it to them, and what you give them matters. The third is where their strengths lie. They are strong at reading, summarising, reorganising and drafting, and weaker at exact arithmetic across long tables and at knowing what has changed recently. For a team whose work is calculating scorecards from long tables under rules that are amended, that last point deserves attention.

The Codes make a good teaching exercise. Ask a general AI tool a detailed question about a scorecard element, then check its answer against the gazetted text. In half an hour a team sees both the usefulness and the limits for itself.

Practise on real tasks, with safe data

Skills stick when they are practised on the work people actually do. List the team's recurring tasks and choose two or three where AI could plausibly help. Good candidates are summarising a long policy, gazette or verification report into one page for a committee; drafting a first version of the monthly commentary from figures that have already been confirmed; turning meeting notes into a list of actions with owners and dates; and checking a draft document for consistency with the team's agreed definitions.

Notice that none of these asks AI to produce a figure. Practise with fictional or anonymised data until the team is confident, and only then move to approved tools with real data, under the rules above.

Teach review as a skill

The most important AI skill in a transformation team is checking output, and prompt writing comes a long way behind it. Checking means comparing each figure with its source, testing each legal statement against the Act or the Codes, and noticing when a fluent paragraph says nothing specific. These are the same skills a good verification analyst already has. AI makes them more valuable.

Give the practice an owner and clean data

Give one person responsibility for the team's AI practice: keeping the rules current, collecting methods that work, noting where AI went wrong and sharing what was learnt. This need not be a technical person. It should be someone who knows the transformation work well and is willing to experiment carefully.

That person will soon find that AI works best on organised data. A team that keeps clean, consistent records, with agreed definitions and one source for each dataset, will get far more from any tool than a team working from a dozen versions of the same spreadsheet. Structuring a table properly, using consistent identifiers and keeping evidence linked to the record it supports are part of AI readiness, even though they long predate it.

Count it where it counts

Training your employees in these skills may also contribute to your skills development score, depending on how it is structured and who attends. Check how a course fits the learning programme matrix in the Codes before assuming it counts, and keep the evidence your verification agency will ask for, such as attendance records and proof of cost.

Where we can help

Our 2027 training calendar includes practical AI and data courses, built on the same principles as the systems we design: practical tasks, personal information protected and a person checking every result. If you would like to hear more, or want to talk about your team's starting point now, write to us at sales@bsvision.co.za.