South Africa’s AI Problem Is No Longer Adoption. It’s Proving Value.
A claims handler opens a repair quote, a claim form and three emails from the same customer. An AI assistant summarises the file in seconds. The handler still has to find the policy number, check the quote and put the case back in a supervisor’s queue. By Friday, the insurer can count every summary the team generated. Can it tell whether customers got a decision any sooner?
That question matters in corporate South Africa. In October 2026, World Wide Worx and Altron Digital Business surveyed 100 decision-makers at South African organisations with more than 200 employees. In that sample, 93% reported using AI in some form. Yet 73% lacked tools to assess its value, and 71% could not quantify any cost savings. Giving staff access is now common. Finding out what changed for the business takes a different kind of work. SA AI Index 2026 findings
What does the AI dashboard actually tell you?
An executive reviewing a pilot needs three kinds of evidence. Each answers a different question.
| Measure | What it tells you | What to count in claims |
|---|---|---|
| Activity | Did people use it? | Licences assigned, active handlers, prompts per case |
| Output | What work got done? | Files checked, cases resolved, minutes of handling time released |
| Outcome | Who benefited, and by how much? | Customers receiving decisions on time, overtime spend avoided, fewer incorrect payments |
These numbers reach the boardroom at different speeds. A vendor can count prompts on day one. To find out whether a claimant received a correct decision on time, someone has to join the claims record to the customer notification and check whether the case reopened. The licence dashboard fills that reporting gap by default. It can even point in the wrong direction. A handler who needs five attempts to extract one policy number looks more active than a colleague who gets it on the first try.
Four ways a pilot loses its business case
The team buys a tool before finding the delay. “We need an AI assistant” leaves procurement with a specification and operations with an unanswered question. Take ten overdue claims and trace where they waited. If most sat in an approval queue after the paperwork was complete, changing who can approve routine cases deserves a trial of its own.
Nobody owns the result. IT can deliver a reliable service while customers still wait a week. Give the head of claims responsibility for turnaround, quality and staff adoption. Bring finance in before the pilot to agree what counts as a saving and where it would show up.
The baseline arrives too late. A claim of “60% faster” needs a starting time, a defined task and a comparable set of cases. Time a representative sample before rollout. Include the difficult files, reopened cases and work people do outside the claims system.
The old workflow survives intact. Asking handlers to paste every document into a chat window and copy the answer back adds another handover. Give them the extracted details inside the claim, with a link to the source. Then retire the duplicate capture step. That change gives the team a specific piece of work it can measure.
The scorecard to agree before funding
We’d ask the executive sponsor to put these seven answers on one page. If finance and operations disagree about a measure, the cheapest time to find out is before anyone builds the pilot.
| Field | Question for the sponsor |
|---|---|
| Business objective | Which delay, cost, service failure or missed sale warrants investment? |
| Baseline | What does one completed case cost today, how long does it take and how often does it go wrong? |
| Target outcome | By day 90, what improvement would justify the spend, and which quality measure must hold? |
| Accountable owner | Which executive can change the workflow and answer for the result? |
| Adoption measure | Which staff and cases are eligible, and what share actually use the new workflow? |
| Risk controls | Who may access the documents, which decisions need human approval and how does the team revert to manual handling? |
| Value review | On what date, and against which thresholds, will we scale, redesign, pause or stop? |
Follow one claim all the way to the customer
Consider a hypothetical motor insurer. A model reads the repair quote and claim form, extracts the registration number and quoted amount, and links each field to its source. Software checks that required fields are present and compares the registration against the policy record. A mismatch sends the file to a handler. An authorised assessor still makes the settlement decision.
Suppose the team agrees the following targets. The figures are illustrative.
| Measure | Baseline | Day 90 target |
|---|---|---|
| Average hands-on time per completed claim, including rework | 30 minutes | 20 minutes |
| Operating cost per completed claim, including review and technology | R180 | R155 or less |
| Claims returned for correction | 8 in 100 | 5 in 100 or fewer |
| Median time from receipt to customer notification | 4 working days | 3 working days |
| Customers who have to chase an update | 18 in 100 | 10 in 100 or fewer |
At 600 comparable claims a month, ten minutes released per claim would free 100 staff hours. Finance can book an overtime saving when the overtime bill falls. If handlers use those hours to clear a backlog, report the extra cases closed and how long those customers waited. The same hour can only be counted once.
Record which handlers used extraction, which fields they corrected and when the approval route changed. Where volumes allow, compare teams using the redesigned route with and without extraction on a similar mix of claims. Otherwise, report the combined effect of the model, the workflow change and staff uptake. That gives the next branch an honest account of what it would have to repeat.
Use the first 90 days to make an investment decision
Days 1–30. Select and baseline. Shortlist three workflows with a material cost or service problem. Give each an executive owner and record its current performance and risk conditions. Choose one to test first. Agree the eligible cases, targets and stop conditions with finance and the people doing the work.
Days 31–60. Put the smallest workable change into use. Start with one team and one claim type. Track eligible cases, actual use, corrections, exceptions and time spent in each queue. Keep a manual route open. Review a sample of cases that passed the automated checks as well as the ones handlers escalated.
Days 61–90. Decide where the next rand goes. Compare like cases against the baseline, including review time and running costs. Scale when the economics and controls hold. Redesign if faster capture has simply moved the queue to approvals. Pause if a control fails. Stop if even a well-used workflow cannot deliver enough value to cover its cost.
A small portfolio gives an executive team room to make those calls. At the next AI review, ask the owner of each initiative to bring one completed case, its baseline and the result the customer or the business experienced. That is enough to start a more useful conversation about which initiative earns another rand.