At the Finance Indaba that took place at the Sandton Convention Centre on 22 October 2026, Naeem Seedat, group executive of digital strategy and venturing at Telkom, explained why AI investments often fail when they are judged with tools built for traditional technology.
During a session titled Piercing the veil: Why most AI business cases fail and what finance leaders must do differently, Telkom’s Naeem Seedat said appraisal methods must change before money is committed because AI behaves differently from earlier technology.
Naeem kicked off the session by asking audience members to raise their hands if their organisation uses AI. While almost every hand went up, only a few attendees said AI had cut their total costs, fewer even said that it had grown revenue, and only a handful said AI had done both.
This mirrors a PwC survey of about 4,454 CEOs in 95 countries, in which slightly over a quarter reported that AI had cut costs and about three in 10 said it had raised revenue. Only 12 percent saw both, while 56 percent of respondents said they saw no change to the business.
Naeem warned that Boards will look to finance for answers. "When the pressure mounts and the money starts flowing, they don't go to IT; they come to finance," he pointed out.
Understand what you are paying for
Naeem said understanding the technology is the starting point. “AI is not magic. It’s very smart maths, research and engineering,” he said, adding that these three concepts drive the business case.
The first one is tokens. AI reads and writes in small chunks of data called tokens. Users pay for the tokens going in and coming out and the output costs more. Cost therefore moves with volume rather than the number of licences. The price per token is falling, but rising volume can hide a growing bill.
The second is deterministic versus probabilistic systems. A spreadsheet or general ledger gives the same answer every time. Most AI models give the most likely answer, so it can also be confidently wrong. Naeem urged leaders to ask whether a problem needs certainty or a good estimate, which can prevent expensive over-engineering.
And the third is compute. This covers the chips, data centres and power behind AI. Organisations can own it, rent it or sit in between and that choice decides where costs appear in the financial statements. Owning moves costs below EBITDA (earnings before interest, taxes, depreciation and amortisation) into depreciation, while renting keeps them in operating expenses. Steady, high use favours owning, while spiky demand favours renting.
Where traditional appraisal breaks down
Past methods worked for physical assets, traditional IT and – with some adjustment – cloud. Naeem said AI pushes them past their limits in three places.
The first is cost. Usage does not grow in proportion. "If you double the agents, you're not going to double the cost. Chances are you can quadruple it, even six times your costs," he said. Agents also “burn tokens based on what they want to do, not what you have asked them to do”.
The second is benefit. Because AI answers are probabilistic, people must check the work. This adds cost, which means staff cannot simply be removed and requires a manual fallback if the service is switched off.
Third, is an asset. AI can appear as an asset, an operating expense, or a feature inside a product, where it can quickly erode margin.
"We are appraising a metered probabilistic capability with a toolbox that we built for licensed predictable ones," said Naeem.
While everyone can buy the same models, Naeem said he suspects the difference lies in planning. He believes the organisations that see results budgeted for consumption, redesigned the business before introducing AI and made each pilot earn its next round of funding instead of releasing all the money upfront.
Above all, they named the return before funding the project. “You have to name the business value before you release the money,” he concluded.
