In July 2025, the Commonwealth Bank of Australia confirmed it was cutting 45 call-centre roles, tying the reductions to a new artificial intelligence voice bot that the lender said was making it faster and easier for customers to get help. Within a month, the bank reversed the decision, apologized to affected staff, and called the cuts an “error,” admitting it “did not adequately consider all relevant business considerations.” According to the Finance Sector Union, which had challenged the redundancies, incoming call volumes were actually rising, and the bank had been offering staff overtime to clear the queues.
Months earlier, Swedish payments group Klarna had gone through a similar arc. In February 2024, Klarna announced that its AI assistant had handled 2.3 million conversations in its first month, representing two-thirds of all customer-service chats. The company said the tool was doing the equivalent work of 700 full-time agents, had cut resolution times from 11 minutes to under two, and would drive a US$40 million profit improvement that year. By May 2025, Klarna’s chief executive told Bloomberg that “cost unfortunately seems to have been a too predominant evaluation factor” and that “what you end up having is lower quality,” as the company began recruiting human agents to work remotely.
Industry-wide studies capture the same gap between trial demonstrations and live profit-and-loss statements. In August 2024, technology research firm Gartner predicted that at least 30 percent of generative AI projects would be abandoned after proof of concept by the end of 2025, pointing to poor data quality, weak risk controls, rising costs, and unclear business value. A year later, in August 2025, a report from the Massachusetts Institute of Technology’s NANDA initiative drew on 52 interviews, 153 survey responses, and a review of 300 public deployments, finding that 95 percent of organizations reported no measurable profit-and-loss return from generative AI. Even allowing for a modest, self-reported sample, the commercial pattern holds across markets.
Kenyan bank boards are now finalising their 2027 budgets, and many are being asked to fund the move from a promising AI pilot to full-scale operation. In its July 2025 survey of 125 lenders and credit reference bureaus, the Central Bank of Kenya placed 67 percent of respondents in the first two of five stages of AI maturity. Among institutions already using AI, customer service ranked as the third most common use case, with 82 percent of respondents indicating plans to deploy AI in customer support, even though only 56 percent of current adopters had measures in place to make their models explainable. Before releasing 2027 capital, a chief executive and finance director need to know what the automated service costs once it runs on every case, every channel, and every bad day.
Modelling one disputed-payment workflow four ways.
Failed and disputed payments offer a fair test of an automation business case. Customers feel a missing transfer immediately, chase the bank across multiple channels until the money returns, and judge the institution on how fast the balance is restored. According to the Kenya Bankers Association’s 2025 customer satisfaction survey, the share of bank complaints always resolved within two days dropped to 66.5 per cent, down from 75.4 per cent a year earlier.
To test the economics below the income statement,Precursormodelled the disputed-payment workflow at a mid-sized Kenyan commercial bank handling 20,000 customer contacts a month across failed, duplicated, or disputed transactions. The figures draw on published vendor pricing and typical Kenyan banking pay scales so that any finance team can substitute its own numbers.
Under the baseline manual workflow, a front-line customer care officer spends about 10 minutes on each contact. Across the bank, 45 percent of those contacts involve missing funds that must be traced and reversed across the core banking ledger, the card switch, M-PESA, or Pesalink, requiring about 15 minutes from a back-office investigator. Including benefits, supervision, and workspace overheads, staff time is costed at KSh 600 an hour for front-line officers and KSh 1,000 an hour for back-office investigators. Three in 10 customers contact the bank a second time to chase their money, and two in every 100 cases end in an operational error or a goodwill payout averaging KSh 1,000.
In the AI scenario, the bank puts a text assistant on WhatsApp and its mobile app to answer customer queries, look up transaction statuses, and draft reversal requests for investigators.Precursorpriced the assistant at US$0.60 a conversation, or about KSh 77 at an exchange rate of KSh 129 to the dollar. Commercial AI customer-service vendors publicly quote between US$0.90 and US$0.99 per resolved chat, and charge even more for voice bots, so using US$0.60 gives the technology the benefit of the doubt.
Seeing why the three-month pilot flattered the numbers.
In our model, the bank ran a three-month trial on WhatsApp in English, connected read-only to a single payment switch, with the software vendor covering setup. At the end of the quarter, the project report compared the pilot’s KSh 181 cost per contact against the bank’s KSh 251 manual average and claimed a 28 per cent saving.
Two structural blind spots distorted that comparison. First, the customers who opted into the English WhatsApp trial brought simpler problems, with only a quarter of their contacts requiring money to be traced and returned compared with 45 per cent across the wider bank. Handled manually by front-line staff, that easier WhatsApp mix would have cost about KSh 200 per contact, putting the true like-for-like saving during the trial at about 10 per cent. Second, the pilot left out the engineering, compliance, and supervisory bills that arrive only when a system runs at full scale.
Project teams rarely skew those numbers in bad faith. A pilot is built to prove that the technology can work in a controlled setting, and it usually does. A budget has to show that the service pays its way when the hard cases, the telephone calls, and the legacy integrations all arrive together.
Counting what the live production bill adds.
Going live means connecting the assistant to the core banking system, the card switch, M-PESA settlement files, and Pesalink while cleaning three years of ticket data, testing security, and tuning the model for Kiswahili and the mixed-language Kenyan customers actually type. Spread over 36 months, that integration and data work costs KSh 45 million, or KSh 63 per contact. Under Section 31 of Kenya’s Data Protection Act, processing likely to pose a high risk to individuals also requires a data protection impact assessment before it starts, and that legal and technical work belongs in the same budget line.
Once connected, the software needs a dedicated team to run it. The model budgets about KSh 1.05 million a month, or KSh 53 per contact, for three full-time staff (a product owner, a quality analyst, and an engineer) alongside monitoring tools. Supervisors sample one in 10 AI conversations for accuracy, and back-office investigators check every reversal the assistant drafts before money moves. That manual check on reversals is more than an internal precaution. Section 35 of the Data Protection Act gives customers the right not to be subjected to a decision based solely on automated processing that significantly affects them, and a refused payment reversal sits squarely inside that rule.
The assistant also makes mistakes of its own. It can promise a refund that bank rules forbid, misread a duplicated charge on a statement, or route a fraud report to the wrong queue. The model assumes 1.5 percent of digital conversations go wrong at an average cost of KSh 1,500 each, while one in five digital customers still phones the call centre afterwards to check on their money.
Most of the underlying workload stays where it was. Four in 10 contacts still arrive by telephone, which a text assistant never touches. Every case requiring money to be returned still goes to a back-office investigator. Having the assistant prepare the file trims the investigator’s handling time by about a fifth, yet the customer waits almost as long for her money as before because the settlement step that returns it has not changed.
At full volume, the bank ends up paying KSh 384 a contact against KSh 251 under the manual process, adding KSh 2.7 million a month to its operating costs.
Doing the arithmetic of a Kenyan wage.
Why a bot that saves money in Western Europe adds cost in East Africa comes down to the local wage ratio. At a Kenyan bank, a front-line officer’s time costs about KSh 10 a minute. Priced in US dollars, the AI assistant costs KSh 77 a conversation. In a market where a call-centre agent costs a dollar a minute, paying 60 cents for an automated chat buys large payroll savings. In Nairobi, 60 cents buys roughly the same human time the bank was already paying for.
Strip out the KSh 77 vendor charge and every fixed integration and run-team cost from the live model, and the remaining operational steps still save only about KSh 28 a contact against the manual process. That means the assistant would need to cost less than KSh 46 (about US$0.36) per digital conversation simply to break even on running costs, before recovering a single shilling of the upfront integration bill. A bank could build its own assistant on a cheaper foundation model at around US$0.10 a conversation, though it would then need about 115,000 disputed-payment contacts a month to repay its fixed engineering and compliance costs, nearly six times the volume of our modelled bank.
Billing in foreign currency adds a second exposure. Customer-care salaries are paid in shillings, whereas AI vendor charges and the cloud capacity behind them are billed in US dollars. A finance director who signs a three-year AI contract is also taking on a three-year currency position.
Fixing the underlying payment process first.
The most effective alternative uses no language model at all. Instead of deploying a bot to field complaints, the bank runs an automated nightly reconciliation of failed and duplicated transactions across its payment switches and reverses clear-cut failures automatically before most customers notice. Front-line officers are given read access to switch records and the authority to reverse disputed amounts up to KSh 5,000 on the spot, with a second officer checking a daily sample, while back-office investigators handle only the complex cases.
In our model, nightly reconciliation clears 5,000 failed payment cases a month before anyone calls the bank, dropping monthly contacts from 20,000 to 15,000. Front-line calls run slightly longer, averaging 12 minutes instead of 10, because officers are resolving the problem on the call rather than logging a ticket. Only four in 10 of the remaining money disputes ever reach a back-office investigator. Writing the reconciliation scripts, configuring access controls, updating policies, and training staff costs about KSh 12 million spread over 36 months, or KSh 22 per contact.
That operational fix brings the monthly bill down to KSh 2.9 million, which is 42 percent below the manual baseline and KSh 4.8 million a month cheaper than the live AI service. Just as important, most customers hit by a small failed transfer get their money back the same day, directly improving the two-day resolution metric measured by the Kenya Bankers Association survey.
Once the underlying process is fixed, AI can still earn a place in a much narrower role. Using a model strictly to draft case files for the 1,600 complex investigations left each month would save about KSh 160,000 in investigator time and cost roughly the same amount to run. That trade is worth making only if it also shortens the customer’s wait.
Deciding whether to expand, narrow, or stop.
Every bank will find a different break-even point depending on its wage bill, monthly call volume, in-house engineering costs, and customer churn. What should stay constant across every boardroom is the set of cost lines a chief executive and finance director ask to inspect before approving the 2027 budget.
- Expand to full operationonly when three conditions hold together. The fully loaded cost per resolved case, measured across the live mix of channels and complex disputes, beats both the manual process and the cheapest non-AI process fix. The time customers wait to recover their money falls. And the bank owns its data, prompts, and test cases cleanly enough to switch suppliers without rebuilding from scratch.
- Narrow the deploymentwhen the real savings sit in one specific slice of the work, such as answering basic status queries or preparing investigators’ paperwork. Keep that slice, measure its unit cost every month, and stop the rest.
- Stop the investmentif the running cost per case exceeds the manual cost before fixed integration bills are counted, if repeat contacts or refund errors climb, or if the apparent payroll saving comes from making customers wait longer for an answer.
Commonwealth Bank of Australia made the mistake of announcing its headcount savings before counting the live call queues. A Kenyan bank board can avoid that trap with a single request before signing off on its 2027 budget. Ask the project team for the fully loaded cost per resolved case across the live customer mix, set right beside the cost of fixing the underlying payment process first. A pilot report that cannot supply both numbers is not ready to become a budget.
