The Banker AI Still Needs.

A uniform maker in Nakuru, a hypothetical borrower, earns most of her income in three bursts, before each school term. In late November she asks her bank for a top-up loan to buy fabric for January. The bank’s scoring model reads three thin months of M-Pesa inflows as a business in decline. A credit manager with 15 years at the branch knows the school calendar and approves the loan. In a bank that has replaced that manager with the model, the same request is declined in seconds, and the model never learns why it was wrong.

The model can learn the school calendar only if someone who knows it records the exception. That knowledge sits with staff who have 10 to 20 years’ service, in the management and supervisory grades. Banks are investing in AI for the decisions those grades already make, and the grades are the cheapest way to make it work. Kenyan banks’ financial years start in January, so their 2027 headcount is being budgeted now.

The middle layer shrank in 2025

Kenya’s banks employed 41,124 people at the end of 2025, up 5.9 percent from 38,840 a year earlier. The figures are from the Central Bank of Kenya (CBK) Bank Supervision Annual Report 2025, published on 22 September 2026. Clerical staff rose by 2,588 to 14,757, a 21.3 percent increase and the largest in that grade since 2013. Management fell by 224 to 12,183 and supervisory staff by 303 to 8,722. The two grades together went from 21,432 to 20,905, a fall of 527, or 2.5 percent.

CBK’s first survey of artificial intelligence in banking was published in July 2025 and restated in the same report. Half of the institutions surveyed used AI, including 66 per cent of commercial banks. Credit risk assessment was the leading use, at 65 per cent of institutions, and 83 per cent planned to use AI for it. Cybersecurity followed at 54 per cent and customer service at 43 per cent. Only 30 percent had a formal AI strategy, and the main obstacles they named included a shortage of AI-skilled staff.

The 2025 fall reversed 2024, when management grew 8.2 percent and clerical posts fell 5.5 percent, a decline CBK linked to automation. CBK’s figures give no reason for the 2025 change. Managers and supervisors hold the authority to approve credit, restructure arrears and sign off exceptions. Banks are investing in AI for credit decisions while reporting a shortage of AI-skilled staff, and they ended 2025 with 527 fewer of the people who make those decisions.

Experience is what the machines copy.

Employment of 22- to 25-year-olds in the most AI-exposed US occupations fell 13 per cent in relative terms between late 2022 and July 2025. The finding comes from a Stanford Digital Economy Lab study of payroll data, published in August 2025. Older workers in the same occupations held their jobs or gained. Losses were concentrated where AI automates tasks, and employment kept growing where AI supports the people doing them.

A study published in The Quarterly Journal of Economics in 2025 tracked customer support agents given a generative AI assistant. They resolved 14 percent more issues an hour on average. Novices improved by 34 percent, and the most experienced agents barely changed. The authors found evidence suggesting the assistant passed the practices of the most able agents on to newer ones.

Both studies use US data. The mechanism in the second depends on a firm’s own experienced staff, and every bank has them. A bank that cuts those staff before its tools have absorbed their practice leaves the tools less to pass on.

Each kind of AI needs a different checker.

Rules engines run fixed logic, such as anti-money laundering thresholds and credit policy limits. They fail by going stale. Their checker knows which exceptions are legitimate.

Predictive models score risk, as in credit decisions, the leading use in CBK’s survey. They fail by drifting, and by missing what the data leaves out, such as the uniform maker’s school calendar. Their checker reviews overrides and knows when the model is wrong.

Generative models draft credit memos, customer letters and call summaries. They fail by writing plausible errors in faultless grammar. Their checker must know the right answer faster than they could write it.

Agents act on their own, reversing payments, sending collection messages or changing limits. They fail by repeating a wrong action many times before anyone notices. Their checker sets what the agent may do and handles what falls outside it.

Each checker needs years of seeing which exceptions are real, the experience held in the management and supervisory grades.

Someone still has to carry the blame.

A generative tool can draft a dozen versions of a restructuring proposal before the mid-morning tea break. Choosing the one that will survive the credit committee takes someone who has sat on it, and who will be asked about it if the loan turns bad.

A model cannot be summoned to the board risk committee, questioned by CBK’s examiners or served a show-cause letter. When a loan book sours, the bank needs a person who can say the decision was theirs and explain why they made it.

A contractor does not restructure a KSh 40 million facility because a model sent a courteous reminder. Borrowers stay with the manager who kept a promise when calling in the loan would have been easier. A model risks nothing when it is wrong, and its licence fee falls due either way.

Carrying the blame, choosing among options and keeping promises fall to the same grades that check the machines. If machines make options cheap, judgement becomes the thing a bank pays for. Kenya’s banks moved the other way in 2025, adding 2,588 clerks and losing 527 managers and supervisors.

The boring option costs less.

Where a bank plans to run branches with fewer credit managers because a scoring model will make their decisions, the cheapest non-AI alternative should be tested first. Experienced managers write down their exception rules, such as which sectors are seasonal and which cash cycles only look like distress, and juniors apply them as a checklist. The cost is a few weeks of senior time. The same rules can then become inputs to the model, so the uniform maker’s school calendar reaches every branch.

Cutting the manager saves a salary. Wrong declines and approvals then have to be found and corrected. Generative and agentic tools need a reviewer who knows the right answer. Borrowers who are wrongly declined, like the uniform maker, take their business elsewhere. Where the model is licensed, a vendor that withdraws or reprices leaves the bank needing people who can decide without it.

Kenya’s Employment Act, 2007 sets the minimum price of the cut itself. It requires at least 15 days’ pay for each completed year of service, a month’s notice or pay in lieu, and accrued leave. It also requires selection on seniority, skill, ability and reliability. For a hypothetical manager with 15 years’ service on KSh 250,000 a month, severance is at least 225 days’ pay. At KSh 8,333 a day, that comes to about KSh 1.9 million. Notice adds KSh 250,000, and collective agreements can raise both.

Replacement failed at Commonwealth Bank.

Commonwealth Bank of Australia employs more than 55,000 people. In July 2025 it told 45 customer service staff that an AI voice bot would take over their work. On 21 August 2025 it reversed the decision and apologized. A spokesperson said its assessment had not adequately considered all relevant business considerations and “the roles were not redundant”. The Finance Sector Union said calls to staff had risen, overtime had been offered and team leaders were answering phones.

Running a bot alongside the team for a quarter, and counting the calls that still need a person, measures the gap before anyone is dismissed.

Klarna, the Swedish payments firm, said in early 2024 that its AI assistant was doing the work of 700 agents. In May 2025 its chief executive told Bloomberg that cost had weighed too heavily and quality had suffered. The firm began hiring people for customer service again.

DBS, Singapore’s largest bank, chose retraining. In February 2025 it said it would let about 4,000 contract and temporary roles lapse over three years as AI took on their work. It planned about 1,000 new AI roles and earmarked about 13,000 staff for retraining. Its chief executive at the time, Piyush Gupta, said that for the first time he was “struggling to create jobs”.

Redeployment needs jobs to move people into. The checker jobs for rules, models, generative tools and agents can be defined before a bank buys the tools.

Three changes carry an owner and a test.

The chief people officer should list, before the 2027 headcount plan is signed, every decision a mid-career grade owns and every AI tool that touches it. The cost is a few weeks of staff time. The result is a decision register the board risk committee reviews by March 2027, with no grade cut where a tool depends on its judgement.

The head of credit should log every model override with a reason and check those loans after six and 12 months. The cost is one field in the loan system. If overridden loans perform as well as the model’s approvals, the overrides add value the model lacks.

The chief executive should allow no redundancy justified by AI until the tool has run in parallel for a quarter. The cost is one quarter’s patience. The result is a measured count of the work the tool cannot do.

Three dates will show the choice.

CBK says it is drafting AI guidance, and 93 percent of the institutions it surveyed asked for it. Watch whether that guidance names a person accountable for each model decision. From March 2027, banks’ annual reports for 2026 will show staff numbers and staff costs. The CBK Bank Supervision Annual Report 2026, due around September 2027, will show whether managers and supervisors fell for a second year.

If they did while AI use climbed, the banks will have kept the models and lost the people who know when the models are wrong. The first sign will be a uniform maker in Nakuru, declined in seconds in November, buying her fabric somewhere else. No CBK table counts that loss.

Disclosure. FINTAK, which publishes Precursor, opens a talent exchange for verified fintech professionals in November 2026.