The middle manager's job after AI

For the past two years, enterprise software vendors have pitched artificial intelligence to chief executives on a tempting organisational promise: automate routine knowledge work, flatten the corporate hierarchy and trim the middle-management payroll. Global headlines reinforce the pitch every quarter, as technology firms and banks from San Francisco to Singapore announce white-collar restructurings framed as AI efficiency.

Walk into a Kenyan bank branch or operations hub six months after automation goes live, and you will see the opposite happening. When software swallows the predictable transactions, it leaves behind a thick residue of messy exceptions, angry customers and edge-case judgements, making the middle manager's job substantially harder even as headline ticket counts drop. The institutions cutting that layer are not removing cost. They are removing the only people positioned to catch the machine's mistakes before the regulator does.

Pair the 2025 Quarterly Journal of Economics study of 5,172 support agents with the 2023 Harvard and Wharton field experiment involving 758 Boston Consulting Group consultants. Junior workers used AI to climb rapidly toward average proficiency, while experienced specialists gained little speed. More troubling, when the consultants used generative AI on problems sitting just outside the model's competence boundary, their accuracy fell by 19 percentage points compared with colleagues working unaided, because the machine's confident prose lulled them into switching off their own judgement. Inside an African financial institution, the person standing on that fault line between fluent software output and costly operational reality is the team leader or branch operations manager.

Counting the cognitive load of what survives automation.

Look at what disappears from a banking hall when self-service works properly. Routine balance checks, statement requests and standard PIN resets move onto the phone, leaving human desks with the high-stakes tangles: a deceased estate account with competing family court orders, a cross-border trade-finance discrepancy, a merchant settlement freeze triggered by an anti-money-laundering alert.

When an operations director keeps a middle manager's span of control at 15 or 20 direct reports based on old transaction volumes, the supervision mathematics breaks down. Every surviving ticket now demands cross-departmental coordination, regulatory judgement and careful de-escalation. Overloaded supervisors still graded on average handling time react the only way rational employees can: they rubber-stamp whatever recommendation flashes on the screen, while frontline clerks quickly learn that questioning a suspect model output ruins their weekly productivity score. Task dashboards look green right up until an unchallenged algorithmic mistake triggers a central bank fine.

Giving supervisors the authority and skill to challenge the screen.

Managing a hybrid team of people and software requires knowing how different tools break in production. Fixed business rules keep executing faithfully against outdated policy thresholds long after the market has moved. Statistical credit or fraud models stay mathematically valid on past data while missing a sudden supply-chain shock in a borrower's industry. Generative assistants draft authoritative compliance summaries that quietly drop a vital statutory condition. Autonomous scripts with write-access to core ledgers can fire off a string of mistaken fee waivers before a supervisor ever checks the log.

No branch or operations manager can police those boundaries on instinct alone. Supervisors need bounded, auditable permission to pause an automated workflow whenever frontline evidence contradicts the screen, without waiting three weeks for an IT change committee. And often the highest-return upgrade has nothing to do with artificial intelligence at all: cleaning up muddled delegation schedules, publishing one searchable policy manual and raising frontline resolution limits by KSh20,000 clears more managerial bottlenecks than buying a supervisory copilot.

Pricing the hidden managerial tax before cutting headcount.

Trace the hours on a back-office desk where workflow automation frees up 100 frontline staff hours a week. If auditing model exceptions, untangling edge cases, and coaching junior clerks not to trust the screen blindly takes 40 hours of senior supervisory time, the net capacity gain is 60 hours. Because a manager's loaded hourly pay is typically double a junior clerk's, the cash saving shrinks further still. A chief financial officer who books a 100-hour headcount cut upfront leaves those 40 hours of exception oversight unfunded, stripping out the exact layer of institutional experience needed to keep the automated process safe.

Before cutting a single supervisory post in 2027, operations and human-resources leads should rewrite the managerial scorecard so team leaders are judged on first-time resolution quality, verified error detection and complex escalation turnaround, tested each quarter with synthetic files laced with plausible AI errors and conflicting KYC records. By the mid-year staffing review in June 2027, a dashboard showing zero escalations from middle managers should be read as a red flag that supervisors have stopped looking under the bonnet. The hierarchy you flatten this year is the audit finding you answer next year.