Can smaller financial institutions share the cost of AI?

A Tier-3 commercial bank, a microfinance lender and a deposit-taking SACCO in upcountry Kenya rarely have the balance-sheet muscle to outbid Tier-1 banks for data engineers, enterprise security suites or dedicated GPU clusters, and the global economics of AI are widening that gap by the quarter: the useful models demand capital that only hyperscalers and their largest customers command. Pooling demand through an industry association or a shared service bureau therefore looks like the obvious mathematical fix, dividing a heavy capital bill across ten smaller balance sheets so every member gets Tier-1 capability at a tenth of the cost.

What the consortium brochure leaves out is that sharing software also pools operational fragility, regulatory liability and exit risk. When Britain's TSB Bank migrated five million customers onto a shared technology platform managed by its parent group Sabadell in April 2018, testing and governance failures locked up to 1.9 million customers out of their accounts. In December 2022, the UK's Financial Conduct Authority and Prudential Regulation Authority fined TSB £48.65 million for operational risk and outsourcing governance failures. And as the Financial Stability Board warned in its October 2025 report on artificial intelligence in finance, crowding multiple institutions onto a single third-party stack turns a vendor outage into a sector-wide vulnerability. The warning matters doubly in a market like Kenya's, where a handful of core-banking vendors already serve most of the SACCO sector: concentration that begins as a procurement convenience ends as a financial-stability question. Outsourcing the code never outsources the board's statutory responsibility.

Putting the hidden integration work back into the vendor quote.

When a software vendor pitches a shared AI platform to five Kenyan SACCOs or community banks, the price quote almost always covers only the model layer. In reality, the licence fee is the smallest part of the bill because each participating institution still has to pay for data cleaning, core-banking API connectors, role-based access controls and frontline staff training: costs that scale with each member's own legacy mess, not with the consortium's size.

Different automated tools also create different hazards when stretched across competing institutions. A shared generative policy assistant sounds appealing until you remember that SACCO A and SACCO B run different loan-guarantor rules, dividend policies, and recovery bylaws, so one misconfigured retrieval filter can make the bot quote a rival's interest rate or leak confidential member records across tenant walls. A shared credit-scoring model carries an equally sharp trap: repayment patterns trained on salaried urban teachers in Nairobi will misprice the default risk of tea farmers in Kericho or coastal hospitality workers in Mombasa.

An automated collections script shared across a bureau can trigger simultaneous compliance breaches at five lenders the moment a single regulatory threshold is misprogrammed. Before signing a multi-year AI consortium deal, a careful buying group compares the proposal against a simpler shared utility, a rules-based document repository, a standardised reporting schema or a shared fraud blacklist, because most of the time the basic utility solves four-fifths of the operational headache at a fraction of the governance risk.

Stress-testing the consortium arithmetic when members walk away.

Consortium pricing looks unbeatable in year one and turns punitive in year two whenever the contract shifts volume risk onto the remaining buyers. Walk through the arithmetic of a shared compliance and analytics bureau serving four smaller lenders at a fixed monthly platform charge of KSh400,000, or KSh100,000 per institution before variable compute.

If two members merge with larger rivals or pull out after twelve months, and the master services contract includes a joint-and-several minimum revenue guarantee, the monthly bill for each surviving lender immediately doubles to KSh200,000. Even per-member pricing hides traps unless the contract defines what counts as an active record. Does a SACCO member holding a FOSA current account, a BOSA savings account and an emergency loan count as one person or three billable identities? And how much does the vendor charge simply to hold the dormant and archived files that central bank retention rules oblige members to keep? The arithmetic that made the consortium attractive was built on everyone staying; the contract must be built for someone leaving.

Buying the exit door before moving into the building.

A shared steering committee that meets once a quarter over tea cannot manage a live 3.00 am data breach or a model-drift incident. Shared infrastructure requires a named operational service lead inside each member institution who holds the contractual right to inspect audit logs, veto unvalidated model updates pushed by the biggest member, and cut data feeds instantly. Above all, a consortium contract is only as safe as its exit clause, because paper promises of reasonable migration help mean nothing when a vendor holds your historical credit scores, feature transformations and audit trails inside a proprietary schema and demands millions of shillings in professional-services fees to export them.

Before signing any shared technology agreement in 2027, the chief executive, chief financial officer and risk lead should lock two conditions into the contract: strict single-tenant cryptographic data isolation verified by an independent penetration test, and a mandatory six-month dry-run exit drill in which the institution extracts 1,000 customer records, model decision logs and audit trails and proves an independent system can reconstruct them without the vendor's help. Shared cost is a real prize for small institutions, but only the members who can leave a consortium were ever really its customers. The rest are its collateral.