Skip to content

Architecture-first platforms for startups and growing organisations

Estimate a programme

Automation · 28 July 2026 · 10 min

Agentic AI for operations — without surrendering the record

Multi-agent systems can take the tedium out of how organisations actually work. They must not take the accountability with it.

The request, the approval, the invoice, the cashbook: Nigerian organisations are still documentary operations. That is not an insult. It is the reason a founder or a finance lead can still be asked, years later, who approved a payment. Any automation that cannot survive that question should not be plugged in.

Agentic systems are well suited to the tedium around that work — classifying inbound requests, drafting first responses, reconciling columns, spotting omitted codes, preparing the dockets an officer will actually sign. They are poorly suited to being the officer.

Human in the loop is not a slogan

A human-in-the-loop design means a named person disposes. The agent proposes, with the evidence it used sitting next to the proposal. Maker-checker is not a UI flourish; it is how money is supposed to move.

We refuse designs where an agent posts to a ledger, sends a customer a determination, or changes an access right without a reconstructable human act. Speed that cannot be audited is not efficiency. It is a future inquiry.

Start where the errors already live

The highest-yield automations in our experience are not chatbots on the website. They are the reconciliation, the coding, the completeness check, the routing of a request to the right desk. They live in the operations that already hurt — and they leave a trail.

Related practices

More notes

Next step

Bring us the hard programme.

If the work needs architecture, security, and a team that will still be there at go-live, we should talk. Discovery conversations are with a senior architect — not a queue.