AI agents belong in the back office first

Most companies introduce AI agents at the most exposed point in their business: the customer conversation. It is the wrong place to start. Customer-facing agents carry brand risk, demand near-perfect accuracy, and fail publicly. Meanwhile the work that actually consumes the organisation sits one floor away, unautomated.

Where the hours actually go

Look at what your operations, finance and procurement teams do in a week. Reading documents and copying fields into systems. Reconciling two records that disagree. Chasing approvals. Assembling the same report from the same four sources. This work is repetitive, rule-shaped, high-volume, and almost entirely invisible in strategy decks.

It is also close to ideal for agents: the inputs are structured enough to verify, the outputs are checkable, and a mistake is caught internally before it reaches anyone outside the company.

Design for the handoff, not for autonomy

The useful question is not whether an agent can complete a task end to end. It is where the agent should stop and hand to a person. A system that processes eighty percent of invoices cleanly and routes the ambiguous twenty percent to a human — with its reasoning attached — delivers more value than one attempting full autonomy and requiring every output to be re-checked.

• Give every agent a confidence threshold below which it escalates rather than guesses
• Log the full tool-call trace, not just the final output, so decisions can be audited
• Let reviewers correct an outcome in one action, and capture that correction as training signal
• Keep a human owner accountable for each workflow — agents assist, they do not own

Measure the thing your CFO already tracks

Cycle time, touches per transaction, exception rate, cost per document. If your agent does not move a number that already appears in an operations review, it is a demo.

Back-office automation has an underrated advantage: the baseline is already instrumented. Finance knows what invoice processing costs today. Operations knows the current turnaround. You do not have to invent a measurement framework to prove value — you inherit one.

Then earn the right to go customer-facing

Teams that start internally arrive at customer-facing AI with working evaluation harnesses, a tested escalation model, real latency and cost data, and an organisation that has learned to supervise a probabilistic system. That is a far better position from which to put an agent in front of a customer than a greenfield pilot and optimism.