AI is good at drafting and checking and bad at being accountable. In any business where a mistake costs money, trust or a regulator's attention, a named person should approve every AI output before it reaches a customer, and the system should record who approved what. Done well, this makes the work faster, not slower.
The accountability gap
When a person writes a letter, everyone knows who is responsible for it. When a model writes it, responsibility goes fuzzy: the vendor points to its terms, the firm points to the tool, and the customer is left with the error. Regulators have noticed. In UK financial services the senior manager regime makes a named person accountable for outcomes regardless of what produced them, and Consumer Duty asks for evidence that outcomes were good.
The simplest way to close the gap is to make the approval explicit: the model proposes, a person disposes, and the system writes it down.
What human approval means in practice
Take a common case: a platform that takes in new products from provider documents. The model reads the document and proposes every term: the underlying, the barrier, the coupon, the dates. Each proposed value is shown next to the passage it came from. A member of staff checks and approves every field. Nothing is saved until they do, and the approval is logged with who, what and when.
That is the pattern I build to: propose, show the source, approve, record. It is faster than typing the terms in by hand and far more accurate, because the person is checking rather than transcribing.
Six design patterns that make it work
- Propose, do not commit. The model never writes to the system of record directly. It produces a proposal that a person turns into a fact.
- Show the source. Every proposed value links to the document, passage or record it came from, so checking takes seconds.
- Make approval cheap. One screen, one click per field or per document, with the differences from the template highlighted. Approval that takes longer than drafting will be skipped.
- Log it. Prompt, model, output, approver and timestamp, queryable later. This is your audit trail and your Consumer Duty evidence.
- Limit what the agent can do. An AI acting on a person's behalf should only be able to do what that person could do, enforced by permission checks, with spend caps so a loop cannot run up a bill.
- Defend against injected instructions. Anything that reads documents or email will eventually read one that tries to give it orders. Treat all read content as data, never as instructions.
Does it slow everything down?
No, and the numbers say so. Approving a good draft is faster than producing one, and far faster than producing one and then fixing the errors a customer found. In Mini Steps, my own goal-planning product, the owner approves actions before they run. Review steps are not where the time goes; rework after a customer finds the error is.
The question is never “human or AI?” It is “which steps does the model do, and which decisions does the person keep?”
When full automation is fine
Not everything needs a signature. Work that is low-stakes, reversible and internal can run without approval once the measurements show it behaves: sorting an inbox, drafting an internal summary, moving a record between two of your own systems. Even then, log it, and keep the approval step for anything that leaves the building.
A checklist for your own AI
- Can the model change a system of record without a person approving? If yes, fix that first.
- Does the approver see the source of each proposed value?
- Does approval take seconds, not minutes?
- Is every approval logged with who, what and when?
- Can the agent only do what the signed-in person could do?
- Is there a spend cap, and does someone get told when it is hit?
- Is content the model reads treated as data rather than instructions?
If you would like this built into your own systems, that is my integration and agent build service, and the first conversation is a 30-minute call.