Let AI Draft, Let People Sign: Automation That Fits How Enterprises Approve Work

In February 2024, the British Columbia Civil Resolution Tribunal ordered Air Canada to pay a passenger CA$812.02, including interest and fees, in Moffatt v. Air Canada. The airline's website chatbot had told him he could apply for a bereavement discount after his trip, which was never the airline's policy. Air Canada argued that the chatbot was responsible for its own actions. The tribunal member called that "a remarkable submission" and held the airline responsible for everything on its website, chatbot included.

For an enterprise planning a customer-facing assistant, the case raises a practical question: who checks the promises it makes? Assign that responsibility while designing the workflow, when the team can still change what the assistant is allowed to send.

An Approval Chain Is Already a Specification

Large organizations look slow from the outside because work passes through people who have to approve it. A purchase order goes to a manager and then to finance. A public report goes to the team that wrote it, then legal, then an executive. The steps may reflect financial controls, legal obligations or earlier mistakes. Some will also be outdated, so check their purpose before reproducing them in software.

That chain is the most useful document you have when you automate. It tells you which decisions carry risk, who owns them, and what evidence each owner looks at before saying yes. Those details give the automation a starting specification. The team still needs to check whether the existing controls cover the new system and how reviewers will inspect its output.

For a reporting platform, start by mapping who approves what and what evidence each reviewer checks. Define the AI’s job against that map: draft a section, gather supporting records and leave approval with the person who owns the report.

Sort Work by What a Mistake Costs

For each task, the useful question is how much a mistake would cost and how easily it could be undone.

Internal tags and first drafts can be candidates for automatic processing when errors are contained and the output is clearly provisional. A medical summary or confidential search result can carry much greater risk, so classify the use of the output as well as the task name. Payments, emails to customers, published figures, permission changes and deleted records are expensive or impossible to undo, so they should wait for a named person every time, at least until the system has a long record of getting them right.

A customer-service draft shows how the boundary works. The assistant can locate the refund policy and propose an answer. If the policy is ambiguous or the response promises an exception, a person with authority over refunds should review it before it reaches the customer.

Reviewing Has to Be Faster Than Doing

A human approval step protects you only if people actually review. When checking the AI's output takes as long as doing the work, reviewers start approving without reading, or they quietly go back to doing the work themselves and the automation becomes a cost with no benefit.

In METR’s early-2025 trial, 16 experienced open-source developers completed 246 tasks on familiar projects. AI access increased completion time by 19%, although participants later estimated it had made them about 20% faster. This small study of particular tools and repositories is a reason to measure the complete workflow, including review; it does not establish how every AI-assisted team performs.

An approval screen should show what changed, where each number came from and which checks failed. Group similar items only when they share the same risk and acceptance criteria. A standing rule needs explicit limits and an owner who can revoke it.

The Process Has to Exist on Paper First

Most enterprise processes live partly in people's heads. The checklist says "verify supplier data," and the senior analyst knows that means comparing three spreadsheets and calling one person in procurement. An AI system can follow only the version that is written down.

Writing the process down exposes missing ownership before automation makes it harder to see. For supplier verification, name the three source records, the fields to compare and the person who resolves a mismatch. “Check the supplier” leaves all three decisions to whoever happens to run the task.

The resulting procedure can guide the automation and help a new employee understand the same task. Keep it with the workflow so it gets updated when the source records or responsibilities change.

What the Approval Log Tells You Next

After a few months, count approvals by type and inspect the edits. A run of accepted drafts is evidence to examine, alongside rare failure cases and the cost of a mistake. Move a task to automatic execution only after testing those cases, defining a rollback path and assigning someone to monitor it.

Schedule that review with the workflow owner. For each proposed change, record the evidence, the new limits and who will respond if the change causes problems.