From Task to Workflow: Who Drafts and Who Decides
Designing AI-Supported Workflows · 5 min read
An AI tool does not arrive as a workflow. It arrives as a capability, and a leader's job is to turn that capability into a process in which each step has an owner. The most useful first move is plain: break the task into steps and decide, step by step, what the tool drafts and what a person decides. The answer is rarely all or nothing. A tool can prepare a summary, propose a category or draft a reply, while a named person accepts, changes or rejects it. The split is a design choice, and it should be written down before anyone starts using the tool.
Drafting for review is not the same as sending on its own
There are two quite different designs. In the first, the tool drafts wording from permitted records and an employee reviews, edits and sends it. In the second, the tool sends on its own when its confidence score passes a threshold. The second removes the human check from the step that matters most, and the confidence score is the tool's own estimate of itself. A high score can sit on a wrong answer, and it says nothing about how costly the error is for the person who receives it. Where autonomy is considered, treat it as a bounded exception: a low-cost message type, tested on unusual and incomplete cases as well as routine ones, sampled afterwards, with a way to stop it.
Decide what data may enter
Every workflow needs a boundary on its inputs: which record types the tool may read and which it may not. Leave this to individual judgement under time pressure and it will drift. The boundary should be written, kept as narrow as the task allows and enforced in how the tool is connected, not only in a policy note. Data collected for one purpose should not slide into another because it happens to be available. Swiss data protection law lists purpose limitation among its processing principles, and the principle is sound practice anywhere.
Put review where errors cost most
Review is a scarce resource, so place it with care. Ask of each step whether a wrong output can be undone once it has left the building, how much harm it could do to the person affected and whether a later step would catch it. A payment, a filing or a letter to a customer deserves a strong check before release. An internal working summary that is easy to fix deserves a light one. Spreading review evenly dilutes attention at the costly step, and checking only at the end finds errors after the harm is done.
Do not automate judgement in high-consequence decisions
Some decisions deny a service, affect income or concern a child or a patient. In these the tool should prepare information and a person should decide. This is not a refusal to use AI, because the drafting and sorting around the decision still save time. Swiss law adds a specific rule for decisions based exclusively on automated processing that have legal consequence or significant effect: the person must be told, may state their position and may ask for review by a natural person. A notice alone does not meet those duties. Keeping a person as the decider avoids the question for most cases.
Simplify before you automate
Automation makes a process faster, including its faults. If an approval chain has signatures that never change a decision, or a form exists in three layouts, the first job is to remove the waste. Often the process then needs a smaller tool or none. Measuring days saved by automating a poor process counts activity and not outcome.
A worked example
Halden Housing Association receives repair requests by email, phone notes and a web form. The sponsor wants a tool to cut handling time. The team first maps the process into seven steps: receive, extract details, classify urgency, check the tenant's record, decide whether the repair is a landlord duty, book a contractor and write to the tenant. They agree that the tool drafts the extraction, proposes an urgency class for an officer to confirm and drafts the tenant letter for an officer to send. The duty decision and the booking stay with named officers, because a wrong call leaves a tenant without heating and is hard to reverse once a contractor has been sent. Tenant medical notes are excluded from the tool's inputs because a repair letter does not need them. Along the way they notice that two forms ask the same questions, and they merge them before building anything. The result is a smaller design than the sponsor first imagined, and every step has a person who is answerable for it.
As of 24 September 2026.
You can read every lesson without an account. Signing in keeps your place and unlocks the assessment.