AI & Data
Designing AI workflows people trust
Why people trust some AI workflows and ignore others, and how we design for evidence, review, and a safe exit.
Written by
Sofia Marquez
Design Director
Key points
- Start with the decision the AI improves and the person who owns it. The model comes later.
- Evidence beats a confidence score. Operators need sources. A percentage with a progress bar tells them little.
- Shadow mode and a first-class manual path are how trust is earned before automation acts.
A model can be accurate and still be useless. It becomes useful when a person can see why it made a recommendation, judge whether the evidence applies, and overrule it without the whole workflow collapsing.
That is a design problem, and it is where most enterprise AI pilots quietly fail.
Start with the decision
We begin every AI engagement by writing down the decision the system should improve: approve the claim, route the ticket, flag the asset, draft the response. Then we ask who owns that decision today, what they look at, and what happens when they are wrong.
If the answer is “nobody owns it,” the project needs an operations review before it needs automation.
Evidence beats confidence
A percentage next to a recommendation is almost meaningless to an operator. What they need is the source: the three prior tickets, the clause in the contract, the sensor reading that crossed a threshold. Confidence scores are for model evaluation. Evidence is for decisions.
We design the review surface so the evidence is visible before the action, and the action is reversible. A recommendation that cannot be traced is a rumour with a progress bar.
Ship in shadow mode
The most reliable adoption tactic we have found is boring: run the model alongside the existing process without letting it act. Compare its recommendations with what the team actually did. Show the disagreements to the team and ask which side was right.
Shadow mode turns a model launch into an operational conversation. It also produces the evaluation set you need for the next version, labelled by the people who will use it.
Build the exit
Every automated path needs a manual path that is just as usable. People need a first-class route for the case the model cannot handle. Hiding that route behind a support ticket does not count. The existence of that route is what makes people comfortable letting the model handle the rest.
Trust is the accumulated experience of being right, being explainable, and being easy to overrule. Design for that and adoption follows.
FAQ
Common questions.
Why do enterprise AI pilots fail?
Many pilots optimise the model before defining the decision, the owner, and the evidence an operator needs. If nobody owns the decision or can trace a recommendation, accuracy alone will not produce adoption.
How do you build trust in an AI recommendation?
Show the evidence behind the recommendation, make the action reversible, and give operators a first-class manual path. Trust grows when people can verify the system, overrule it, and see that being wrong is recoverable.
What is shadow mode in AI deployment?
Shadow mode runs the model alongside the existing process without letting it act. Teams compare its recommendations with real decisions, review disagreements, and build a labelled evaluation set before automation is allowed to influence operations.
Keep reading
All postsEngineering
Why we still choose Angular for enterprise platforms
Why we still choose Angular for long-lived enterprise platforms, and what we look for before committing to a framework for years.
Commerce
The real cost of a slow checkout
How small checkout delays turn into lost revenue, and the four numbers we track before changing the design.