Four Questions That Decide Whether AI Earns Its Place in a Workflow

When 27 radiologists read mammograms alongside a simulated AI advisor in a 2023 experiment, incorrect suggestions cut diagnostic accuracy at every experience level. The least experienced readers fell from 79.7% correct categorisations to 19.8% when the AI suggestion was wrong, according to research by Thomas Dratsch and colleagues. The finding anchors a newly published four-question audit designed to help product teams decide whether an AI feature should be built at all.

The audit, published by Smashing Magazine, opens with a health insurer case. A product team built AI that wrote full benefits-appeals determinations for specialists. The technology worked as specified, yet specialists abandoned it, and those who kept using it took longer per case than those who ignored it. A short evaluation found the cause: reading an AI-written determination carefully enough to sign it took roughly as long as writing one, plus the specialist had to check for mistakes they would never have made themselves. A rescoped second version — surfacing relevant policy provisions and prior decisions, leaving the writing to the specialist — removed the search step and cut review time. The technology was identical in both versions; only the task division changed.

What is the break-even line?

The article's central claim: every interface change costs the user something before it pays anything back. It grounds this in Nelson Cowan's 2001 review of working memory research, which put the limit at about four items held in mind at once. An AI feature in enterprise software demands more than a redesigned menu because it changes what the user does all day. Hand synthesis to AI and the user becomes a checker; give it the recommendation too and the job shrinks to signing off.

The article defines the break-even line: "An AI feature earns its place when what it returns beats what it costs the user to adjust, for one user doing one task." Returns come in three forms — effort removed, errors caught that the person would have missed, and consistency across people doing the same work. A 2025 study by Sophia Ackerhans and colleagues found that physicians asked to sign AI-assisted decisions reported a more serious threat to their professional identity, and explanations of the AI's reasoning raised both trust and that threat together.

What are the four questions?

• Question 1: Does this reduce total effort for this user? Fewer clicks prove nothing on their own. Pre-filling a loan-number field trades a low-effort action for high-attention checking. The same AI-drafted risk summary gains a junior analyst a starting point, while a senior analyst spends four minutes reading it and learns nothing. Design condition: name the segment served and give other segments an exit.
• Question 2: Does this beat the current method? AI wins on ambiguity, scale, speed, consistency or catch rate. An HR succession tool that surfaces internal candidates by unusual skill combinations wins on ambiguity and scale; drafting the two-minute notification email loses to the method in place. Design condition: remove AI from steps the current method already handles well.
• Question 3: Can the user catch the errors? Raja Parasuraman and Dietrich Manzey's 2010 review of decades of human factors research showed people defer to automated output even when it is wrong, across aviation, medicine and process control. A trip planner suggesting a museum on its closed day produces an invisible error; a flight departing before its connection lands is visible on screen. Design condition: high-consequence, low-visibility error surfaces ship with provenance display and verification prompts, or do not ship.
• Question 4: After a year, can the user still tell when the AI is wrong? The Dratsch experiment showed experience softens but does not prevent automation bias — the most experienced radiologists, averaging 15 years, still dropped from 82.3% to 45.5% accuracy under wrong suggestions. Design condition: build a named mechanism such as periodic blind review into the workflow itself.

How long does the audit take?

The article prescribes a 60-minute version: state the task specifically (10 minutes), answer the questions with someone who does the work (25 minutes), map the error surface across three likeliest error types (15 minutes), and write conditions into the brief (10 minutes). It cites Everett Rogers's 2003 concept of pro-innovation bias — the assumption that new capability means progress before anyone verifies the benefit — as the pressure the audit counters.

Run on AI meeting summaries, the audit reaches a different verdict: total effort drops for non-note-takers, summaries beat half-listening typists on completeness, but low-visibility misattribution errors require a linked source timestamp behind every action item. Question 4 needs no mechanism, since nobody loses a skill worth protecting by giving up meeting notes.

The article closes on a fixed constraint: the attention a person can spend in a working day stays fixed no matter how good the model gets, so teams should write down what a feature will cost the worker and which of the three returns pays for it — before the next feature reaches design.