One brief, three dashboards. Only one was built for the CFO.
Same prompt to Claude and Lovable. Same brief to me. The AI tools built a thorough audit tool. They missed who was using it.
A CFO, not an auditor.
Elena is a CFO. She opens this between meetings to check where the quarter's audit stands before a 9 AM status call. She is an expert, but her expertise is finance and decisions, not working a findings register line by line. An audit manager lives inside the detail. She does not.
The brief was one requirement. In ten seconds she has to know what needs her this week and what is running fine without her. The trap is obvious only if you know the user: an audit dashboard can be impressively complete and still be the wrong tool for a CFO.
I gave that brief to two AI tools and designed it myself. Here are the three results.
Five choices, all about altitude.
None of these were aesthetic. Each one came from one fact: Elena is an executive, and an executive's screen is about decisions, not records.
Lead with the decision, not the data.
The screen opens with a one-line status and the few items that need her this week. Not 89 findings, not a transaction count. A CFO arrives with one question, what needs me today, and the layout answers it before anything else.
A two-minute brief for a ten-second scan.
The AI engagement brief is a short summary she can take before the 9 AM call, framed as the handful of decisions the team needs from her this week. It respects that she is between meetings, not sitting at a workstation ready to read.
Density calibrated to a CFO, not an auditor.
Five process areas, plain status pills, an on-time figure, an issue count. Enough to judge health in a glance. I deliberately left out the finding IDs, transaction volumes, and confidence percentages an executive never cites. Restraint was the work.
Status as a verdict, not a worksheet.
Each process area resolves to one Overall column: on track, or needs attention. She reads the exceptions, not the rows. The grid does the triage so she does not have to.
A calm register for a high-stakes tool.
One accent for action, a single large number for the next milestone, room to breathe. The screen should feel composed, because the person reading it is signing off on financial statements under a deadline. Confidence, not clutter.
Both tools were genuinely good.
This is not a strawman. A year ago neither tool would have produced a credible audit dashboard. Both did, and Claude's in particular is impressive.
- Claude built a real audit tool. Severity-ranked queue with finding IDs, every AI finding sourced to a document and version with a confidence score, cross-document checks, recurrence history. As an instrument for an audit manager, it is excellent and I would not redesign most of it.
- The information architecture is sound. Attention, engagement status, AI analysis, quick links. Both tools reached for the right regions in a sensible order.
- The visual systems hold up. Type scale, spacing, component states are coherent. Lovable's serif headers and numbered sections are genuinely tasteful.
- They spoke the domain. Engagements, issues by file version, controls testing, sign-off readiness. The vocabulary is correct, so the prompt was understood, not just pattern-matched.
On craft and completeness, the AI is there. The gap is somewhere else entirely.
Built for the analyst. Used by the CFO.
Every gap below is the same gap seen from a different angle. The tools designed for the person who works the audit, not the person who has to make a decision about it in ten seconds.
- Wrong altitude. Claude surfaces 89 findings, 18,440 transactions referenced, finding IDs, confidence percentages. That is the audit manager's screen. A CFO does not work findings, she works decisions, and the tools optimized for completeness instead of her job.
- Density without prioritization. More data is not more clarity. A seven-row engagement table with budget actuals, aging, and ratios is a lot to parse while scanning between meetings. The one decision she owns is in there, weighted the same as everything around it.
- Wrong reading pattern. A CFO scans for exceptions in seconds. Both tools assume someone who will sit and read a dense table top to bottom. Neither built the layer that says, here are the three things that need you, above the detail.
- Comprehensive, not briefed. The thing a CFO most wants before a status call is a short summary of what changed and what needs a decision. Claude can generate that content, but neither tool designed the brief as a first-class part of the screen.
- No editorial judgment. Designing this meant deciding what to leave out, and that is exactly what a generator will not do unprompted. Both tools added until the dashboard was thorough. Neither subtracted until it was right for Elena.
Generating an artifact and designing for a user are different jobs.
Claude produced a sophisticated audit tool. Lovable produced a tidy one. Neither produced the CFO's view, because the gap was never craft, where the AI is now genuinely strong. The gap is knowing the user.
Knowing that Elena is a CFO and not an audit manager. That she scans for exceptions instead of reading rows, that she wants a briefing before a call, that the right move is to cut 89 findings down to the few decisions she actually owns. That judgment does not come from a prompt. It comes from understanding one specific person and having the nerve to leave most of the data out.
The hard part wasn't adding data. It was knowing what to leave out.
AI gave me a strong, complete first draft in minutes. It did not give me the restraint to design down to the one person this screen is for. That part is still the job.
What I'd do differently.
My insights panel reused one finding across several check types as placeholder content. Before this ships it needs real per-check copy, the kind of detail Claude actually handled well. The right version of this borrows the AI's rigor on content and keeps my discipline on hierarchy.