The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Back to feedPost
Three ways AI products fail that have nothing to do with the model:
I've spent the last year designing conversational interfaces, triage flows, copilots, and agent UIs. The pattern I keep hitting: teams ship a genuinely good model wrapped in an experience that undoes it.
The blank box problem. You give someone an open text field and no idea what to type. "Ask me anything" is the least helpful instruction in software. Users don't want infinite capability; they want to know what this thing is for. The fix is rarely more onboarding copy but rather constraining the first turn.
Errors that read as dead ends. Deterministic software fails loudly and recoverably. AI fails plausibly confidently wrong, or vaguely unhelpful. If your interface treats "I didn't understand" as a terminal, you lose the user at the exact moment they were still willing to try.
Trust spent instead of built. Every confident wrong answer draws down a balance you never explicitly topped up. Products that survive show their work early: where the answer came from, what it's unsure about before they need the credit.
None of these show up in your evals. They show up in churn.
I've just opened a 5-day audit that looks specifically for this class of failure: https://contra.com/s/3ZBRWBck-conversational-ux-audit-for-ai-products-in-5-days?r=PereXDesign
Happy to argue about any of the three
contra.com
Conversational UX Audit for AI Products — in 5 Days by Josue Perez
Read more about Conversational UX Audit for AI Products — in 5 Days by Josue Perez on Contra.
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Related posts
I stopped writing my first draft of any data report myself. Here's why.
Over the past few months, I've started using Claude as the first pass on every client communication — not to do the analysis, but to structure the findings once I already have them.
My workflow is now:
Clean and analyze the data myself in Python/R (this part stays 100% me)
Feed the key findings to Claude and ask him to draft a plain-language summary
Rewrite and tighten it in my own voice, cut out anything generic
The result: reports that used to take me an hour to write up now take 20 minutes, and they read clearer because I'm editing instead of staring at a blank page.
The part AI can't do: determining what finding actually matters to the client's business. That's STILL 100% WITH ME.
What's one part of your workflow you've handed off to AI, and one part you'd never trust it with?
3 mistakes that will eventually make this board useless. 📉
I've reviewed several client dashboards over the years, and the same 3 problems show up over and over:
1️⃣ Too many charts, no clear takeaway — if I can't tell what to do after looking at it, it's decoration, not a dashboard.
2️⃣ Colors with no consistent meaning — red means "unacceptable" on one chart and "category 3" on the next. Pick a system and stick with it.
3️⃣ No comparison point — a number alone means nothing. "Revenue: $12,480" tells you nothing. "Revenue: $12,480 (+18% vs last month)" tells you everything.
Fix these three, and your data stops being numbers on a page — it starts being something people actually use to decide.
What's the most common dashboard mistake you've seen? 👇
