Framework, Decision Tree

    How I Decide What to Ship When AI Is in the Loop

    A working framework I use, and teach my teams to use, when building products in an AI-native loop. Five questions, in order, from validating the problem to trusting the funnel.

    Author

    Ashwarya Subhluxmi

    Audience

    Design leaders and PMs

    Format

    5-step decision tree

    Stress-tested on

    Little Studio Archive

    Why This Exists

    AI did not change what makes a product good. It changed what makes a team slow.

    For most of my career, the bottleneck on shipping was capacity. Designers waiting on engineers, engineers waiting on specs, specs waiting on review. AI removes most of that latency, and what gets exposed underneath is the part that was always the real work, deciding what to build, for whom, and on what evidence.

    This framework is what I use to keep that work honest. It is five questions, in order, that I run myself and run my teams through before we let speed turn into motion without progress. None of them are about AI. All of them are sharper because of it.


    The Framework

    Five questions, in order

    Walk them top to bottom. If a step fails, fix it before moving down. Skipping order is the most common way I see AI-native teams ship beautifully built products that no one needs.

    01

    Have you validated that the problem is worth solving?

    Start Here

    Before a single line of code, before the first prompt, before the Figma file. AI lowers the cost of building, which means the only remaining moat is being right about the problem.

    Yes

    Good. Now define the smallest user flow that proves the value, not the full product.

    No

    Stop building. Spin up a landing page in an afternoon, run two hundred dollars of ads, and watch what happens. Demand is cheaper to test than to assume.

    Lived experience, UX Risk Scanner

    UX Risk Scanner started the right way. Before building anything I tracked the 2.5 billion dollars in dark pattern fines levied in 2025, talked to founders shipping AI-generated landing pages, and confirmed the gap between 5K dollar manual audits and enterprise accessibility scanners. The product followed the demand, not the other way around.

    02

    Can you describe what you're building in one sentence?

    AI is a clarity amplifier. Vague intent produces vague output, and vague output is what most teams ship when they confuse motion with progress.

    Yes

    Start there. That sentence is your first prompt, your first hire, and your first design review.

    No

    Write the brief first. If you cannot defend the product in one sentence, the team will not be able to defend it in a quarter.

    Lived experience, Credit Karma Tax

    Credit Karma Tax came down to one line, a free, mobile-first tax filing experience that a first-time filer can finish in under fifteen minutes. That sentence settled hundreds of downstream calls, from where the refund card lived to which forms we cut from the funnel.

    03

    Are you designing in Figma or building in production?

    AI-native building collapses the design and engineering gap. Every static spec is a translation layer, and every translation layer is a place where intent gets lost and timelines slip.

    Production

    You're in the right loop. Design decisions live where users will feel them, on real data, real components, real performance.

    Figma

    Ask why. If the answer is review, alignment, or fidelity, those problems are better solved in the running product than in a frame.

    Lived experience, UX Risk Scanner

    UX Risk Scanner shipped end to end without a single Figma file. The scan flow, the results page, the edge functions running on Gemini, all of it iterated in production. The first paying user signed up in the same week the design system stabilized.

    04

    Are you iterating on real user feedback or your own assumptions?

    AI makes polish nearly free, which makes self-deception nearly inevitable. Without users in the loop, every iteration just compounds your own taste.

    User feedback

    Keep going. Refine on what is actually true, even when it contradicts the roadmap you already wrote.

    Assumptions

    Ship what you have, even if it embarrasses you, and put it in front of ten real users this week.

    Lived experience, Credit Karma Tax

    On Credit Karma Tax, my analyst and I assumed the W-2 step was the biggest leak. Session playback proved otherwise, fifty percent of users were dropping at the dashboard, stuck in a click loop on a locked card. The fix that recovered the funnel was the one I would not have prioritized without the data.

    05

    Have you smoke-tested the critical paths?

    AI-generated code feels finished long before it is. Confidence is part of the output, not a measure of quality. Trust the funnel, not the vibe.

    Yes

    Ship it. Deploy. The product only exists when it is being used.

    No

    Walk every core flow yourself, on a real device, on bad wifi, signed out and signed in. Then ship.

    Lived experience, Credit Karma Tax and UX Risk Scanner

    On Credit Karma Tax, the average mobile fold sat at 624 pixels, so the primary CTA was invisible to most users on first paint. The dashboard reviewed beautifully on a 27 inch monitor and quietly failed in the funnel. The lesson stuck. On UX Risk Scanner I now walk the full scan on a throttled phone before any release.


    Operating Principles

    What I want my teams to internalize

    The framework is the surface. Underneath are five principles I hold teams to. They predate AI. AI just made them load-bearing.

    01

    Validate before you build

    The cheapest design is the one you do not have to design twice. Demand first, fidelity second.

    02

    Brief in one sentence

    If the product cannot be described in a sentence, it cannot be defended in a meeting or shipped in a quarter.

    03

    Build in production

    Static specs are translation layers. Translation layers leak intent. Work where users will feel the work.

    04

    Listen louder than you ship

    AI makes iteration almost free. The discipline is choosing what to iterate on, and on whose evidence.

    05

    Trust the funnel, not the vibe

    AI-generated work performs confidence. Confidence is not a quality signal. Real flows on real devices are.


    Closing

    The next decade of design leadership

    The designers and design leaders who will define the next decade are not the ones writing the most detailed specs. They are the ones closest to the problem, fastest on real evidence, and most willing to ship something honest before it is polished.

    AI-native design is not about speed. It is about clarity, applied earlier, on better evidence, with fewer translation layers in between.

    Newsletter

    New writing, frameworks, and lessons from AI-native product work, straight to your inbox.

    By subscribing you agree to receive occasional writing from UX by Ash. Unsubscribe anytime. See the Privacy Policy.

    Advisory

    Startup advisory & fractional leadership.

    Partnering with founders and leadership teams to build AI-first products, scale design organizations, and drive measurable business impact.

    Request Availability Book A Discovery Call
    AI-Native Product Strategy

    From idea to roadmap, I help define the right problems and scalable solutions.

    Design Leadership On-Demand

    Embed executive design leadership without the full-time overhead.

    Team & Process Building

    Hire, mentor, and align design teams to build high-performing culture and processes.

    Growth & Impact

    Align design with business goals to drive adoption, retention, and revenue growth.

    Ashwarya Subhluxmi

    Design leader for Credit Karma's $3B marketplace. Five business lines, a team of staff-level designers, 140M+ members.

    San Francisco, available for advisory
    © 2026 Ashwarya Subhluxmi

    Legal notice. All content on uxbyash.com, including written work, case studies, frameworks, visual design, code, illustrations, and page layouts, is the original work of Ashwarya Subhluxmi and is protected by copyright and other intellectual property laws. No part of this site may be copied, reproduced, republished, mirrored, scraped, fine-tuned into a model, or used to train AI systems without prior written permission. Short quotations for editorial, educational, or commentary purposes are welcome with clear attribution and a link back to the source page. Unauthorized use, including verbatim reuse of copy or design, will be treated as infringement and pursued under applicable law.