In partnership with
Three dimensions used to be enough
For two decades, design schools taught one evaluation framework for any new product idea.
Is it desirable: do people want it?
Is it functional: Can we build something that works?
Is it viable: does the business case hold?
Pass all three, build. Fail any one, don’t.
That triad was the gold standard for evaluating software, hardware, services — anything we shipped. Then AI moved inside our products, and the framework evolved. The three dimensions still sound right. Desirability still matters. Functionality still matters. Viability still matters.
The problem is that functionality (once a binary) has fractured into eight separate questions, each of which can quietly kill an AI feature even when the other seven are green. Most teams haven’t updated their evaluation framework yet.
They are still using a 2005 triad to evaluate 2026 software. That is why so many AI features ship, demo well, and then quietly disappear.
In today's newsletter:
Why the classic triad (desirability, functionality, viability) is no longer enough when AI lives inside the product
The eight questions you now need to ask instead of just “does it work?”
How Jared Spool uses these dimensions to stress-test AI features before a team commits engineering hours
How the Triad Evolved
Most teams pick AI use cases the same way they picked features ten years ago.
Desirability used to mean “do users want this?” For AI, it now also depends on whether users trust it — and trust depends on observability, predictability, and how much control the user feels they have.
A feature can be wanted in theory and trust-bankrupt in practice.Functionality used to mean “does it work?” For AI, that single question fractures into a long list: is it accurate, predictable, complete, observable, learning over time, fast enough, aware of current context?
Each one can be green on demo day and red on day thirty.Viability used to mean “does the unit economics work?” For AI, viability depends on compute cost per task, on latency-driven adoption, and on whether the model improves over time or has to be re-trained at scale every quarter.
Static cost assumptions break when the underlying compute is variable.
The triad is still where you start. It is no longer where you finish. The real diligence — the part that decides whether an AI feature actually survives its second month — happens one level deeper, inside the functionality lens.
The Eight Dimensions
When you press on “does it work?” for an AI feature, eight separate questions fall out. Run your use case against each of them — score it red, yellow, or green — before you commit engineering hours.
1. Accuracy: Does it pull the right content, or does it miss essential details? One wrong data point in Nadia's leadership report means lost credibility.
2. Predictability: Run it three times. Do you get the same result three times? If not, trust breaks.
3. Ease of Control: Prompting is an art form. If your user needs to become a prompt engineer to get value, you have a design problem.
4. Detail Completeness: AI nails the first 80%. The last 20% is where it falls apart — and that is often the part that matters most.
5. Observability: Can you see what the AI did and why? If Nadia cannot trace where the data came from, she will either blindly trust it or stop using it.
6. Computational Learning: Does it get better over time? Right now, most AI tools start from zero every session. Nadia corrects the same mistake on Monday and again on Wednesday.
7. Computational Speed: Fast enough to fit into the workflow, or slow enough that doing it manually takes the same time?
8. Contextual Awareness: Does it know the product roadmap shifted last month? AI often works with a frozen picture of the world. If it does not know things changed, it is creating work, not removing it.
The problem is rarely a lack of talent or tools. It's a lack of alternatives.
How Jared Spool Uses This
Jared Spool, a longtime UX strategist and founder of User Interface Engineering (UIE), has spent three decades arguing a single point: the gap between “the thing works” and “the thing works for the person doing the job” is where most products die. AI has widened that gap, and his pattern when evaluating an AI feature reflects it.
He pushes teams to demonstrate — not assert — accuracy on their own data. He’ll run the same input three times in front of stakeholders to expose predictability gaps no one had stress-tested.
He pushes hardest on observability, because he has watched too many AI features get quietly abandoned the moment users couldn’t explain results to their stakeholders.
Think of it like Lego.
The triad itself was an evolution.
Designers built it because earlier evaluation frameworks. Usability, manufacturability, marketability were each too narrow on their own.
Combining them into desirability/functionality/viability gave the field one shared language for product decisions across crafts.
Think of it like Lego.
The original set was basic bricks 2x4s and 2x2s that gave every build its shape.
As people started building more complex things — vehicles, robotics, architecture — Lego didn’t throw out the old bricks. It added specialized pieces: Technic gears, Mindstorms sensors, hinges, transparent panels.
The basic bricks still hold every structure together. The new pieces let the structure do things the old set never could.
The triad is the basic bricks. The eight dimensions are the specialized pieces the AI era forced the set to add.
AI is forcing the next evolution.
The teams that win the next three years will be the ones who updated their evaluation framework before the rest of the industry did — and used it to say no, this isn’t ready for AI yet on the whiteboard, (instead of proving it the expensive way after launch.)
How to Run the Build
Lay the baseplate first. Ask the triad questions before you reach for the specialized pieces. Do users want this? Is it technically possible at all? Does the math work?
Snap on each brick, one at a time. For each of the eight dimensions, write a single specific test you’d want to pass at launch. Concrete tests, not abstractions. “Run the same input three times, get the same answer twice” is a test. “Be more predictable” is not.
Watch for cracked load-bearing pieces. Some bricks hold more weight depending on the use case. For an executive report, observability and accuracy are load-bearing.
For a brainstorming tool, ease of control and speed matter more. Identify which dimensions are load-bearing for your use case before you score themRun it as a thirty-minute exercise. Get the team in a room. One whiteboard. Eight columns. Score each one red, yellow, green. Argue.
Key Takeaway
The classic design triad — desirability, functionality, viability — is necessary but no longer sufficient for AI features inside software. Each of the three lenses has expanded.
The biggest expansion is inside functionality, which has fractured into eight dimensions: accuracy, predictability, ease of control, detail completeness, observability, computational learning, computational speed, and contextual awareness.
Use the triad to brainstorm whether a use case belongs on the roadmap. Use the eight dimensions to decide whether it is ready to be built.
They are not parallel — the eight live underneath the three.Most AI features that quietly disappear from products are dying because the team used a 2005 framework to evaluate a 2026 capability.
AI Tools I'm Using This Week
💬 Claude — Now using chat for casual brainstorming and co-work for deep thinking🤖 Manus — Build presentations, videos, and prototypes from a prompt
Write docs 4x faster. Without hating every second.
Nobody became a developer to write documentation. But the docs still need to get written — PRDs, README updates, architecture decisions, onboarding guides.
Wispr Flow lets you talk through it instead. Speak naturally about what the code does, how it works, and why you built it that way. Flow formats everything into clean, professional text you can paste into Notion, Confluence, or GitHub.
Used by engineering teams at OpenAI, Vercel, and Clay. 89% of messages sent with zero edits. Works system-wide on Mac, Windows, and iPhone.