AI Is Startware, Says Nielsen. It Finishes the Day the Failure Inventory Exists.

Jakob Nielsen's roundup this week carries a survey of 443 designers in 43 countries, and the curve it draws is the one every design leader already feels in their own team. AI use peaks in the first phase of the Double Diamond and bottoms out in the last. 72% use it in Discover. 41% use it in Deliver.

Nielsen's explanation for the curve is the best sentence in the piece. His prescription is where I get off, and the difference is a thing you can write in ninety minutes.

The short version

Sepideh Tajarmakan and Khashayar Hojjati Emami surveyed 443 designers, as Jakob Nielsen reports on October 19, 2026, and found AI use falling across the Double Diamond from 72% in Discover to 41% in Deliver, with nobody starting in Deliver except two. Nielsen's reading: early errors are free and late errors ship, so AI is startware for now; ask for 20 alternatives instead of 3, reinvest the hours in convergence, and wait for vendors to build finishware designers trust, with Deliver catching up by 2028. The rewrite: late errors ship because nobody wrote down what a late error is. The failure inventory from my design handbook lists the ten worst plausible outputs of a feature with one of five responses each, prevent, catch, disclose, recover, or accept. The prevents become constraints and the catches become eval cases, and a generated edit that passes the catches ships with the same confidence as a hand-made one. That opens the Deliver phase this quarter, it is design's to write, and the reason to write it now is that Figma's design chief already reported 17x more code and 1.3x more releases. Finishing is where the hours are.

What the survey says

The shape first, in Nielsen's numbers. 80% of respondents used AI in at least one phase, but only 35% of those used it in all four. Adoption accumulated in workflow order: Discover alone was the most common single phase, Discover plus Define the most common pair. Nobody started with Deliver. Well, two did.

Inside the phases the tools change. In Develop, standalone generators, mostly for images, took 58% of all tool picks. Only 9 of the 73 AI tools named were general-purpose assistants, and they won most picks in three of the four phases; ChatGPT alone took 52% of picks in Define. Near the finish line the pattern flips: AI built into familiar software, Photoshop's Generative Fill is his example, peaked in Deliver at 38% of picks. Nielsen's reading of that is right: at the end, designers want AI inside a tool they trust, working on a selection they control.

The holdouts are the most useful group. Of ten barriers, only one separated the 84 non-users from the adopters. 40% of holdouts found AI insufficiently useful, against 16% of adopters. Ease of use did not distinguish them at all, 12% against 10%. The holdouts are not confused by the tool. They are unconvinced it does the job.

Nielsen flags the sample honestly. Respondents came through social media and the authors' networks, 47% from Iran, 60% in their twenties, and enthusiasts answer AI surveys, so the 80% is probably high. He trusts the shape, not every point, and so do I.

Where Nielsen lands

His explanation: early errors are free; late errors ship. A weird AI idea in Discover costs nothing, you discard it or it sparks a better one. A wrong AI edit in Deliver reaches customers. That is correct and it is the whole problem in seven words.

Then the prescription. AI is startware, for now. Designers trust it to open a project far more than to close one. So exploit free ideation, ask for 20 alternatives rather than 3, and reinvest the saved hours in convergence, winnowing, testing with real users, finishing. For vendors, the untapped market is finishware that designers trust to deliver. His prediction: repeat the survey in 2028 and Deliver will have caught up with Discover.

I argued the 20-alternatives half in Kill the Batch of Forty: Count the Unopened, Not Review Time, so I will leave it. The half to argue with is the wait. It puts the Deliver phase on a vendor's roadmap, and it treats "designers trust it" as something a tool earns on its own.

The rewrite: late errors ship because nobody wrote them down

Late errors do not ship because the generator is immature. They ship because the team never wrote down what a late error is, and so there is nothing to check a generated edit against except a senior designer's eye.

The Failure Inventory(coming Nov 25) is the handbook chapter for this, and it predates the survey. Ninety minutes, three people, ten rows. Each row names the worst plausible output in the product's own words: not "the AI could be wrong," but the actual sentence the feature would show a customer on a bad Tuesday. Then who gets hurt, what it costs, whether it recovers, and one of five responses: prevent, catch, disclose, recover, or accept. Plausible means it could happen with the system working as designed, which rules out meteors on one side and bugs on the other.

Then two things happen to the list. The prevents become constraints, the lines a surface must never cross, which belong in the spec. The catches become eval cases, which is the sentence that opens the Deliver phase. Once the ten worst outputs exist as test cases, a generated edit is not an act of trust. It is a candidate that passes the catches or does not, and the check is the same check a hand-made edit would get.

Design wrote the list, in ninety minutes, before anyone designed the good version. That is finishware. It is a document, not a product, and nobody can ship it for you.

The senior curator still holds the veto, and Nielsen is right to keep it: once AI has generated a mountain of options, someone senior has to judge them. The inventory is what the veto reads. A veto with ten written rows behind it is a decision. A veto without them is a mood, and a mood is what 40% of the holdouts were reporting.

Why this quarter, not 2028

17x the Code, 1.3x the Releases. Design's Job Is the Spec Now has the number: Figma's design chief reported 17x more code and 1.3x more releases. Generation stopped being the bottleneck. Deciding what is good enough to ship became it, and that decision was always design's. The survey draws the same picture from the designer's chair: the first diamond got cheap and the second did not.

Waiting for finishware spends the two years in which the finish is the whole job. A design team that writes ten rows per feature this quarter is running AI in Deliver while the survey is being repeated. One that waits for a trusted tool will be measured in 2028 as the holdouts were measured in 2026: not confused, just unconvinced, with the releases still at 1.3x.

Under AI Product Management, the argument is that AI collapsed the parts of the job people were trained for and left the parts nobody was. Finishing is the part nobody was trained for, because it used to be absorbed into the drawing. Now it needs a list. Figma's Trial Closed 70% of the PM Gap. The Gate Is the Rubric, Not the Designer. made the same move on the PM side last week: a person as the gate caps throughput, and the written artifact is what scales.

This week: pick the feature your team is finishing by hand. Three people, ninety minutes, ten rows. Hand the catches to whatever generated the first draft and run it. Count how much of the Deliver phase was waiting on a tool, and how much was waiting on the list.

Related answer: Why do designers use AI less in the Deliver phase?

Sources: Jakob Nielsen, "UX Roundup: AI by Design Phase | Product Managers vs. Designers | Individualized Tutoring | ...," UX Tigers, October 19, 2026, the AI by Design Phase section read in full. Sepideh Tajarmakan and Khashayar Hojjati Emami's survey of 443 designers (arXiv preprint) is cited only as Nielsen reports it; the preprint was not opened. Falk Gottlob, "The Failure Inventory(coming Nov 25)," "17x the Code, 1.3x the Releases. Design's Job Is the Spec Now," and "Figma's Trial Closed 70% of the PM Gap. The Gate Is the Rubric, Not the Designer."