Field notes from shipping.

Rolf Nelson · Design by Proxy

Most AI adoption advice gets written by people who haven't shipped any. I have, and most of what I learned doesn't fit on a slide.

I've spent the last few years putting AI into real products: suggestion patterns, explainable reports on demand, content review for a backfill of over 20,000 images, a handful of small apps I built to keep my hands dirty. This is what I'd tell a leadership team trying to do the same without torching user trust or their own people on the way.

The short version: move fast, but don't break the user.

A few concepts I'll put my name on:

Adopt where it earns its place. AI belongs where it makes the product measurably better for the person using it. Trend pressure is not a use case. If you can't name the user problem it solves, you're shipping novelty.

Tell your teams what this actually is. The fastest way to lose your designers, researchers, and engineers is to let them believe a model is coming for their jobs. It isn't. Say so, early and plainly, and mean it. Morale is a feature.

Build trust into every layer. Explainability, grounding, and a human in the loop do real work. They keep a confident-sounding wrong answer from reaching a user with your logo on it.

Let teams experiment, give them rails. People should feel free to prototype. They should also know where the edges are. Freedom without a frame is just everyone learning the same lesson separately, and expecting them all to teach from each other or learn the same ways isn’t a likely scenario.

Lead with clarity. Hype without direction breeds friction. Give people examples, guardrails, and a reason to care, and adoption feels like momentum instead of noise.

None of this slows you down. It's what lets you go fast without leaving a mess for someone else to clean up.

But most AI adoption breaks earlier than any of this, before a single feature exists. It breaks in how the decision gets handed down.

The decree is not the work

Most AI adoption fails before anyone writes a line of code. It fails the moment leadership says "go" and hands the teams nothing else.

A mandate sets direction. It does not create alignment. When "adopt AI" arrives as a decree with no workflow attached, product, design, and engineering each fill the gap with their own read. The PM scopes one thing, design imagines another, engineering builds a third. Everyone is moving, which looks like progress, and the result solves for novelty because nobody agreed on the problem first.

I have watched this happen more than once, and the technology is rarely the bottleneck. The missing piece is a shared way for three disciplines to decide together before anyone builds. That is a workflow, and it comes down to three things.

A shared problem statement. Turn the mandate into one written sentence the whole team signs: the specific user problem the AI solves, and what good looks like. If the three disciplines can't agree on that sentence, you aren't ready to build. Shipping anyway just moves the disagreement to a more expensive moment.

Clear ownership. When everyone owns the AI decision, no one does. Give each call a single owner. The PM owns the problem and the bar, design owns the experience and the controls, engineering owns model behavior and the guardrails. Write down who decides before the disagreement, not during it.

Hand-offs that survive. An AI feature changes meaning every time it crosses a discipline. Intent drifts from spec to flow to implementation until the thing that ships isn't what anyone agreed to. Carry the user problem through every hand-off, not just the ticket. Prototype the behavior first so engineering builds against something real, and review the result against the original problem.

Get this layer right and the product decisions in the rest of this paper get easier, because the team is making them together. Skip it and you get speed in three directions, which is slower than going nowhere.

What we're actually talking about

A quick glossary, because half of AI conversations stall on people using the same word to mean four different things.

• Artificial intelligence (AI): the broad field of building machines that do things we'd call intelligent.

• Machine learning (ML): algorithms that learn patterns from data instead of following rules someone hand-wrote.

• Large language models (LLMs): ML models trained on enormous amounts of text to predict and generate language. The thing behind most of what people mean when they say "AI" today.

• AI agents: systems that perceive, decide, and act toward a goal with some autonomy. The model stops answering questions and starts doing things.

Prompting, briefly

The quality of what you get out of an LLM tracks the quality of what you put in. Vague prompts produce vague output. Give the model context, tell it the outcome you want, and specify the format. "Summarize this" gets you mush. "Summarize this report in five bullets focused on user pain points" gets you something you can use. This is most of the skill, and most people skip it.

Designing with AI: the principles I hold

Make it explainable

When the AI shapes an outcome, the user deserves to know why. A recommendation with no reasoning attached is a black box, and black boxes don't earn trust, they spend it. Show the work: what the model saw, why it landed where it did. This sets honest expectations and lets people calibrate how much to rely on it.[1]

Augment human judgment

AI earns its place by making people better at their work: faster, sharper, freed from the tedious parts. The best implementations take the grunt work off someone's plate so they can spend their attention where a human is required, on the creative call, the strategic read, the judgment. Aim the tool there.

Finding the right places to use it

Start with the user problem

The bar for adding AI is the same as the bar for adding anything: does it solve a real problem for the person using the product? Start from user needs and success criteria, then ask whether AI gets you there better than the alternative. Personalize content, automate a workflow nobody enjoys, surface an insight you couldn't reach by hand. Those are reasons. "Competitors have it" is not.

Control versus autonomy

There's a real tension here, and pretending otherwise produces bad features. People want their tools to be predictable and to feel in control. They also want the boring, repetitive work handled for them. Those pull in opposite directions. Decide case by case: does this feature wait for an explicit command, or does it act on its own? An assistant that auto-sorts your inbox is great until it buries something you needed, at which point the same behavior is a betrayal. Same feature, different autonomy, opposite reaction.

The hybrid usually wins

Where you can, give people control and let the AI offer help on top. Direct manipulation for the things users want to own, suggestions and initiative for the low-risk grunt work. Roll it out gradually, show clearly what the AI is doing, and people learn when to trust it and when to step in. You end up with users who feel in charge while the machine handles the parts they'd rather not.

Ground it in real data

A generative feature is only as reliable as what you feed it. Supply the model with current, domain-specific context (your knowledge base, the user's actual situation) instead of letting it freewheel. Retrieval-augmented generation (RAG) is the common pattern: the system pulls the relevant facts and hands them to the model as part of the input. Grounding is what keeps a confident model from inventing things.[2] Skip it and you ship hallucinations.

Where it actually saves time

This is where AI is already pulling its weight, in my own work and across teams I've watched.

Documentation and specs

Drafting from a blank page is the slow part. AI closes that gap: rough notes into a structured first draft, repetitive sections filled, terminology held consistent across a pile of documents. You start from something to react to instead of an empty file. The reported gains are real (one clinical-research team cut documentation time by roughly 40%),[3] and they track with what I've seen. The model gets you to a draft worth editing, fast.

Research synthesis

AI is genuinely good at the part of research everyone dreads. Feed it interview transcripts and survey responses and it'll transcribe, cluster, and surface the recurring themes and pain points. Teams report synthesis running an order of magnitude faster this way.[4] The catch, and it's the important part: this is a collaborator, and the verdict stays with you. The model finds the patterns. You decide what they mean. Done right, it hands the researcher their time back to do the interpretation, which was always the actual job.

Wireflows for AI-driven UX

Wireflows marry wireframes to flowcharts, and they're the right tool for AI features because AI features branch. [5] The screen changes based on what the model returns: different states, different content, different next steps. A wireflow lets you lay all of that out, screen by screen, and catch the broken edge case before anyone builds it. Mark the interactive elements and label what each user action or AI event triggers, so anyone can follow the logic without guessing.

Wizard of Oz: fake it before you build it

Before you spend a quarter building a model, put a human behind the curtain. Users interact with what they think is an intelligent system while a person generates the responses out of sight. You learn fast whether people want the help at all, how much autonomy they'll accept, and what tone earns their trust. It's the right move when the range of possible responses is too messy to pre-script, which covers most AI features worth building. You get real reactions at a fraction of the cost, and you find out early if the idea was wrong.

Prototyping with AI tools

AI builds the prototype now, it doesn't just star in it. Hand a model a design or a rough sketch and get back a working, clickable thing in an afternoon. I lean on this constantly, including a stretch where I shipped one AI-assisted app a day to see how far the tooling could go. The value isn't a perfect first draft. It's that I can spin up three versions of a flow, click through all of them, and kill two before the idea calcifies.

• One caveat, earned: these prototypes look finished long before they are. The code reads as done and isn't. Treat the output as a starting point and do the real UX pass yourself. The model gets you to the interesting questions faster. It won't answer them.

The ways this goes wrong

Every failure I've seen with AI on a product team traces back to one of these.

Trusting it blindly

The danger isn't that AI is wrong sometimes. It's that it's wrong confidently, in fluent complete sentences, and people stop checking.[6] Automation complacency is real: hand someone a tool that's right most of the time and they quit verifying the rest. Treat AI output the way you'd treat a sharp colleague's first draft, useful and worth a second look. When a model surfaces a surprising insight, go back to the source data before you act on it. Skepticism keeps the tool honest.

Presenting output with no context

AI output needs its receipts attached, especially in front of stakeholders or users. "The AI recommends changing feature A" is useless and slightly alarming. "An analysis of user feedback points to feature A, since 60% of comments mentioned confusion there" is something people can evaluate. Show where it came from, how confident it is, what it's based on. Without that, people either over-trust the output or dismiss it, and both are expensive.

Cutting the human to move faster

Human oversight is the thing you protect, even when it's the slow part. Keep a person at the decision points that matter. If a model flags user content as inappropriate, a human reviews the borderline cases before anything gets deleted. The reviewer is how bias and error get caught before they reach a user. The whole industry keeps relearning this: human involvement matters at every stage, design through monitoring.[7] Pull it out to save time and you'll spend the savings, with interest, on the cleanup. Autopilot is not the goal. A person with better tools is.

Move fast, don't break the user. Everything above is just how you keep both halves of that sentence true at once.

Sources

1. Google People + AI Research (PAIR), People + AI Guidebook, on explaining AI decisions to build user trust. pair.withgoogle.com

2. Microsoft Tech Community, on grounding generative AI with retrieval-augmented generation. techcommunity.microsoft.com

3. Saama, "Behind the AI-Powered Document Generator," reporting up to a 40% reduction in clinical documentation time. saama.com

4. Strella, AI-moderated interviews and real-time synthesis, reporting roughly 10x faster time to insight. strella.io

5. Nielsen Norman Group, "Wireflows: A UX Deliverable for Workflows and Apps." nngroup.com/articles/wireflows

6. Nielsen Norman Group, on AI tools producing confident but incorrect output. nngroup.com

7. Thomson Reuters, on maintaining human oversight across the AI lifecycle. thomsonreuters.com

Rolf Nelson is a product design leader with two decades across SaaS, fintech, and consumer products. He runs Design by Proxy, a fractional design practice, and has spent the last few years putting AI into things people actually use.