Playbook/Stage 02

Ship

Get it live

Building an AI Product

If AI is what your user interacts with, this is how to build it right. Patterns, UX, trust.

This section is for builders whose product has AI at its core -- where the user interacts with AI directly. If AI is your construction tool and the user never sees it, skip ahead to Ship It.

Five Patterns That Work

Most successful AI products fit one of these. Pick the one that matches what you already know about.

PatternHow It WorksReal Example
The Expert AdvisorUser gives it information → AI gives back expert analysis they would normally pay a consultant forA financial analysis tool, a legal document reviewer, a marketing audit tool
The Content CreatorUser gives it context → AI generates personalized content at scaleEmail writer for realtors, social media posts for restaurants, proposal generator for freelancers
The Smart DirectoryA free searchable database that attracts visitors → premium features behind a paywallA directory of youth sports programs, local contractors, or niche tools -- enriched by AI
The Process GuideA complex, multi-step process turned into an AI-guided walkthroughTax prep assistant, onboarding system, compliance checker for a specific industry
The Personalized AssessmentYour expertise turned into an interactive, personalized experienceA quiz that gives tailored recommendations, a diagnostic tool, a coaching platform
The secret: AI is the delivery mechanism. Your knowledge is the product. Anyone can ask ChatGPT a question. What makes your product valuable is the specific way you have structured the questions, the context you feed the AI, and the format of the output. A financial advisor who builds an AI analysis tool brings 20 years of judgment. A parent who builds a youth sports assessment brings lived experience no AI has.
Define the job, not the feature. Write one sentence: "When [situation], the user needs to [motivation], so they can [outcome]." For example: "When a freelancer gets a new lead, they need to write a personalized proposal in under 10 minutes, so they can respond before the client moves on." That sentence tells you exactly what the AI should do. If you cannot write it, you do not know what you are building yet.

The Weekend Sprint

You have validated the idea. Now build it. A working AI product in a weekend if you keep the scope tight. The constraint is the point -- it forces you to focus on the one thing that matters.

WhenWhat to Do
Friday eveningPick your pattern. Write one paragraph: who is this for, and what is the one thing the AI does for them?
Saturday morningBuild the AI part: what goes in, what comes out, what format. This is the core -- get it working before anything else.
Saturday afternoonBuild the website around it. A simple form, a results page. Add login if people need to save results.
Sunday morningPolish: what happens when it is loading? When it fails? Does it work on a phone?
Sunday afternoonPut it online. Send it to 5 real people. Watch what happens.

Making AI Feel Good to Use

Builder's checkWatch one person use your product for five minutes without helping them. Do not explain, do not jump in, just watch. It will be uncomfortable and it will teach you more than any UX article, including this one. Your users do not care about your model, your token count, or how clever your pipeline is. They care whether their problem went away. And here is the irony of building an AI product: the best version of it right now might be you doing the work by hand. Send the first ten users their results personally. Write the recommendations yourself, using AI to help but not to replace your judgment. You will learn more in those ten manual deliveries than in ten weeks of prompt engineering. Automate after you know exactly what good looks like.

When AI is thinking (loading)

AI takes 5-15 seconds to respond. That is an eternity on the internet. Users will leave if they see a blank screen.

  • Show the response as it is being written -- words appearing one by one feels fast, even if it takes the same time.
  • Show progress messages -- "Analyzing your input..." → "Generating recommendations..." → "Almost done..." feels faster than a spinner. Name what the AI is doing.
  • Tell them how long it usually takes -- "This usually takes about 10 seconds" removes anxiety
  • Show a cursor -- a blinking cursor at the end of the stream tells users "I am still working." Remove it when done.

When AI gets it wrong

It will. AI makes mistakes. Your product needs to handle that gracefully:

  • Add a "Try again" button -- a second attempt often gives a better result
  • Let people edit their input and retry -- better input = better output
  • Add thumbs up/thumbs down -- the simplest way to find out when AI is underperforming. Only useful if you actually read the downvoted outputs and improve from them.
  • Offer alternatives on failure -- "I could not process that. Try rephrasing, search the help docs, or contact support." Never show a blank screen or a cryptic error.

Building trust

The biggest barrier to adoption is not accuracy. It is trust. Users who see one wrong answer may never trust your product again. Trust builds in stages:

  1. Transparency -- the user knows it is AI and can see what it is doing. No black boxes.
  2. Control -- the user can edit, override, or reject the AI's output. They are in charge.
  3. Understanding -- the user starts to see patterns in how the AI thinks. They can predict when it will be right and when to double-check.
  4. Verification -- the user has tested the AI enough to trust it on routine tasks. They still spot-check, but less often.
  5. Delegation -- the user lets the AI handle things without reviewing every output. This takes weeks or months, not minutes.

Most builders design for stage five on day one. Your users are at stage one. Design for that:

  • Suggestions, not automation -- AI suggests, human decides. Never commit an action without the user's confirmation.
  • Show the reasoning -- "Based on what you told me about X and Y, I recommend..." is much better than just giving an answer.
  • Show confidence -- "3 matches found, top match 92% confidence" tells the user something real.
  • Never pretend the AI is human -- be clear it is an AI tool that is using your expertise

Trust regresses. One bad output can drop a user from stage four back to stage two. The "Try again" button and the thumbs down are not polish -- they are trust infrastructure. If you are coming from traditional software, your instincts about testing, shipping, and user trust will actively mislead you here. Building AI Is Not Building Software explains why.

Go deeper: Building AI Is Not Building Software — why your software instincts will mislead you when building AI products.