What Can You Build With AI? 10 Assets for Your Expertise

Most of us who sell what we know own far more than we can sell, because the valuable part lives in our heads and only comes out during delivery — the questions you ask on a first call, the order you walk somebody through a decision, the thing you spot in minute four that she has been circling for six months — and for years the only two containers we had for any of it were the calendar and the content. You sell your hours, or you publish something people read and admire and then do absolutely nothing with. (I say that with love, as a woman who has written a LOT of admired content.)

The answer, and it took me an embarrassingly long time to see it, is that one piece of your method can now become something another person uses without you in the room — an assessment, a guided planner, an intake process, a calculator, a proposal draft, a small app that does one useful job — and you can make the first version yourself, this month, inside the AI you already pay for, with no product launch and no year of development.

There is a trap on the other side of that, which is why a good chunk of this piece is about what not to build — an expert with a working AI tool and a free Saturday can lose the whole weekend making something clever that nobody asked for (read: expensive, in the only currency none of us can buy more of) — so we are going to be picky about which one you build and very small about the first version. If you have not decided yet which part of your business deserves the attention first, go read how to choose your first AI use case and then come back, because the ten builds below go a lot better when you already know which problem you are solving.

What makes an AI-assisted business asset worth building?

An asset is worth building when it does a repeatable job for a specific person without you rebuilding the whole experience from scratch every single time. That is the whole test, and it rules out most of the fun ideas, which is the point of having a test.

What makes the asset worth anything is your questions, your criteria, your examples, the distinctions you draw that a smart outsider would miss, and the judgment you have built over however many years of doing this work — the software around all of that is close to free now and getting cheaper. When I watch a build go wrong it is almost never a technical failure; it is that the expert handed over a vague idea and got back a beautiful, confident, entirely generic version of something that already exists in twelve other places.

That is also the difference between using AI and building with it: when you use AI you sit down, work with it, and walk away with a result, and when you build with it you spend the time once so the result shows up again next week without you sitting down at all. I wrote more about how those two modes feel different in practice if you want the longer version.

How do you pick which one to build first?

Pick the one where you already have the raw material and a real person who would use it next week. Two conditions, and if a build idea fails either one, it goes on the list for later instead of onto your Saturday.

Having the raw material means you have done this by hand enough times to know what a good result looks like — which questions matter, what the answers mean, where people go wrong — because AI can turn your criteria into a working tool very fast but cannot supply the criteria, and a scoring system built out of guesses produces confident nonsense at scale. The second condition means somebody has already asked you for this, so you are not testing on an imaginary audience.

If both of those feel shaky, the first job is writing your method down, and the build comes after that. That sounds like a consolation prize and it is not, because when a client gets her process out of her head and onto a page, she usually discovers the app was already in there, fully formed, waiting for somebody to type it up.

What can you build with AI for an expertise business?

Ten things, and every one of them starts as something you already do by hand: a website, a small app, a survey or assessment, a client intake process, a calculator or estimator, a guided planner, a searchable library of your own work, an interactive FAQ, a proposal generator, and a personalized roadmap. Under each one is what it is, what it looks like in practice, and the first thing to do — and none of those first steps involve code.

1. A website

AI can plan your pages, draft the copy from positioning you have approved, propose a visual direction, and produce a working first version you can put online this week. I built the site for my VIOS app with Claude Code, and my own site is vibe-coded, which is a sentence I would not have believed about myself two years ago.

The caveat I owe you: your first AI-built website will probably look like everybody else’s, because the models have opinions about what a professional site looks like and those opinions are extremely samey, so you will want a designer and a real brand eventually. But if the alternative is the site you have been meaning to rebuild for years, put something honest up now and refine it later — and if you want it sounding like you while you do it, I wrote about building tech that feels like you rather than like a template.

First step: decide the job of the site before a single page exists. Is it explaining your work, attracting speaking, qualifying prospects, or getting people into a conversation? Then show the draft to one person in your audience, ask her what she thinks you do and what she would click next, and fix the message before you fuss with animations.

2. A small app

A small app is one user, one job, one useful result, and the building tools will take a plain-English description of that and hand back a working prototype to poke at. My rule of thumb is that when I catch myself asking Claude to do the same thing over and over it becomes a skill, and when I use several skills together over and over, I start asking whether the whole thing wants to be an app.

That is how my LinkedIn Profile Writer happened. We used to charge $3,500 for a profile optimization, which was fair for the hours it took, and the tool now does the heavy first pass in minutes. (Oops. Now you know what we used to charge.) I build with ChatGPT, Claude, and Lovable.dev depending on the job, and you do not have to write every line yourself — though anything public still needs real help with accounts, payments, security, and the unglamorous work of keeping it running.

First step: write the app’s job in one ordinary sentence, and if you cannot, the first version is too big. Then have the AI propose a manual version — the one where you do the steps yourself with a document and a phone call — and run that with one real person, because if the process is not useful when you guide it by hand, software will only make it faster and more expensive.

3. A survey or assessment

An assessment turns your criteria into a set of questions, a scoring range, and a plain-language explanation of what the result means for the person who took it. This used to cost real money, because you paid somebody to design the instrument and a platform to host it, and paid again every time you changed a question.

I built my Authority Gap Audit in a few hours and it lives on my site, and when I ran a coding class at my house where we played with Lovable and Claude Code, I was impressed by the surveys people walked out with by dinner. The freaking useful part is that an assessment gives a stranger a reason to tell you about her situation, and gives you a reason to follow up with something specific.

First step: write down what the assessment can identify and what it absolutely cannot diagnose, before you write a single question. Then test it on someone who resembles the person you built it for, and when she reads a question differently than you meant it, fix the question rather than the scoring. Half of my early questions were clear in my head and ambiguous on the screen.

4. A client intake process

Intake is everything that happens between “yes” and the first session, which in most businesses means forms, scheduling, a welcome email, document collection, and a pile of information somebody has to organize before you can do good work. AI can design the questions, catch incomplete answers, summarize what came in, and build your prep checklist.

This was the thing that pulled me in as a builder, and I put together an intake process early on that I already want to rebuild, because the tools have improved so much since — which is either a sign of progress or a sign of my personality, possibly both.

First step: map the sequence and mark, at every step, what the client submits, what the software may organize, what you review, and what triggers a personal note from you. Then walk through it as a client, on your phone, because a form that looks civilized on a laptop can become an endurance event on a six-inch screen. Nobody wants to write an essay with her thumbs.

5. A calculator or estimator

A calculator takes a handful of inputs and returns a number, along with the assumptions that produced it. I avoided these for years because Excel intimidates me and I assumed the category belonged to people who understood pivot tables, and it turns out I do not have to become a Microsoft expert to give somebody a useful number.

I worked with a client recently who wanted an estimator for his prospects, and where I would once have sent that to a developer, we made a working first version together in a short session — he may take it to a designer eventually, but he has something functional in front of customers now, which beats perfect-someday most of the time.

First step: make the AI restate the formulas and the assumptions in plain English, walk through three test cases including one weird edge case, and show its arithmetic step by step before it builds any interface at all, because a polished result can hide bad math beautifully and your eighth-grade math teacher was right about showing the work. And if you are anywhere near a regulated industry, call it an estimator rather than a calculator, put the assumptions on the screen, do not let it promise a financial or physical outcome, and bring on the disclaimers.

6. A guided planner

A guided planner asks one question at a time, uses your framework to choose the next question based on the last answer, and ends with a draft plan that shows which input produced each recommendation. It is the closest thing on this list to sitting across from you, which is why the questions have to be yours.

One of my favorites started as a physical thing, The Legacy Code Journal, and became an online version. I still believe in the handwriting — something happens with a pen that does not happen with a keyboard — and the digital version does what paper cannot, which is that you can photograph a handwritten page, have it transcribed, and ask for a synopsis of a whole month of entries. I had never once gone back and reread my own journals, and when the planner reviewed them it surfaced patterns and half-started ideas I had completely forgotten: it gave me my own thinking back.

First step: complete the planner yourself, as a user, start to finish. You will find the missing instructions in about four minutes, when it misunderstands your own answer and you know exactly why.

7. A searchable library of your own work

A searchable library lets you ask a question of everything you have ever made — articles, recordings, transcripts, decks — and get an answer grounded in your own material instead of a filename you half remember. I have not built mine yet and I freaking love this idea; it is sitting in my sandbox of things to make.

The version I want would let me ask “what have I taught about rebuilding visibility after leaving corporate” and get an answer drawn from my own teaching with the source named, and the useful ones say plainly when the answer is not in there, because a library that fills the gaps with general knowledge stops being your library and becomes the internet wearing your name.

First step: start with a small group of current documents, strip the duplicates and the outdated versions, keep private client material out entirely, and test the whole idea inside a Project in the AI you already have before you go shopping for a platform.

8. An interactive FAQ

An interactive FAQ lets a visitor ask her question in her own words and get an answer built only from information you have approved — your offers, your process, your policies, your boundaries, your next steps — instead of making her hunt for her question in a list, which she will not do.

I have a regular FAQ on my site that AI helped me write, and the interactive version is on my list. The part people underestimate is the handoff rule: anything about exceptions, contracts, guarantees, or whether she is a fit for you goes to a human being, and that human being is you.

First step: test it with badly phrased questions, questions missing half the details, and questions you do not want it answering at all, because polite refusal is part of the build and you want to watch it decline to invent a price before a stranger gets there first.

9. A proposal generator

A proposal generator takes your intake information and your approved template, picks the sections that fit, organizes the prospect’s stated goals in her own language, flags what is missing, and hands you a draft. I built mine right after the intake process and OMG, the time it has given me back on higher-end proposals is significant.

What it does not do is decide anything, because I set the recommendation, the scope, the price, the timeline, and every promise in the document, and I am the one who sends it. Revenue-related work deserves a visible human approval step, so your first version should produce a document that waits for you rather than anything that goes out on its own.

First step: compare the draft against the intake notes line by line before you admire how polished it looks, because polish is persuasive even when it is wrong, and a proposal that quotes a goal the prospect never mentioned is worse than no proposal at all.

10. A personalized roadmap generator

A roadmap generator organizes somebody’s goals, answers, constraints, and current stage, then selects from recommendations you have already defined and explains which of her inputs produced each step. I hate the word roadmap, because AI has decided to put it on absolutely everything, and unfortunately for my vocabulary preferences the roadmaps are useful anyway.

I built one into my VIOS app and I am consistently impressed by what comes back, for my clients and for me. Yes, I use my own tools on myself, which I recommend, if only because it is the fastest way to find out that step six makes no sense.

First step: run it against several past clients where you already know what you recommended and why, then study the differences, because the gaps between what it suggests and what you would have said are a precise map of the judgment still sitting only in your head.

What do you write down before you build any of them?

One page, five questions, answered before AI sees a single request for screens or code or copy. It takes about fifteen minutes and it is the difference between an impressive answer to the wrong problem and a small useful thing that works.

Hand that page over before you ask for anything else, and say plainly what the tool may not do — no invented credentials, no invented results, no testimonials, no client stories, no promises about outcomes. Most times Claude will push back on something in there, or ask a question you had not thought about, and that exchange is worth more than the build. You will also burn tokens getting to the brief that produces what you wanted, which is normal rather than a sign you are doing it wrong.

How do you review something AI built for you?

Review it like the expert you are, which means checking the logic against real situations and real clients where you already know the right answer rather than against tidy test data, so you can look at an output and say yes, that is what I would have told her, or no, that is nowhere close.

Then make sure every output traces back to information you approved, because an asset that pulls from general knowledge when your material runs thin will eventually say something in your voice that you do not believe. Feed it missing inputs, contradictory inputs, and outright strange ones, since the people who use it will hand it all three within the first week.

And then hand it to a person who is not you and do not help her — sit on your hands and watch where she hesitates, where she backs up, where she rereads an instruction, because every one of those is a place your build assumes knowledge that exists only inside your head. It is the least comfortable part of the process and it is where the improvements come from.

What about privacy and the information your asset collects?

If the asset touches private information, decide before launch where that data goes, who can see it, which AI provider processes it, how long it is kept, and what happens when somebody asks you to delete it. Someone will eventually ask, and after the fact is a terrible time to be working it out.

The easiest protection is collecting less, so cut every field that exists because it seemed interesting rather than because the work requires it — an intake form does not need her whole history to prepare one session. Anything public-facing may also need legal, accessibility, security, or professional review depending on what you do for a living, and if you are in a regulated field, go ask your compliance person before your Saturday rather than after.

What does one of these look like from start to finish?

Take the assessment, since it is where I would send most people first. The pressure point was the same conversation on every discovery call — an accomplished woman describes a visibility problem, I ask my handful of diagnostic questions, and by minute twelve we both see where the gap is — which is a fine use of a call, except that nobody got to that understanding without forty minutes of my time.

So the job was to let her find the gap herself, in about five minutes, with a result specific enough that she recognizes her own business in it. The material it needed was my diagnostic questions, the ranges I use to sort the answers, and my explanation of what each result means — all of it out of conversations I had already had hundreds of times, which is why the build took a few hours and most of those hours went to deciding the scoring rather than to anything involving software.

The first version had questions that made perfect sense to me and confused an outside reader, because I was using my own shorthand without noticing, and the second scored so generously that almost everyone came out in the middle, which helps nobody. The version that worked came from testing on people who resembled my audience and rewriting whatever they misread, which was not a triumphant montage, it was a tedious loop of watch, fix, watch again.

What I have now does a piece of my thinking for a stranger at two in the morning while I am asleep, and gives us both a much better starting point if she books a call — it did not replace the conversation, it moved the conversation forward twelve minutes before it starts.

When is building the wrong answer?

When you have not done the work by hand enough times to know what good looks like, when nobody has asked you for it, or when the build is standing in for a business decision you have been avoiding. I say this as somebody who builds constantly and still tells clients no on this regularly.

If you have run the process by hand exactly twice, do it ten more times first — deliver it as a service, charge for it, take notes on where people get stuck — because those repetitions are the specification, and without them you are asking AI to invent your method, which it will do confidently, out of everybody else’s. And if nobody has asked you for it, go have the conversations first, because an unused asset costs more than an unbuilt one in the way it sits there reminding you of the weekend.

And if your positioning is still fuzzy, a beautiful assessment will produce fuzzy results faster and in greater volume, which is a worse problem than the one you started with. I have watched brilliant women spend a month building the machine so they would not have to make the decision underneath it, and I have been that woman, so this comes with no judgment — but the fix there is the decision. Write the one page from earlier, and if you cannot answer the second question, close the laptop and go work on that instead.

What if the first build takes longer than doing the work yourself?

It will, and that is the correct trade rather than a sign you picked wrong, because the first one runs long while you make decisions you have carried in your head for years and never once said out loud — the scoring, the boundaries, the exceptions, what you would tell her if she answered this way instead of that way — and that work is being done once instead of every time.

The return shows up on the second use and every one after it, when the thing is sitting there already decided, and the decisions turn out to be worth having made separately from the tool, because that is your method finally written down in a form you can hand to a team member or a collaborator.

Your first version does not need to prove that you can build software, which very few of us have anything to gain from proving anyway. What it needs to show is that one specific piece of what you know becomes more useful to another person when she can interact with it directly, and that is a much smaller thing to demonstrate than it sounds.

Pick the one you have explained the most times. Build the manual version this week. If you want help deciding which piece of your expertise is ready to become something, that is most of the work I do with clients, and it usually takes less time than people expect.