Summarize with AI
A good how it works section for an AI product follows one real input through the product, step by step, and shows what happens to it at each step that a general chat tool could not do. Three numbered circles reading Connect, Analyse and Get insights describe every AI product ever shipped, so they tell a stranger nothing about yours.
That matters more than it sounds. The how it works block is the only place on a landing page built to explain the mechanism, which makes it the only place your eighteen months of engineering can show. If this section is generic too, a visitor leaves with one guess about your company, and the guess is usually wrapper.
Upload, Classify, Export: the section a customs tool ships
Picture an AI tool for importers. This is an example, not a client. A clothing brand uploads the commercial invoice for a shipment from its factory, and the tool proposes the tariff code for every line, the code that decides how much duty the brand pays at the border.
Its how it works section sits third on the homepage. Three circles with icons: a cloud with an arrow, a sparkle, a download tray. Under them, Upload your documents, Our AI classifies every product, and Export to your broker. Each card has one line of grey text. The button below says Start free trial.
Now picture what the team spent eighteen months on. The tool matches each invoice line to the brand's own product catalogue. It checks the code against every past shipment the brand has cleared. It applies the textile rules, where the fibre that weighs the most usually decides the code. It cites the published classification rulings it relied on. And it holds back any line it is not sure about, with the reason, for a licensed broker to decide. None of that is in the three circles. The only step that mentions the hard part says Our AI.
Why this is not the same job as your features section
Features and how it works get treated as two versions of one block. They are not. A features section makes parallel claims, side by side, and we cover how to write those as mechanisms in features vs benefits on an AI product page. A how it works section has a clock in it. It follows one thing from the moment it arrives to the moment the buyer gets it back, in order.
Most advice says keep it to three short steps so the product looks easy. For a settled category that works. Nobody doubts an invoicing app can send an invoice. For an AI product, looking easy is the problem. The visitor already suspects it is a model call with a nice font, and three effortless steps confirm the suspicion. The section should look exactly as easy as the product is to use, and exactly as hard as the product was to build.
The rewrite: one invoice line, four steps
Here is the same section rebuilt around a single line from a real looking invoice: Men's crew neck pullover, knitted, 60% cotton 40% polyester, 1,200 pcs. Every step shows what happens to that line.
Step one, Reads the invoice the way your broker does. The visual is the scanned invoice with that line highlighted, and the fibre split and the word knitted underlined, because those two facts decide the code.
Step two, Checks it against your own history. The visual is a matched record: SKU KN-2231, entered as 6110.20 on fourteen past shipments, none queried. This is the step a chat tool cannot copy, because it has never seen your shipments.
Step three, Tests the code against the rules that decide it. The visual is a short trail: Knitted, so chapter 61. Cotton weighs the most, so cotton. Code 6110.20, with the ruling it matched linked underneath. This is the step that took the longest to build, so it gets the most room on the page.
Step four, Sends close calls to a broker, with the reason. The visual is the review queue: 3 of 41 lines need a person. One row reads Line 17: invoice says knitted, your catalogue says woven. Chapter 61 or 62 depends on which is right. Under it sits one button, Approve 38 lines.
Nothing in the four steps says AI, and nothing names a model. A broker reading it can check every claim against what she already knows.
Two more how it works examples, rewritten
The same move works for any AI product with real work behind it. Take a tool that plans the weekly food order for a restaurant group. Before: Connect, Forecast, Order. After: Reads twelve weeks of till data and next week's bookings. Turns expected covers into kilos using each dish's recipe card. Drafts one order per supplier at last week's prices. Holds any line that doubles from last week for the head chef to confirm.
Or a security tool that triages alerts for a small team. Before: Ingest, Detect, Respond. After: Pulls the alert and the logs from the same host around it. Checks whether the same pattern was closed as harmless before, and who closed it. Writes a one paragraph case with the log lines it used. Never closes a critical alert on its own, an analyst does.
In both, each step names something the buyer handles every week, and the last step says where a person comes in.
A visual per step, and it has to be the real thing
An icon per step is a decoration. The visual for each step should be the artefact the product produces at that point: the highlighted input, the matched record, the rule trail, the queue. Crop each one tight so the label you want read is legible on a phone, where the steps stack into a single column.
The Dualite case study shows why this lands. Dualite is an AI builder that turns a prompt into a working app. The product shown on that page does not stop at a box and a result. It shows the prompt, then the files the build created by name, such as game.ts, maze.ts and useGameState.ts, then the terminal finishing yarn install, then the live preview. A visitor watches the work happen in order, which is the whole argument a how it works section is meant to make.
Static frames usually beat a looping video here. A visitor can read a frame at their own speed and come back to step three. A video makes them wait for it.
Show the human check, and keep it short
Founders often cut the broker step, because it seems to weaken the story that the product is automated. For a careful buyer it does the opposite. The person evaluating a customs tool is the one who answers to the border agency when a code is wrong. A step that says exactly when a person decides is the step that makes the other three believable. If your product acts on its own, what an agentic landing page must show covers approval moments in more depth. The rule here is simpler: if a person checks anything, it is a step, not a footnote.
On length, three to five steps is the range. Each step gets a verb led heading under eight words, one sentence, and one visual. If you need a sixth step, you are describing the architecture, and that belongs on a separate technical page linked from the section.
What the wrapper reading costs the founder
Watch what the founder of the customs tool does instead of fixing the section. On sales calls he skips the website and shares his screen on the review queue within the first five minutes, because that is where prospects lean in. After every demo he attaches the same spreadsheet of past entries the model was tested on. He has rewritten the hero twice this quarter. The three circles have not changed since launch week.
When an investor asks what stops a large model provider from adding tariff codes to its chat app, the honest answer is steps two to four. He starts on the rules engine and is still explaining it when the meeting runs out. The question a prospect asks is a version of the same thing, usually phrased as whether it is a chat tool that has read the tariff book.
None of this shows up in a funnel report. It shows up in the price. A seed-stage AI company often carries a valuation roughly 42 percent above a non-AI peer, and the next round now expects around 3.5 million in ARR, up from roughly one million three years earlier (Carta, Q1 2026). When a partner files the company as a thin layer on someone else's model, the multiple follows the filing, not the code. Hiring follows it too. The engineer he most wants opens the homepage before the job post, and three circles tell her the hard problems live somewhere else.
Rewrite your how it works section this week
You need an afternoon, thirty minutes of your lead engineer's time, and one person who has never seen the product.
- Pick one real input from last week. One invoice, one contract, one alert. Not a made up demo case.
- Ask your engineer to narrate what happens to it, start to finish, and record the call. Write down every place they say then it checks.
- Circle the step that took the longest to build. It is usually a check, a threshold or an eval. That step gets the most room.
- Write each step as a verb plus a noun your buyer uses. Reads the invoice, not Ingests data.
- Capture the real screen at each step and crop it until one label is readable on a phone.
- Add the human step with its trigger, in plain words, even if it fires rarely.
- Show the section to the outsider for thirty seconds and ask what happens to their invoice after upload. If they say the AI processes it, the steps are still generic.
If you would rather hand the whole page to someone, our ranking of landing page design agencies for B2B AI SaaS companies compares studios that do this kind of work.
At Studio Maydit we design websites for AI founders in the US, UK and Europe, and the how it works block is often where we find the company's real story sitting unused. The founder can walk a single input through the product on a call, and the page shows three icons. We sit with the team, pick that input, and build the section around it in Framer, Webflow or custom code, usually inside a fixed scope of three to four weeks. If your section still says Connect, Analyse, Get insights, book a 30 minute call and we will sketch the steps for your product with you.
