AI gets you 80% of a website. Here's what lives in the other 20%

The demo is genuinely impressive. Describe the site in a paragraph, and 40 seconds later there’s a homepage, an about page, a contact form, and a set of images that look broadly on-brand. For a lot of small orgs, that is already more than they had.

Then the site goes live, and the remaining 20% starts showing up.

What does AI actually do well for a website?

AI handles structure, layout, first-pass copy, and placeholder imagery at a pace no human matches - the templated parts of web design that look the same across ten thousand similar sites.

If your site doesn’t exist yet, or exists in a state that actively embarrasses you, an AI-generated site is probably a step up. I’d rather see a clean AI-built site than a WordPress install from 2017 that nobody is maintaining. The floor has been raised, and for a lot of small orgs that’s the win.

What AI doesn’t do is the 20% that decides whether the site earns trust and leads to action.

What lives in the 20% AI doesn’t handle?

Five things, in order of how often I see them missed.

1. Actual search behavior. Who is typing what into Google to find an org like yours, and does the copy on your page match those phrases? AI defaults to smooth, generic language. Search users don’t talk that way - they type fragments, typos, local modifiers. Closing the gap takes looking at real queries in Google Search Console, not a generator’s assumption of intent.

2. The messy parts of trust. Board page, financials, annual report, program outcomes, named people with real bios. These are the pages that convince a funder or a new client you are a real thing, run by real humans, doing real work. AI produces templated versions of all of these, and most of them will read as placeholder until a person writes them.

3. Email deliverability setup. A new domain sending email needs SPF, DKIM, and DMARC configured correctly or the contact form notifications go to spam. The website generator does not do DNS work. Nobody tells you this until a donor mentions they never heard back after filling out your form.

4. Accessibility beyond the defaults. Color contrast, form labels, focus states, alt text that isn’t “image of a person smiling.” WCAG 2.1 AA - the published web accessibility standard - is a set of practices the generator doesn’t apply by default. “It looks fine to me” is not the test.

5. What to actually say to someone who lands there. The hardest 20%. AI writes confident-sounding copy about whatever you give it. It won’t tell you your about page buries the mission statement. It won’t tell you your donation page asks for a ZIP code before the person has decided to give. It won’t tell you your headline is a statement when it should be a question. Those decisions come from knowing your actual audience, not from being good at sentences.

When is the AI 80% enough on its own?

Sometimes. A volunteer committee launching a one-time event microsite. A solo practitioner putting up a placeholder while they figure out what they actually offer. A temporary landing page for a campaign that will sunset in three months. AI is a great fit when the stakes of being wrong are low and the stakes of not having anything at all are high.

It’s a bad fit when the site is the main thing people use to decide whether you’re fundable, hirable, or trustworthy. For those cases, the 80% is the starting point, not the finish line. Someone still has to close the last 20%, and the site won’t tell you where that 20% is hiding.

What does the 80/20 split look like in a real project?

Here’s how I’d sequence a small nonprofit website project that uses AI at the start - roughly four to six weeks of total work.

• Week 1 - AI generates a complete first draft in a morning. Structure, layout, placeholder copy, placeholder imagery. Two weeks of traditional design work compressed into an afternoon.
• Weeks 2-3 - human work on the 20%. Real copy for the pages that carry trust. Search-intent research and headline revision. Email deliverability setup. Accessibility audit and fixes.
• Week 4 - review, test, polish, launch. Same as a traditional project, but with sharper focus because the scaffolding came together faster.
• Week 5+ - the part the AI demo never shows: what happens when something needs to change, who knows how to change it, and whether the handoff documentation exists.

The AI compresses the scaffolding, not the judgment or polish.

---

AI shortens the first half of a website project from weeks to hours. The second half - judgment, trust pages, deliverability, accessibility - takes the time it always did. If anything, it gets sharper, because once the scaffolding is done, the judgment calls are no longer hiding under it.