Build vs. buy: when to add AI to your website (and when to wait)

A consultant friend asked me last week: "Should we add AI to our website?"

The answer in the abstract is "probably yes, eventually." The answer in the specific is "depends on five things, and three of them are about your business — not the AI."

This post is the five questions, in the order I'd ask them.

The framework, briefly

Before adding AI to your website, answer:

Do you understand the inquiry shape — what visitors come to do?
Do you have a structured definition of a "good lead" vs. a "bad lead"?
Are you currently shipping intake well without AI?
Will the AI live for ≥18 months, or are you experimenting for 3?
Can you afford the maintenance budget — or is this a one-time project?

If the answers are yes/yes/yes/yes/yes, add AI. If two or more are no, the AI project is probably premature; you have smaller, cheaper work to do first.

Let's go through each.

1. Do you understand the inquiry shape?

Most websites don't. They have "contact us." They don't have "we serve four kinds of inquirers, in this order of frequency, with these characteristic needs."

If you can't articulate the inquiry shape in a paragraph, AI cannot articulate it for you. AI is a generalist that becomes a specialist by reading specifics. Without a paragraph that says "our visitors are usually trying to do X, Y, or Z; the most valuable ones are doing X; here's how to tell," the AI ends up being a generalist on your site — slightly worse than a thoughtful static form.

Quick test: write the paragraph in 30 minutes. If you can, you're ready. If you can't, do the homework first. Spend two weeks watching your incoming inquiries and writing the paragraph. The paragraph is the project; the AI is the artifact.

2. Do you have structured "good lead" vs. "bad lead"?

This is a related but separate question. The first asks "what comes in." This asks "what should we do about it."

Without structured criteria — concrete, written down, agreed-upon — the AI can't help you triage. It will produce summaries and sort attempts that sound plausible but don't match your actual decision-making, because your actual decision-making is partly a feeling. AI can't model a feeling well.

Make the feeling explicit. Write the criteria as a checklist. Test it on the last 30 inquiries — does the checklist correctly predict which ones the team picked up? If not, refine the checklist. Repeat until 80%+ accurate. Now you have a definition the AI can encode.

This is also genuinely good work for the business; you'll find your team's gut-feel was inconsistent, and codifying it tends to improve close rates with or without AI.

3. Are you shipping intake well without AI?

If your current intake is broken — abandoned forms, slow response times, leads that disappear into email — adding AI is fixing the wrong layer.

The same problems will exist with AI; they'll just be expressed in fancier ways. The form is still abandoned (now it's a chatbot that's abandoned). The response time is still slow (now the AI summarizes faster, but the human still ignores the summary). The leads still disappear (into a different inbox, or worse, into a JSON blob nobody reads).

Fix the human-process layer first. Then bring AI in to multiply what's working. We say this constantly; it's the single hardest piece of advice for founders to internalize because the AI sounds like a fix for the human problem. It usually isn't.

4. ≥18 months of life, or 3-month experiment?

AI implementations are heavier than they look. The model is the easy part. The maintenance — prompt updates, schema drift, vendor changes, model upgrades, edge cases that slowly accumulate — is the part that surprises everyone.

If the AI you're considering is a 3-month experiment, just don't. The integration cost won't pay back; the team will spend its time hand-holding and its sanity in burnout-by-incident. Use the same budget on a thoughtful static form, or on better human-process design, or on a one-time content project.

If the AI is something you'll live with for ≥18 months, the maintenance investment makes sense. You'll get a year of compounding benefit.

For us, the litmus is whether the founder commits to the team a rough two-hour-per-month maintenance budget for the AI, or to a vendor relationship that does it for them. Without one of those two, the AI rots.

5. Maintenance budget — or one-time project?

A static form, well-built, can run for five years untouched. An AI form needs attention. Concretely:

Prompt updates: ~30 min/month for tone drift, edge cases, new product mentions.
Schema reviews: ~30 min/quarter to confirm the brief shape still matches the team's needs.
Vendor changes: ~2 hours/year (one model upgrade, one pricing change, one breaking API change — at least).
Edge case fixes: ~30 min/month for the surprises ("the AI told a patient we accept Aetna; we don't").

Total: ~2 hrs/month of focused attention by someone who understands both the business and the AI. If you have that person — staff or vendor — proceed. If you don't, you're building a beautiful machine that will quietly break.

This is also the question that decides "build" vs. "buy."

Build vs. buy

If you have the maintenance budget and the AI is core to your competitive advantage, build (or commission a build).

If you have the budget but the AI is a nice-to-have rather than a moat, buy off-the-shelf.

If you don't have the budget, neither — defer.

Most small businesses fall into "buy off-the-shelf" or "defer." That's fine. Bespoke AI is for the small fraction of businesses where the AI itself differentiates them; for everyone else, the static-form-plus-good-process beats the bespoke AI 80% of the time.

When to wait six months

Sometimes the right answer is "you're 80% ready; do the 20% first."

Common 20%-still-needed work:

The inquiry-shape paragraph (1–2 weeks, free)
The good-lead checklist (1 week, free)
The data foundation engagement (2 weeks, $2k–$8k — see the dedicated post)
A simple static form that runs for 90 days to validate the schema (1 week, included if you're hiring a studio)

Total: 4–6 weeks of preparation. After that, the AI engagement is dramatically more likely to succeed and dramatically cheaper to maintain.

If a vendor tells you you don't need any of this, find another vendor. The work doesn't go away; it gets paid for in failure mode after the fact.

Our position

We sell AI integration services. We're also frequently the studio that tells founders "you're not ready; do these three smaller things first; come back in two months."

That sounds like leaving money on the table. It's not. The clients we tell to wait come back twice as committed and ready, and our project succeeds; our reputation compounds.

If you'd like a candid take on whether your business is ready, the way to find out is a 30-minute call. We don't sell on those calls; we tell the truth about your stage. Sometimes the truth is "yes, ready, here's what we'd build." More often it's "wait six months; here's the smaller thing to do first."

Curious where you stand? Book the 30-minute call. The worst case is honest feedback; the best case is a scoped project.

Like this kind of writing?