"If AI can draft a website in an afternoon, why am I paying an agency six figures?" That's the question buyers are typing into search bars right now, and every RFP written in 2026 needs to answer it honestly before it goes out the door.

Here's the short version. AI has reshuffled where the effort in a web project lives. It hasn't made the project cheap, and it has changed what a competent agency should be willing to promise. The old RFP template, built around wireframes and page counts and template selection, now asks agencies to bid on the parts of the work that got easier while ignoring the parts that got harder. Send that template in 2026 and you'll get proposals that look reassuringly familiar and miss the actual risk in your build.

What Did AI Actually Change About Web Project Costs?

AI compressed the repetitive parts of a web build: wireframe variations, first-draft copy, standard component scaffolding, content migration scripts, and a good share of the automated testing. Those line items didn't vanish. They take fewer hours than they did two years ago, and the good agencies are pricing them that way.

Everything that requires judgment stayed put. Positioning. Information architecture that reflects how your buyers actually think. Brand voice that survives a redesign.

Accessibility remediation on legacy content. Performance work on a slow database.

A Harvard Business Review piece on generative AI at work calls this "judgment integration" — the human discernment layered on top of AI output to make it trustworthy — and that's what a good agency now sells. For a longer walkthrough of which line items shrank and which held, the DEV.co discussion of website Design Costs in 2026: What AI Actually Changes and What It Doesnt maps it out clearly.

So What Should the New RFP Actually Ask?

Throw out the parts of your template that assume the agency is quoting hours nobody spends anymore, and add the questions that surface how they handle AI output. A few worth putting near the top:

Why Ask So Much About the Generated Code?

Generated code is where the invisible risk lives. A large 2025 analysis of code produced by more than 100 models found that 45% of AI-generated code samples failed basic OWASP Top 10 security checks, with certain enterprise-heavy languages failing far more often than that. That isn't an argument against using AI in a build. It's an argument for asking who reads the output before it becomes your production site.

The RFP question flows from there. You aren't asking whether the agency uses AI; assume they do. You're asking what layer of review sits between generated code and your customers.

"Senior engineer review on every PR, static analysis on the pipeline, dependency scanning at build" is a real answer. Silence on the question is a red flag.

What About Accessibility, Since Nobody Wants to Talk About It?

Accessibility is the line item most likely to get dropped from an AI-accelerated bid, and it carries the clearest downside. Generated components aren't automatically compliant. Alt text a model wrote isn't automatically useful. Contrast ratios in an AI-suggested palette aren't automatically legal.

Ask the agency for their accessibility standard in writing, the tooling they use, and the manual audit step. Ask who signs off. If the answer is "the AI checks it," get another proposal.