Most of what’s been written about AI and web development misses the actual story. The narrative arc that dominated 2024 and 2025 — “AI will write all the code, developers are obsolete, your website will build itself” — turned out to be wrong on every count. The 2026 reality is messier, more interesting, and considerably more valuable to SMEs paying attention.
Here’s what’s actually happening: AI didn’t replace web developers. It compressed the timeline of what one good developer can ship in a week. The agencies that figured this out early are charging less per project while taking on more projects. The agencies that didn’t are losing engagements to competitors who can deliver in three weeks what used to take six. And the SMEs caught in the middle of this transition are the ones with the most to gain — if they understand which parts of “AI-powered web development” are real and which are still mostly marketing.
What AI is actually good at, in production
The wins are concentrated in five areas. They’re unglamorous individually. Together, they reshape how a project gets built.
Code scaffolding. Setting up a new Next.js project with authentication, database schema, deployment config, and a component library used to take a senior developer most of a day. With Claude or GPT generating the initial scaffolding from a prompt, it takes about an hour. The output is rarely production-ready as-is, but the structural decisions — how to organise routes, where to put utility functions, how to wire up the auth flow — are handled, leaving humans to make the choices that matter.
API integration boilerplate. Connecting a Shopify storefront to a third-party inventory system used to be a multi-day exercise in reading documentation, writing transformation logic, handling edge cases, and debugging. AI handles the tedious 80% of that work — request signing, response parsing, error handling, retry logic — leaving developers to focus on the genuinely tricky parts (rate limiting strategy, idempotency guarantees, what happens when the third-party API goes down).
Test coverage. Writing tests is the work most developers least want to do. AI is excellent at it. Generating unit tests for existing functions, writing integration tests against a documented API, producing realistic test fixtures — all of these have moved from “we’ll get to it later” to “trivially included in every project.” Sites we ship in 2026 are better-tested than equivalent sites we shipped in 2023, despite spending less time on testing.
Performance optimisation suggestions. Lighthouse audits are easy to run; understanding what to fix and in what order requires expertise. AI is now genuinely useful at translating a Lighthouse report into a prioritised list of code-level fixes. It catches things humans miss — the unused JavaScript imports, the synchronously-loaded font that’s blocking first paint, the layout shift caused by an image without explicit dimensions.
Documentation. Every project we ship now includes documentation that would have been “out of scope” three years ago. AI generates it during development, not as a separate post-launch task. The result: clients can actually understand the systems they paid for, internal team members can be onboarded faster, and the next agency that touches the work has a fighting chance of doing so without asking us to explain everything.
What AI is still bad at
The list is shorter, but each item matters.
Architectural judgement. Choosing between WordPress, Webflow, Next.js, and Shopify for a given project requires understanding business context that AI doesn’t have. When we describe a project to Claude or GPT and ask “what should we build this in,” the answer is consistently reasonable and consistently wrong. The right choice depends on the client’s team capability, content velocity, growth trajectory, integration requirements, and twenty other things that don’t fit in a prompt. We make these calls in person, with senior judgement, and have for thirteen years.
Performance at the architecture level. AI is great at micro-optimisations within an existing architecture. It’s terrible at the architectural decisions that determine whether a site can hit its performance budget at all. Choosing static generation versus server-side rendering, deciding what to cache and where, structuring database queries to avoid N+1 problems — these require systems-level thinking that AI hasn’t reliably learned.
Security. The security holes in AI-generated code are subtle and recurring: SQL injection in stitched-together database queries, missing authorisation checks on internal API endpoints, secrets accidentally committed to repositories. Senior developers catch these by default; AI introduces them by default. Production code from AI gets a security review by a human, every time, no exceptions.
Aesthetic judgement. AI-generated UI is recognisably generic. It defaults to the design patterns it’s seen most in training data — which means rounded corners, blue gradients, and the same hero-section layout every other site uses. The brands winning visually in 2026 are the ones whose websites don’t look like everyone else’s. That requires designers with taste, working with developers who can implement it. AI can accelerate the implementation; it can’t generate the taste.
What this means for SMEs hiring an agency
The practical implications for buyers are bigger than most agency pitches acknowledge.
Project timelines should be shorter. A small business website that took eight weeks to build in 2023 should take four weeks in 2026 — assuming the agency is using AI effectively. If your agency is quoting eight-week timelines on small projects, ask why. Either they’re not using AI tools at all (a red flag), or they’re using them and pocketing the time savings instead of passing them to clients (also a red flag, just for different reasons).
Quality should be going up. Faster development should not mean lower quality. The opposite is what good agencies are delivering: more thorough testing, better documentation, more performance optimisation, and more time spent on the architectural choices that AI can’t help with. If quality has dropped at agencies you’ve worked with — more bugs at launch, less testing, sketchier security practices — they’re using AI as a cost-cutting shortcut, not a leverage tool.
Pricing should reflect value, not hours. The agencies still billing hourly are the ones with the worst incentive structure under AI: they make less when AI makes them faster. The agencies pricing on outcomes — fixed-price projects with clearly-defined scopes, retainers tied to business metrics — have aligned incentives with their clients. They keep the time savings as margin while delivering more value per project. This is sustainable; the hourly model isn’t.
Senior involvement should be more, not less. The expectation that AI would let agencies replace senior developers with juniors operating AI tools turned out backwards. Junior developers can’t recognise the wrong AI suggestions; senior developers can. The most productive setup we’ve found is one senior developer, AI as accelerator, no juniors in the production codebase. Agencies trying to run the inverse — many juniors armed with AI, supervised by one senior — are shipping more code, but it’s worse code, and the maintenance debt is starting to surface.
What we’d do if we were on the buying side
If we were a small business hiring a web development partner in 2026, three questions would make or break the engagement before signing.
First: ask the agency to walk you through how they use AI in production. The answers separate the agencies that have integrated AI thoughtfully from those that are either ignoring it or pretending. Vague answers (“we use AI tools to be more productive”) are a signal of ignorance. Specific answers (“we use Claude for initial scaffolding, GPT for test generation, neither for security-sensitive code, all of it gets reviewed by a senior developer before merge”) signal seriousness.
Second: ask what the agency would build for you in WordPress versus a custom framework, and why. The agencies that immediately pitch a custom build are usually solving for their own engineering preferences, not your needs. The agencies that immediately pitch WordPress for everything are stuck in old patterns. The right answer is contextual — and an agency that can’t articulate why they’d choose one over the other for your project specifically isn’t doing the architectural thinking AI hasn’t replaced.
Third: ask how they’ll involve you in performance and security decisions. Both are areas where AI has compressed the work but not eliminated the need for human judgement. An agency that handles them as black-box deliverables you don’t see is hiding a risk. An agency that walks you through their performance budget at kickoff and shows you the security review at launch is treating you like a partner.
The transformation in web development from AI is real, but it’s not the transformation we were sold on. It’s quieter, more useful, and concentrated in the agencies that figured out how to use new tools without compromising the standards that mattered already. For SMEs, the practical effect is a market full of agencies that look more capable on paper than they actually are — and a smaller number of agencies that have genuinely raised their game. The work of hiring well in 2026 is telling them apart.