Our homepage says we're an AI-augmented software delivery agency. So does almost every agency you're evaluating. The phrase has been through the same cycle as "agile" and "full-stack": useful for about eighteen months, then adopted by everyone, then worth nothing as a signal.
That's a problem for you more than for us. You're trying to work out which vendors have actually changed how they build software and which ones added a line to their about page. The adjective doesn't help you tell them apart.
So here is the specific version. Where AI sits in our delivery, where it doesn't and where a senior engineer signs their name regardless.
Where we use it
Code review, first pass. Every pull request gets an automated review before a human opens it. It catches the boring things: unhandled null cases, inconsistent error handling, a migration that will lock a table under load. The value isn't that it finds bugs a senior wouldn't. It's that the senior starts their review at the interesting part instead of spending fifteen minutes on lint-adjacent commentary. Tuning this so it doesn't drown the team in noise took real work, because the first version did exactly that.
Test generation for known-shape code. CRUD endpoints, form validation, data transformations. Anywhere the correct behaviour is obvious and writing it out is tedious. Coverage on this kind of code used to be the thing that quietly slipped when a deadline moved. Now it doesn't.
Discovery synthesis. Twelve stakeholder interviews produce a lot of transcript. Clustering it into themes is genuinely faster with a model, and it surfaces contradictions between what two departments told us that we'd otherwise notice in week three instead of week one.
Migration and refactor grunt work. Renaming a concept across four hundred files, converting a component library, updating deprecated API calls. High volume, low judgement, easy to verify.
Documentation that would otherwise not exist. Not the architecture decisions, those need a human who was in the room. The reference material: endpoint documentation, component props, environment setup. The stuff every team agrees is important and nobody writes.
Where we don't
Architecture. A model will confidently propose a multi-tenant design without knowing that your biggest prospect has a data residency requirement that rules it out. The information that makes architecture decisions correct usually isn't in the codebase. It's in a conversation someone had with sales.
Anything touching money, health data or identity. Not because a model can't write a payment integration. Because the failure mode is unacceptable and the review burden ends up larger than the writing burden. If we have to read every line with full attention anyway, generating it first saved nothing.
Scoping and estimation. Estimates carry commercial commitments. A model has no stake in the number and no memory of the last three projects that ran over for the same reason.
Client-facing writing without a human rewrite. Proposals, technical explanations, this article. A first draft, sometimes. Shipped as-is, never: partly on quality, mostly because the voice flattens into something that reads like every other vendor.
The line that doesn't move
A senior engineer approves every merge to main. A design lead approves every screen that reaches a user. A project manager owns every date given to a client.
That's not a hedge. It's the actual product. What you're buying from an agency is not typing. Typing has been getting cheaper for two years and will keep getting cheaper. You're buying the judgement that catches the thing which was plausible and wrong.
AI has made production dramatically cheaper. It has not made being wrong any cheaper. If anything the opposite: it's now possible to generate a large volume of confident, well-formatted, structurally sound code that solves the wrong problem. The bottleneck moved from writing to verifying, and verification is a senior activity.
This is why we run a senior triad on every project: tech lead, design lead and project manager. Not a pyramid with juniors at the base. The pyramid was built for an era where the base did the volume work. The volume work is now largely automated. What's left is the part the pyramid was never good at. The cost arithmetic of that shift is its own article.
What this should change in how you buy
Ask any vendor claiming AI augmentation three questions.
"Which parts of your delivery use it, specifically?" A real answer sounds like a list with boring items on it. An unreal answer sounds like a philosophy.
"What have you decided not to use it for, and why?" This is the more revealing question. A team that has actually integrated AI into delivery has hit its limits and formed opinions about them. A team that hasn't will tell you it's useful everywhere.
"Who signs off, and on what?" If the answer involves reviewing AI output in batches, or a review process that got lighter rather than more focused, that's the tell.
You should also expect the savings to show up somewhere you can see. Ours show up in scope: an MVP that was a fourteen week build two years ago is an 8 to 12 week build now, at a similar team size. If a vendor tells you AI has transformed their delivery and their timelines and prices are identical to 2023, one of those two claims is wrong.
The honest caveat
Some of this will be outdated within a year. The line between what a model can be trusted with and what it can't has moved several times since we started, always in the same direction, and we've moved with it each time. The list above is where we've landed as of now, not a permanent position.
What won't change is the structure of the claim: a vendor should be able to tell you which parts, and should be able to tell you what they've ruled out. If the answer is the adjective on its own, you've learned nothing about them, which is itself worth knowing.
Building something where the AI question actually matters? Tell us about your project and we'll tell you which parts apply to yours.
Was this useful?
