This is not an “AI is amazing” post
You’ve read enough of those. Instead, here’s what actually changed in our delivery workflow in the last 18 months, stage by stage — and why clients get the benefit without having to care about the details.
We’re an agency that builds custom systems: ERPs, CRMs, e-commerce, mobile apps, RAG wikis. For that kind of work, AI tools shift the economics of certain stages a lot, and other stages almost not at all. The 30–40% improvement in overall timeline is the weighted average of a much more uneven picture.
Where AI saves the most time
1. Discovery and specification writing.
After a scoping call, one of our senior engineers used to spend a day to a day-and-a-half writing up the functional spec, data model sketches, and user flow diagrams. That’s now closer to two hours. The engineer still does the thinking, but the typing, formatting, and “turn notes into a structured document” part is done in minutes.
2. Scaffolding and boilerplate.
A new Laravel module with auth, CRUD, validation, tests, and admin UI used to be half a day of structured-but-boring work. Now it’s 30 minutes. Same for a React + TypeScript component library starter, a FastAPI + Pydantic schema layer, or a Flutter screen stack. The code is still reviewed and shaped by a human — but we skip straight to the interesting parts.
3. Translating requirements across layers.
Given a clear spec, generating the matching SQL schema, the ORM models, the API endpoints, and the front-end types is exactly the kind of repetitive, error-prone work AI excels at. What used to involve four files kept in sync manually now happens in one flow.
4. Test coverage.
Writing unit tests used to be a discipline issue — everyone knew they should, few teams did. AI-assisted test generation makes the first-draft tests cheap. Human review makes them real. The net effect: better coverage than before, at less cost.
5. Documentation.
README files, API docs, user guides, handover documents. All of these used to come at the end of a project when the team was tired and the budget was tight. Now they’re written alongside the code — because the marginal effort has collapsed.
Where AI does not help
This is the part most agency blogs skip. Being honest here:
1) Architecture decisions. “Should this be a monolith or microservices? Should we use Postgres or Mongo? Should this be event-driven or request/response?” AI can discuss these. It can’t decide them for your specific business. That remains senior engineering judgement.
2) Understanding the actual problem. If you don’t understand your client’s business deeply, AI will confidently help you build the wrong thing faster. Discovery conversations still matter.
3) Security-sensitive code. Anything touching authentication, authorisation, payments, personal data, or privilege boundaries gets human-first treatment. AI can assist with drafts, but production sign-off is manual.
4) Final code review on production-facing systems. The last 20% of the review is still a senior engineer reading code carefully. This is where we find the subtle issues AI also missed.
5) Design judgement. “Is this UI clear? Is this brand consistent? Does this flow feel right?” — that’s still human work. AI can generate candidates. Taste is still ours.
How this shows up for clients
Three concrete ways:
1. Shorter timelines. Most medium-sized projects that were 10–12 weeks two years ago now run 7–9 weeks. The front-of-project discovery phase shrinks the most; the back-of-project QA and polish phase shrinks the least.
2. Tighter scope on fixed-price work. Because boilerplate and testing cost less, fixed-price engagements can include more coverage for the same number — features that would have been “phase 2” stay in phase 1.
3. Better documentation at handover. The READMEs, architecture diagrams, and admin guides are actually produced, not promised. This is probably the most durable benefit for clients long after the project ends.
Note what’s not on this list: “cheaper.” Sometimes it is, for certain project shapes. But we don’t lead with that, because the lion’s share of project cost is senior thinking, client conversation, and careful QA — none of which collapse just because the scaffolding phase shrinks.
The things we still insist on
For anyone shopping for an “AI-powered agency” as a cost-cut: be careful. The market is filling with shops that use AI to skip steps rather than accelerate them. The failure mode shows up six months after launch, when something breaks and nobody understands the system that was shipped.
What we still insist on, internally:
1) Every AI-generated piece of code is read and understood by a human before it’s committed.
2) Production deployments require human sign-off on a written deployment plan.
3) Security-sensitive paths are covered by human-written tests, not just AI-generated ones.
4) Client calls are done by engineers who built the system, not by handing it off to whoever’s free.
What you should ask your development partner
If you’re evaluating agencies right now, ask the following two questions. The answers tell you whether they’re using AI well or just marketing it:
1. “Can you show me an example where you chose not to use AI for a particular stage of a project, and why?” — an agency that can’t answer this is probably not exercising judgement.
2. “How do you handle the quality bar on AI-generated code for security-sensitive paths?” — an agency that says “the same as any other code” is being careless.
Done right, AI makes a good team noticeably faster at the boring parts so they can spend more time on the parts that actually matter to you. That’s the whole story.
Next step: Curious how this would translate to your specific project? Send us a quick description and we’ll come back with a realistic timeline — with and without AI-assisted delivery, so you can see the difference.
