Think Build Implement Repeat
London, UK +44 7367 067226
WhatsApp FOLLOW f in X
AI Apps

Realistic Timelines for an AI Application

Last updated:

What the weeks are spent on

PhaseDurationWhat happens
Scoping and evaluation set2–3 weeksNarrowing, and agreeing correct answers on real cases
Core build4–8 weeksIntake, retrieval, model, validation, integration
Review interface2–3 weeksFrequently underestimated, determines adoption
Shadow running2–4 weeksProducing output nobody acts on, compared against humans
Launch and tuning2–4 weeksThreshold tuning with real traffic

The evaluation set is on the critical path

Agreeing correct answers for a few hundred real cases requires a domain expert's time, and it is the single most common cause of delay in AI projects. Book that time before the project starts.

It also frequently produces the most valuable finding: that your own experts disagree on 20% of cases, which caps what any system can achieve.

What actually causes delay

  1. Real data arriving late, so testing happens on invented examples
  2. Nobody available to say what correct means
  3. The review interface not in the original scope, discovered at week eight
  4. Third-party access — API keys and permissions from someone else's IT team
  5. Scope drifting from one document type to four during the build

How to compress safely

Cut scope, never evaluation. One document type instead of three. One team instead of the whole business. Suggestions instead of automatic actions.

Also: start shadow running before the interface is polished. Real output compared against humans is worth more than a finished screen.

What you can do to make it faster

Provide real data in week one, including the awkward cases. Book your domain expert for the evaluation work. Answer questions within a day. Request third-party credentials before they are needed.

Clients who do these finish two to three weeks earlier on identical scopes.

Frequently asked questions

Can it be done in four weeks?

A proof of concept on real data, yes, and it will not be production-ready. Distinguishing the two clearly at the start prevents a difficult conversation at the end.

Why does the review interface take so long?

Because it has to be genuinely fast to use, which is a design problem rather than a coding one. Getting it wrong means people stop reviewing, which defeats the system.

Does more people make it faster?

Rarely. Small AI projects have limited parallelism and adding people mid-project slows things while they learn the domain.

What if the model improves during the build?

We rerun the evaluation set and take the improvement if it is real. That is an afternoon, because the provider is abstracted and the test set exists.

Keep reading

Have a date you need to hit?

Tell us the date and what must work by then. We will tell you what fits and what should wait for phase two.

Book a free 30-minute call Get a project estimate WhatsApp us

Related services

What we build for problems like this one

AI AgentsCustom Software DevelopmentSaaS Development