How long does custom AI software take to build?

Integrating AI into your existing operations, connecting your current systems and automating the workflows between them, typically takes 4 to 8 weeks, and building a full custom platform typically takes 10 to 16 weeks. Those ranges hold for most projects we deliver, and in this article I want to explain what happens inside those weeks and what makes a project land at either end.

The first thing worth saying is that these timelines are dramatically shorter than custom software used to take, and shorter than much of the advice online still suggests. Guides written even two or three years ago talk about six to twelve month builds as standard. AI-assisted development has compressed that, not just because code gets written faster but because the discovery and iteration cycles that used to consume months now happen in days. A working version of a workflow can often be demonstrated within the first fortnight, which changes the whole rhythm of a project, since you are reacting to something real early rather than approving documents for months and seeing software at the end.

What happens inside a typical project

Inside a typical engagement, the first week or two is discovery: mapping how the process actually works today, including the exceptions, and getting access to real sample data. The middle weeks are the build proper, usually delivered in working increments so you can test against reality as it takes shape. The final stretch is refinement against the edge cases that only surface once real work flows through the system, plus handover so your team owns what they have.

What extends a timeline, and what does not

What pushes a project toward the longer end is rarely the AI itself. It is integration with older or unusual software, since every system the build touches adds connection work and the poorly documented ones add more. It is process ambiguity, because if nobody can describe how the work is done today, the early weeks get spent establishing that before anything can be built against it. And it is decision speed on your side, which is the quietly decisive factor: a project where questions get answered in a day runs to schedule, and one where every question waits a fortnight for a meeting does not, which is one of several reasons I have argued that ownership of an AI project should sit with the business unit rather than a committee.

What does not, in our experience, extend timelines the way people expect: the state of your data, which modern AI is built to handle as it is, and the size of your business, since a well-bounded workflow takes similar effort in a 10-person firm and a 50-person one.

How to read a quote

The practical advice is to be suspicious in both directions. A quote promising a full custom platform in a fortnight has skipped the discovery that makes the thing fit your operation, and a quote stretching a workflow automation across nine months is padding, priced by the hour. The right shape is working software early, refined against reality, finished in weeks not quarters.

Questions we get asked

Can we run the project alongside normal operations? Yes, and you should. A well-run build needs a few hours a week from the people who know the process, not a seconded team.

When do we see something working? For most projects, a demonstrable working version of the core workflow inside the first two to three weeks.

What happens after launch? A refinement period as edge cases surface, then ongoing maintenance you should budget properly for, typically well under 20 percent of build cost annually for a focused workflow build.