How long an AI project really takes

The estimate says eight weeks. It is month five.

Almost none of that slip is engineering. The parts that overrun are the parts nobody put on the plan, and they are the same four parts in most companies.

The honest durations#

An assessment: two to three weeks. Watching how the work is actually done, timing it, and coming back with a ranked short list. What that includes is a separate question.

One automation, clean systems: four to eight weeks. A repeating job, moved into tools that have usable interfaces and somebody who understands them.

A feature inside your product: three to five months. Longer because shipping to customers means evaluation, monitoring, a rollback path and a security review.

A single workflow, deliberately small: about a week. This is what a fixed seven-day piece of work buys, and it exists because the honest answer to "how long" is sometimes "longer than you want to commit to before seeing anything".

The four things that add the months#

Access#

The most common, and the most predictable.

Getting credentials to a production system in a company of any size takes weeks, not days. There is a ticket, an approver who is on holiday, a security questionnaire, and sometimes a vendor who has to be asked. Four weeks is normal and nobody puts it on the plan because it feels like admin rather than work.

What to do. Start the access requests the day you sign, before scoping finishes. It is the one task that can run entirely in parallel and it is almost always on the critical path.

Data that is worse than described#

Everybody believes their data is reasonable. It usually is not, in a specific way that nobody has had to confront because no process depended on it before.

Duplicates, three formats for the same field, a column that means two different things depending on when the row was written. None of it is fatal and all of it is time.

What to do. Look at a thousand real rows during scoping rather than a schema. An afternoon of this moves more estimates than a week of planning.

The security review nobody scheduled#

Anything touching customer data goes through a review in most companies, and in some it goes through a customer's review too. It is a queue with its own timetable and it does not care about yours.

What to do. Find out in week one whether this applies and how long the queue is. If a customer contract requires notification before a subprocessor changes, find that out in week one too.

No decision-maker#

The quiet one. A project stalls for eleven days because a question needs answering and the person who can answer it is in a different part of the business with no stake in this.

What to do. Name one person on your side who can decide, before anything starts. Not a steering group. A person.

Why "when is it live" is the better question#

"How long is the build" has a tidy answer that is frequently irrelevant. "When will this be in production, used daily, with somebody on our side who can change it" has a longer answer and is the one you actually care about.

Ask firms for the second number. The spread between the two is where most disappointment lives, and a firm that quotes the first and delivers the second is not lying, it is answering the question you asked.

Compressing it honestly#

You cannot make the engineering much faster. You can take the four things above off the critical path, and that is usually worth more than any amount of pressure applied to the build.

Start access on day one. Look at real data during scoping. Find the security queue before you need it. Name the decider.

If you want the shortest honest path from nothing to a working thing with a number attached, it is one workflow in seven days. The call is thirty minutes and we will tell you which of the four is going to bite you.