How long should an AI project take?

Two useful numbers and one warning.

A diagnostic engagement should be measured in weeks. If somebody proposes months to tell you what to automate first, they are either padding or they have not done this before, and both are your problem.

A first build should also be weeks, not quarters, because the first thing you build should be deliberately small. If the initial project is scoped at half a year, the scope is wrong regardless of the estimate.

What each phase involves

Discovery. Interviews with the people who do the work, a look at the systems, and a reconstruction of how something moves through your company. A week or two of elapsed time for most small and mid-sized operations, and the constraint is calendar availability rather than effort.

Analysis and roadmap. Ranking candidates by frequency, cost, and risk, making build-or-buy calls, and writing something specific enough to act on. Another week or so. This is desk work and it does not need you, which is why the whole diagnosis fits comfortably inside a month.

First build. Depends entirely on how many systems have to talk to each other. Something self-contained at the edge of your operation is a couple of weeks. Something spanning three systems with a reconciliation problem in the middle is a different project and should be quoted as one.

Testing against reality. The phase people cut and should not. Running the thing against real cases from your business, particularly the awkward ones, before it touches a customer. Days rather than weeks, and skipping it is how you find out about the exception in front of a client.

Live and tuning. Ongoing. The first month of real use always teaches you something the discovery did not, and a supplier who treats launch as completion has left before the job finished.

The five things that cause overruns

Note where four of the five sit.

Nobody senior is available. The biggest one by a distance. Discovery needs answers, and questions that wait a week each add a month to a four-week project. If your leadership cannot commit real attention, delay the start rather than absorbing the slippage.

Your data does not agree with itself. The moment work requires information from two systems, somebody has to decide which record wins when the identifiers do not match. This is the largest technical cost driver in AI work and it is almost never visible in the original estimate, because nobody knows how bad it is until they look.

Access takes weeks to arrange. Credentials held by one person who is away, a security review nobody scheduled, a platform where only an external administrator can grant permissions. Sort this out before day one. It is free to fix in advance and expensive to fix during.

Scope grows because the diagnosis worked. A good discovery finds more than you expected, and the temptation is to fold it all into phase one. Resist. Put it in the roadmap for later and keep the first build small, because a first project that swells is a first project that fails visibly.

Nobody decided who owns the decision. Two departments disagreeing about how a process should work will not be resolved by a project. It will be surfaced by one, in month three, at cost.

Only the second of those is really about technology.

Worth knowing what you should be seeing at each stage as well, because the phase that feels like failure is usually month two: realistic results by phase.

What a suspiciously fast quote means

Speed is not automatically good. Two weeks for something spanning your CRM, your finance system, and a spreadsheet means one of three things: they have not understood the integration, they intend to skip the testing, or the deliverable is thinner than you think.

Ask what happens to the awkward cases. The answer tells you whether the estimate includes reality.

The honest range for our own work

The roadmap runs in weeks: one conversation and a working session, then analysis, then a written plan and priority order. Your total time commitment is a few hours.

A first agent build typically lands around ten days once we know what we are building, because we deliberately start with something contained. Anything I would quote in months would be a project I would advise you to break into pieces first.

The thing worth internalizing: elapsed time in this work is mostly determined by how quickly your side can answer questions and grant access. That is not a criticism of clients, it is just where the time goes, and knowing it in advance is how you avoid the overrun. How to prepare so this does not happen to you.

How the engagement is structured.