The gap between a calm AI project and a chaotic one usually opens in the first week, before anyone writes real code. It opens quietly. The client pictures one thing, the engineers hear another, and nobody says the mismatch out loud because everyone is being polite and hopeful. A month later that small gap has turned into a very uncomfortable meeting.
So I treat the kickoff as the most important deliverable of the whole engagement. Not the model, not the first demo. The kickoff. Get that right and most of the drama never happens.
I make everyone describe 'done' in plain words
At the start I ask each person in the room to finish one sentence: 'This project is a success if...'. The client's answer and the delivery team's answer are almost never the same on the first try, and that difference is the gold. Someone will say the win is a system that answers customer questions accurately. Someone else will say it needs to handle the top twenty question types we see every week. Those are different promises, and pinning down which one we are making saves everybody weeks of guessing.
I write the answers on a shared doc while people watch. Seeing their own words land on the page makes stakeholders slow down and correct me, which is exactly what I want. A vague expectation is easy to nod along to. A written one gets argued with, and I would rather have that argument now.
We list what we are not building
The most useful page in any kickoff is the 'out of scope' list. AI work invites daydreaming. Someone always pictures the assistant summarizing reports, translating five languages, and predicting next quarter's demand, all in the first version. None of that is bad to want. It simply cannot all ship at once.
I name those wishes, thank people for them, and park them in a clearly labeled 'later' section that everyone can see. Nothing gets dismissed. It gets sequenced. When a feature reappears mid-project as something the client assumed was always included, I can point back to the list we all looked at together. That single page has cooled more tense moments than any status report I have ever sent.
The data reality check comes first, not last
The fastest way to blow an AI timeline is to discover in week six that the data everyone assumed was clean is actually full of duplicates, missing fields, and three spellings of the same customer name. I refuse to let that be a surprise. During kickoff we look at real samples together, not a tidy example someone prepared the night before.
I would rather deliver honest bad news on day two than a delayed miracle in month three. That habit of surfacing the ugly parts early is the standard we hold at Shanti Infosoft, and it is the one I hold myself to on every engagement. A little more on how we work sits at https://shantiinfosoft.com.
A kickoff done well is a little uncomfortable. People realize the thing is bigger than the sentence they pitched it as, and that realization is the whole point. Everyone walks out holding the same picture in their head.
The habits that keep this from unravelling are in the five things AI projects that don't get cancelled do differently.
What is the one question you wish you had asked in a first meeting, before it cost you a month?
Top comments (0)