Using AI for Project Estimation Without Historical Data.

I’ve experimented with this in environments where historical data was limited, and the key for me was not using AI to produce estimates—but to build a model we could interrogate.

How I used AI (practically):

I started with a structured prompt to generate a first-pass decomposition

Not timelines—just:

Major workstreams
Key dependencies
Integration points
Assumptions

Then I iterated on that with context

Feeding in:

Team structure (who does what)
Known constraints (capacity, skill gaps, external dependencies)
Delivery model (Agile, hybrid, etc.)

From there, we used it to simulate scenarios:

“What happens if this dependency slips?”
“Where are likely bottlenecks?”
“What requires coordination across teams?”

At that point, it stopped being estimation and became risk shaping.

What worked well:

Rapid decomposition of ambiguous work
Surfacing dependencies earlier than we typically would
Giving teams a starting point instead of a blank page

What didn’t work:

Timeline accuracy was consistently optimistic
Effort estimates ignored coordination overhead
It had no concept of how decisions actually flow in the organization

What made the difference:

We didn’t ask AI, “How long will this take?”

We asked:

“Where is this most likely to break?”
“What assumptions are we making?”
“What would cause rework?”

Then we built estimates around those answers, not from the initial output.

So in my experience:

AI is useful without historical data—but not because it replaces it.

It helps you expose assumptions faster.

And when you don’t have history, assumptions are all you really have.

Curious if others are using AI more for estimation or for risk modeling—because I’ve found the latter is where it actually adds value.

Yes, AI has been successfully used as a supporting tool for estimating complex software projects without historical data, but not as a standalone estimator. In practice, AI works best to help decompose scope, explore assumption‑based effort ranges, and generate multiple timeline scenarios. It adds real value in identifying uncertainty and risk drivers, but tends to be overly optimistic if used without human validation. The most reliable results came from combining AI outputs with professional judgment, explicit assumptions, and contingency planning—using AI as an input to decision‑making rather than a source of final commitments.

AI can help even without historical data, but it should guide, not decide. I use it to break scope, suggest ranges, and surface risks, then validate with expert judgment. Start with assumptions, build scenarios, and refine iteratively. It works well for structure and speed, but reliability depends on how well you question and adjust the outputs.

Even with historical data I wouldn't implicitly trust the output. However, if I had to choose between an individuals single-point estimate and a Monte Carlo estimate performed by GenAI, using the same data the individual had AND the data the GenAI was trained on, I would be more likely to trust the GenAI response, because - and this is critical - the GenAI Monte Carlo analysis would be providing an estimate range with a level of confidence and could likely better take into account where the project is on the cone of uncertainty. The individual providing the single-point estimate might be right, but depending on the project phase, it could still be a highly uncertain guess that doesn't take constraints or dependencies into account.

I'm not saying I would fully trust the GenAI Monte Carlo estimate, but I would think it more likely. In either case, I would refine the estimate as we learn more throughout the project.