Skip to content

03 · Understanding AI Project Lifecycles

An AI project's lifecycle looks nothing like a typical feature build's linear "design → build → ship" flow. It loops, it can dead-end at multiple points for legitimate reasons, and the amount of time spent in each phase is much harder to predict up front. This module gives you a management-level map of the full lifecycle — from a raw idea to a monitored production system — and the specific decision gates you, as the manager, are responsible for at each stage.

1. The AI project lifecycle, stage by stage

Stage Purpose Typical duration Manager's key decision
1. Problem framing Turn a business ask into a specific, measurable ML/AI problem Days to 2 weeks Is this even a good fit for AI, or is a simpler rule-based solution sufficient?
2. Data discovery Figure out what data exists, where it lives, and whether it's usable 1–4 weeks Do we have enough data, or does this become a "collect data first" project?
3. Feasibility / proof of concept A quick, throwaway-quality experiment to test if the approach can work at all 2–6 weeks Go/no-go: does the early signal justify further investment?
4. Model development Iterative building, training, and evaluation of candidate models Weeks to months Have we hit the pre-agreed quality bar, or do we need another iteration (or to reset scope)?
5. Integration & productionization Wrapping the model in real infrastructure: APIs, monitoring, fallback logic Weeks to months (often underestimated) Is this ready for real users, or does it need a staged/limited rollout first?
6. Launch Releasing to real users, often gradually Days to weeks What's the rollback plan if something goes wrong?
7. Monitoring & iteration Ongoing tracking of live performance, retraining as needed Indefinite — this never "finishes" Is the model still meeting the bar, or has drift eroded it?

The critical mental shift for a new AI manager: stages 3 and 4 are expected to sometimes fail or reset. A proof of concept that shows the approach doesn't work is not a failed project — it is the project doing its job of preventing a much more expensive stage-5 failure. Budget explicit "kill criteria" at the end of stages 1–3 rather than treating every project as guaranteed to reach launch.

2. Where AI lifecycles diverge from standard software delivery

Standard software delivery AI/ML project delivery
Requirements are mostly knowable up front Requirements include "is this even achievable," which is answered during the project
A sprint's "done" criteria is binary (feature works or doesn't) "Done" is a quality threshold agreed before work starts (e.g., 90% precision)
QA happens once, near the end Evaluation happens continuously, at every iteration
Ship and move on Ship, then monitor indefinitely — the model is never "finished"
Bugs are fixed by changing code Quality issues are often fixed by changing data, not code

3. The stage-gate decision framework

A practical tool for running this lifecycle: define, in writing, the criteria that must be true to move from one stage to the next, before the stage begins. This turns "the model isn't good enough yet, should we keep going?" from a political argument into a pre-agreed checklist.

Gate Sample criteria to move forward
Problem framing → Data discovery A single, specific success metric is written down and agreed with the business sponsor
Data discovery → Feasibility Data of sufficient volume and quality exists, or a plan (and budget) to collect it is approved
Feasibility → Development The proof of concept beats an agreed baseline (see Module 2) by a meaningful margin
Development → Integration The model meets the pre-agreed quality bar on a held-out test set that reflects real production conditions
Integration → Launch Monitoring, rollback, and a fallback (non-AI) path are all built and tested
Launch → steady state A first monitoring window (e.g., 2 weeks) shows no unexpected degradation or safety issues

Worked example

A logistics company wants to predict delivery delays. The AI manager sets the success metric at "problem framing": predict delays of 2+ hours with at least 75% precision, evaluated on the next quarter's real shipments. During "data discovery," the team finds that delay-cause data is inconsistently logged across regions — the gate isn't met, so the project pivots to a 2-week data-cleanup effort before feasibility work even starts, rather than building a model on unreliable labels. The proof of concept then hits 71% precision — just under the bar. Rather than quietly lowering the bar to call it a win, the manager convenes the team: is 71% close enough that another 2-week iteration likely closes the gap, or is a different feature set needed? They choose one more iteration, hit 78%, and only then move into integration. The stage gates made three real decisions visible and deliberate instead of implicit.

How It Actually Works

Stage gates work as a management tool because they exploit an asymmetry in how information becomes available during an AI project: uncertainty about achievability is highest at the start and only resolves through direct empirical work, not through more planning meetings. A software requirements document can be gotten "mostly right" by thinking harder about it, because the constraints are usually about human/business logic that a smart person can reason through in advance. A model's achievable precision on real data cannot be reasoned to in advance — it can only be discovered by actually training something and measuring it against held-out examples, because the relationship between the available data's signal-to-noise ratio and any model's achievable ceiling is not analytically knowable ahead of time. This is the mechanical reason "kill criteria" at stages 1–3 aren't pessimism — they're the only points where the cost of learning "this doesn't work" is still cheap, because a failed feasibility experiment costs weeks, while discovering the same ceiling after building full serving infrastructure in stage 5 costs months and carries sunk-cost pressure to ship anyway.

The 71%-vs-75% moment in the worked example illustrates the specific trap a gate is designed to prevent: without a pre-agreed number, "close enough" is a social judgment made under the exact conditions (deadline pressure, sunk cost, a team that wants a win) most likely to bias it downward. Writing the threshold down before the proof-of-concept results exist removes the optimization-after-the-fact problem — nobody can quietly redefine "done" to match whatever number the model happens to hit, because the metric was fixed when no one yet knew which way it would cut. This is also why held-out evaluation data matters mechanically at the Development → Integration gate specifically: a model's score on data it was tuned against is not evidence of real-world performance, only of how well it fit that data — the gate criterion is only meaningful if the evaluation set was truly never used to make any modeling decision.

Exercise

Pick a real or plausible AI project. Write out all seven lifecycle stages from the first table, and for each one write: (1) a realistic time estimate given your organization's actual constraints, and (2) one specific, written gate criterion that must be true before moving to the next stage. Then identify which single gate is most likely to be skipped or rushed under deadline pressure at your organization, and write one sentence on how you'd defend keeping it.