
Buying a model is the easy part. A license, an API key and a demo can be in place within days. What decides whether an AI project ships is everything underneath: whether the data is complete, whether people agree on what it means, and whether anyone owns it when it breaks. This checklist covers the work that tends to matter most, in the order we would do it.
1. Agree what the key terms mean
Ask three people for last quarter’s revenue, active customers or on-time delivery and you will often get three answers. Each can be correct under a different definition. A model built on one definition and queried with another will look unreliable when the real problem is the data. Write down the handful of terms the use case depends on, agree one definition for each, and publish it where analysts and engineers will actually look.
2. Find where the data really lives
List every source the decision depends on, including the unglamorous ones: spreadsheets on shared drives, attachments in email, exports from a supplier portal. For each, note the system of record, who updates it and how often. Anything that arrives late, by hand or in an inconsistent format is the first piece of engineering work, because no model can compensate for an input that is missing on the day it is needed.
3. Make the pipelines boring
A reliable pipeline is automated, tested and monitored. It runs on a schedule, validates what it receives, and alerts a named person when data is late, empty or outside the expected range. Boring is the goal: if someone has to rerun a job by hand every Monday, the pipeline is not finished, and the AI feature that depends on it will fail on the Monday nobody remembers.
A model can only be as consistent as the definitions and pipelines underneath it.
4. Name an owner for every dataset
Every dataset that feeds a model needs an owner, a short runbook and a clear access policy. The owner decides what a change means for downstream users, answers questions and fixes issues. Without ownership, quality drifts quietly, and trust is lost long before anyone files a ticket.
5. Check quality where it hurts
You do not need to clean everything. Measure the fields the use case actually relies on: completeness, duplicates, freshness and obvious outliers. Set a threshold for each, track it over time, and decide in advance what happens when a check fails, whether that means pausing the feature, falling back to a person or flagging the output.
6. Settle access and privacy early
Decide who and what may see which data before the first prototype, not after. Use least-privilege access, mask or remove personal data that the use case does not need, and keep an audit trail of what the system read and produced. These decisions are cheap at the start and expensive to retrofit.
Start with one use case
Run this checklist for the inputs of a single use case rather than for the whole company. You will learn what your data can support, produce reusable definitions and pipelines, and have a working baseline to measure against. The second use case then starts from a stronger position than the first.
Want to apply this to your own project?
Talk to our Data Engineering team about where to start.


