Checklist

AI data readiness checklist

A practical pass through the data checks that decide whether an AI feature can ship, from ownership and access to test sets and freshness.

By the Cloudverg engineering team · · 5 min read

Who this is for

AI features fail on data long before they fail on model choice. This AI data readiness checklist is for the people who have to prove otherwise: founders, CTOs, IT leads, and operations or finance managers who are about to spend money on an AI pilot.

Read it if a demo already works on a clean spreadsheet and someone has asked what it takes to run on real data. Read it too if you are picking the first AI use case and want to know which data problems will sink it.

Data engineers and analysts will do most of the work here. Bring them in before the vendor call, not after. The checklist is written so a non-specialist can run the meeting and a specialist can answer the questions.

Before you begin

Pick one use case and write it as a decision, not a theme. A decision sounds like this: flag contracts that renew in the next 60 days and need a legal review. A theme sounds like use AI for legal. The narrower the decision, the fewer data sources you need.

Name a business owner and a technical owner for the data. The business owner decides what correct output looks like. The technical owner knows where the tables live, who can read them, and what breaks when a pipeline fails. If those are the same person, say so, but write both roles down.

Get read access to a sample before the first workshop. A schema dump and a few hundred rows tell you more than a slide. Check whether the sample includes real edge cases: missing values, duplicate records, and records that were later corrected.

Agree on what good enough means for the first release. A model that gets the routing decision right on a reviewed test set is useful. A model that gets it right on data it was tuned on proves nothing. Write the target down and keep it next to the test set.

The checklist

The checklist below runs in order. Skip an item only when you can say why it does not apply. Each item should produce a file, a query, or a decision that someone can point to later.

Start with the source systems, then work through access, structure, quality, and test data. The order matters because access decisions often change what data you can use at all. A table you cannot read is not a data source for this project.

  • List every source system the use case touches and name the table or endpoint that is the system of record for each field.
  • Assign one accountable owner per dataset and write the name in the data catalog or a shared sheet, not in a chat thread.
  • Record the update cadence for each source and the time zone it uses, then convert all timestamps to UTC in the working copy.
  • Map access at the field level, not the table level, so the AI service reads only the columns it needs.
  • Create a service account for the AI workload with the smallest set of read permissions that still lets the job run.
  • Tag personal and sensitive columns in the catalog and confirm the tags match what the access policy enforces.
  • Check primary keys for duplicates and nulls, and fix or document every case where a key is not unique.
  • Profile null rates and value ranges on the columns the model will use, then decide which gaps are acceptable and which must be filled.
  • Standardize units, currencies, and status codes in a view or dbt model so the model sees one version of the truth.
  • Freeze a test set before any tuning, label it by hand, and store the labels with the row IDs so you can rerun the same evaluation.
  • Write a query that returns the expected input for one real request, run it, and keep the output as a baseline.
  • Define a freshness threshold per source and set an alert that fires when the data is older than the threshold.
  • Document the retention window and the deletion path for each source so a removal request can be honored in the AI copy too.

What good looks like

Good readiness shows up as boring artifacts. There is a data dictionary with owners and tags. There is a query that returns a known-good input for one request. There is a test set with labels someone reviewed, stored next to the code that uses it.

You can explain the access model to a security reviewer in one page. The service account reads views, not base tables. Personal data is either absent, masked, or covered by a policy that someone can point to.

Freshness alerts fire before users notice stale output. A failed pipeline does not silently feed the model yesterday's data. When the schema changes, a test fails and a person gets a ticket.

The team can estimate cost per request from the number of records read and the size of the prompt. That estimate is a range, not a promise, and it is tied to a specific source volume.

What to do next

Run the checklist against the top use case in one week. Mark each item as done, blocked, or not applicable. Blocked items become the scope of the first sprint, not a reason to delay the pilot.

Fix access and personal data issues first. They change the shape of the project. Quality fixes come second, because a model can tolerate some missing values but not a permission surprise in production.

If you want an outside pass, the Cloudverg AI readiness assessment is $5,500/month, runs two weeks, and covers workshops, a ranked use case list, a data readiness check for the top use case, a reference architecture, and a cost model. It is one option among others. An internal review with the checklist above is another.

Whichever route you take, end with a written go or no-go decision for the first use case and a named owner for the data work. Then build the smallest version that reads real data and writes a result a person can check.

Frequently asked questions

What is AI data readiness?

AI data readiness is the set of checks that show whether your data can support a specific AI feature. It covers ownership, access, structure, quality, freshness, and a test set you can trust. It is not a general data strategy project, and it should produce artifacts you can review.

How much data do I need for an AI project?

Enough data to cover the decisions the feature will make, including edge cases, plus a labeled test set you can review by hand. A smaller, well-labeled set is easier to trust than a huge unlabeled pile. Start with the smallest set that lets you measure correctness on real cases.

Who should own AI data readiness?

A business owner and a technical owner should share it. The business owner defines correct output and acceptable gaps. The technical owner handles access, pipelines, and test data. One person can hold both roles, but both perspectives need a seat at the review.

Monthly plan

Want help with this? AI readiness assessment

$5,500/month

2 weeks · Starts within 10 business days

Two weeks to rank your AI and agent ideas by value and risk, check your data, and plan a first project with a cost model.

Renews monthly. Cancel any time.
What is included

Related guides

More ai and agents guides   AI and machine learning

Smallest way to start

Get a senior architect on your problem this week

Ninety minutes a month on the cloud, AI or security question that is blocking you, a pre-read of your material and a written summary within two business days.

$450/month

Renews monthly. Cancel any time.