Section 1
What AI-ready actually means
The phrase sounds like it requires a warehouse and an engineer. For most companies it means four ordinary properties. Reachable: the system can get to the data without someone exporting a spreadsheet by hand every Monday. Current: it reflects this week, not last quarter. Labelled: fields mean the same thing everywhere, so status does not mean one thing in the CRM and another in the billing system. Permissioned: you can say who and what may see each field, and enforce it. Nothing there is exotic. It is also the exact work most companies have been deferring for years, which is why data readiness is where automation projects stall. Ground yourself first with [What Is AI Automation? A Plain-English Guide for Founders](/blog/what-is-ai-automation-a-plain-english-guide-for-founders).
Section 2
The data you have versus the data that explains decisions
Most companies have plenty of records and almost no outcomes. You know which leads came in. You do not reliably know which ones closed, because nobody updated the field. You have every support ticket and no record of which resolution the customer actually accepted. That distinction decides what is possible. Records let you automate handling: read, classify, route, draft. Outcomes let you automate judgment: predict, prioritise, recommend. If you want the second kind, capturing outcomes is the project, and it starts months before the model does. The cheapest version is to make outcome capture a byproduct of the work rather than an extra step. If closing a deal requires selecting a reason, you will have reasons. If it requires remembering to add one, you will not.
Section 3
Cleanup is a project with a real cost
Data cleanup is unglamorous, it has no demo, and it is usually the largest line in a first automation. Plan it as work rather than discovering it as a delay. Scope it narrowly. You do not need clean data. You need the fifteen fields this workflow reads to be trustworthy. Fix those, leave the rest, and resist the enterprise instinct to fix everything first, which is how a six-week project becomes an eighteen-month one. Then keep it clean by construction: required fields, controlled vocabularies instead of free text where it matters, and one system that is authoritative per field. Cleanup done once is a task. Cleanup done repeatedly is a symptom of a missing rule. Security implications are covered in [Data Security Concerns in AI Automation](/blog/data-security-concerns-in-ai-automation).
Section 4
What you should never send
NIST files AI risk under four headings: trustworthiness, design, evaluation and use. With data the practical version is minimisation. Send the model the least it needs to do the job. Whole-record prompts are the common mistake. A classifier deciding whether a support ticket is urgent does not need the customer's payment details, home address or contract value. Strip fields before they leave your systems rather than trusting a policy document. Then answer three questions in writing: where is this processed, how long is it retained, and is it used to improve someone else's model. If you handle health, financial or employment data, those answers determine whether the project is legal, not just whether it is wise. On the internal communication of these changes, see [Storytelling in the Age of AI and Automation](/blog/storytelling-in-the-age-of-ai-and-automation).
Section 5
Feedback is the data that compounds
The most valuable dataset in an automated workflow is the one it generates: what the system proposed, what the human changed, and why. Almost nobody captures it, because the reviewer just fixes the output and moves on. Capture the correction and you get three things. A record of where the system is weak. A test set built from real failures rather than imagined ones. And, eventually, training material specific to how your business actually decides. One extra field on the review step is enough to start. Six months of corrections is worth more to you than any amount of generic industry data, because it describes your judgment rather than the average of everyone else's.
Section 6
Measuring data quality without a data team
Four checks, monthly, on a spreadsheet. Completeness: what share of records have the fields this workflow depends on. Agreement: when two systems hold the same fact, how often do they match. Freshness: how old is the median record the automation reads. Correction rate: what share of outputs a human had to change, trended over time. The fourth is the one to watch. A correction rate that is stable tells you the system is fit. One that is climbing means the world moved and your data no longer describes it. That is the earliest warning you will get, and it costs nothing to collect.