Is Your Data Ready for AI? The Audit to Run Before Any Build
More clinic AI projects stall on data than on models, but 'get our data ready' does not mean what most practices fear. A six-question scorecard tells you whether the records behind your first workflow are ready, and what to fix if they are not.
In this article
Ask a room of practice owners why their AI project stalled and almost nobody says “the model was not smart enough.” The honest answers sound like: nobody could say which version of the protocol was current, the scanned letters were three years behind, and nobody was sure who was allowed to see what. The model showed up ready. The records did not.
That failure has produced an equally expensive overreaction: practices that will not start until they have “fixed the data,” a project with no finish line and no budget that ever closes it. Both mistakes come from the same misunderstanding of what ready actually means.
”Ready” is scoped to a workflow, not a company
The question is never “are our records ready for AI.” It is “are the records behind this workflow ready,” and that changes everything about the size of the job.
An assistant that answers front-desk questions from your policies needs those policies: current, deduplicated, and permissioned. It does not need a decade of scanned letters indexed, your old practice management system migrated, or your billing history reconciled. Those might matter for workflow five. They are not the price of workflow one.
This is the same discipline as picking your first workflow in the first place: narrow wins. A scoped readiness job takes weeks and ends. A company-wide one takes years and does not.
What “ready” actually means
Four properties do almost all of the work. For the one workflow you have in mind:
- It exists digitally, and it is findable. Not “somewhere in the shared drive.” A named set of sources: these folders, this table, this system. If finding the data is itself a research project, that is the first fix.
- It is fresh enough to trust. An assistant grounded in your 2023 referral protocol will confidently quote the 2023 pathway. Staleness does not make AI fail loudly, it makes it wrong quietly, which in a clinic is considerably worse.
- You know who is allowed to see it. If the source records carry access rules, and personal health information always does, the assistant needs to respect them from day one and log every access. Untangling permissions after go-live is far more painful than mapping them before, and it is the part a privacy assessment will ask about.
- You have examples of right answers. A handful of real questions with answers you know are correct. This is your ground truth, and it is what turns “seems to work” into a quality number you can actually watch.
Notice what is not on the list: perfect structure. Modern AI reads unstructured records directly, the letters and PDFs and free-text notes where most clinic knowledge actually lives. Messy format is fine. Contradictory copies are not: nine versions of the same protocol with no marked winner will make any assistant contradict itself, and no model upgrade fixes that.
AI does not need your records to be tidy. It needs them to be true, and it needs to know which copy to believe.
Score the workflow you have in mind, honestly:
Scorecard
How ready is the data behind your first AI workflow?
Answer for one specific workflow, not for the practice.
0 of 6 checked
What the audit buys you
One workflow's data, audited
Weeks of scoped work, not a company-wide cleanup
An assistant that tells the truth about this corner of the business
The fixes are smaller than they look
Each unchecked box has a boringly practical fix, and none of them is a platform project.
Data trapped on paper or in one person’s head gets captured by the workflow itself: a one-hour interview written up, a scanning pass, a template people fill from now on. Scattered sources get a named home: not moved, just listed, “these three folders and this table are the sources.” Duplicate versions get a decision: someone with authority spends an afternoon marking winners and archiving the rest. Staleness gets an owner: updating the source becomes part of a named person’s job, because data with no owner rots back within a quarter. And missing ground truth gets an hour of honesty: ten real questions, ten right answers, written down.
That list is the real “data preparation for AI.” Not a lake, not a warehouse migration, not a year of cleanup. Days to weeks of focused work on one corner, and every later workflow inherits part of it.
Where this goes next
Once one workflow’s data is ready and an assistant is answering from it with sources attached, something useful happens culturally: teams see what ready data buys, and the next corner gets cleaned because people want the next assistant. Readiness stops being an initiative and becomes a habit, funded by results instead of patience.
That is the honest sequencing: not “fix the data, then do AI,” and not “ignore the data, the model will cope.” Fix the corner you are building on. Build. Repeat.
Not sure whether your data can support the workflow you have in mind? That is a one-conversation diagnosis. Book a free consult and we will run this exact audit against your first workflow together, and map the shortest path from where your data is to where it needs to be.
Frequently asked questions
How do I know if my data is ready for AI?
Check four things for the one workflow you want to improve: the records exist digitally and are findable, they are fresh enough to trust, you know who is allowed to see them, and you have real examples of what a right answer looks like. If those hold for that single workflow, you are ready to start, whatever state the rest of the practice's records are in.
Do we need to clean up all our data before starting with AI?
No, and trying to is the most expensive way to fail. Readiness is scoped per workflow, not per practice. An assistant that answers administrative questions needs your policies in order, not a decade of scanned charts migrated. Clean the corner you are building on, go live, and let each project pay for the next corner.
What kind of data does AI actually need?
Less than most practices assume, and messier than most practices assume is allowed. Modern AI reads unstructured records, letters, PDFs, free-text notes, directly. What it cannot survive is information that is stale, contradictory across copies, or so scattered that nobody can say which version is true, which is the normal state of a clinic protocol folder.
What is the most common data problem that blocks AI projects?
Duplicates and staleness, not format. Nine versions of the same protocol with no marked winner will make any assistant contradict itself, and in a clinic contradicting itself about a protocol is not a minor annoyance. Picking the authoritative source for one workflow is usually days of work, and it is the highest-leverage cleanup available.
Found this useful?
Share this with your network on LinkedIn, it helps more than you think.
Enjoyed this read? Get the next one in your inbox.
When there is something worth your time, you will be first to know. No spam, unsubscribe anytime.
Keep reading
When Do We Stop? Writing the Kill Criteria Before You Write the Cheque
A clinic pilot with no stated ending does not end, it fades, and keeps billing while it fades. Deciding the number, the date, and who calls it costs nothing on day one and is nearly impossible to introduce in month nine.
Read article
Where to Start With AI in a Clinic: How to Pick Your First Workflow
The hardest part of AI in a practice is not the model, it is choosing where to begin. A simple way to pick a first workflow that is high-value, low-risk, and proves itself fast, and the three candidates a clinic should never start with.
Read article
Your Patient Does Not Experience Your Org Chart
The front desk, the clinician and the biller can each hit their target while the patient has a terrible eleven days. You measure the boxes and they experience the lines between them, and buying clinic AI one department at a time usually makes those lines worse.
Read articleHave software that should be smarter?
Let’s map a free AI-transformation roadmap for your product.