"Get your data in order first" is advice that sounds sensible and lands as paralysing, because it implies a project with no visible end. Most business owners hear it and reasonably conclude that AI is a next-year problem.
The advice is right. The scope people imagine is wrong. Data readiness for a specific workflow is a few weeks of focused work. Data readiness for an entire organisation is the multi-year programme people are picturing, and you almost never need it to get started.
Scope it to the workflow, not the business
The question is never "is our data ready." It is "is the data this specific automation touches ready." Those are wildly different amounts of work.
If you are automating after-hours lead qualification, what matters is the lead record: where it lands, what fields come with it, and whether the system can write back. The state of your seven-year-old project archive is irrelevant to that. If you are automating quote generation, what matters is the pricing data and the customer record. Your email marketing lists do not enter into it.
Every hour spent cleaning data that the automation will never read is an hour that bought nothing.
The four questions
Where does the truth live? For each piece of information the workflow needs, one system has to be authoritative. Not "mostly the CRM except for delivery dates." One. If two systems both claim to hold the current customer address, pick one, make the other defer to it, and fix the sync. This is a decision more than a technical task, and it is the one that blocks everything else.
Is it complete enough? Not perfect. Enough. If the automation depends on a field that is populated in sixty percent of records, you need to know what happens for the other forty. Sometimes the answer is backfill it. Often the answer is handle the gap explicitly, route those cases to a human, and move on. The failure mode is not knowing the field was patchy in the first place.
Is it consistent enough? The same customer entered four ways is four customers as far as a system is concerned. Dates in three formats. Statuses that mean different things to different departments. None of this bothers a human reader and all of it breaks automated logic.
Can a system actually reach it? Data locked inside a PDF, a desktop application with no API, or a spreadsheet on someone’s local drive is not accessible data, however clean it is. This is the one that most often changes the plan, because the fix can be anything from a five-minute export to replacing a tool.
What this looks like on a real timeline
For one workflow in a business of twenty to a hundred and fifty people, the honest range is two to six weeks, and most of that is not technical. It is deciding where the truth lives, agreeing a consistent way of recording a handful of fields, and getting the people who enter the data to enter it the same way from now on.
The technical part, the deduplication and the format normalisation and the export pipeline, is usually the fast bit. The agreement is the slow bit, and it is slow in proportion to how many people have to change a habit.
The part that makes it stick
Cleaning the data once and changing nothing else means you are back where you started in a year. Whatever produced the mess is still producing it.
So the cleanup has to come with something that keeps it clean: a required field instead of an optional one, a single intake form instead of three routes in, a validation rule that catches the bad format at entry rather than six months later. Usually the automation you are building becomes part of that, because it is the thing that now depends on the data being right, which finally gives the standard some teeth.
This is the same reason we treat adoption as part of delivery rather than a follow-up. We covered a version of that in CRM automations that survive week two. A system nobody maintains the inputs for has a predictable lifespan.
How to find out where you actually stand
The fastest honest test: pick a question your business should be able to answer instantly. How many active customers do you have. What is the average time from enquiry to quote. How many jobs did you deliver late last quarter.
Now try to answer it from one system. If you can, your data is in better shape than most. If you need to pull two exports and reconcile them in a spreadsheet, you have just found the exact gap that would have broken your first AI build, and you found it for free.
Working out which gaps matter and which ones you can safely ignore is most of what our operations audit does before anyone starts building. If you would rather just talk it through, book a 30-minute call.