Where most teams are today
Most enterprises have spent the last few years moving data into a lake or warehouse. The idea was to put everything in one place and use it as the foundation for analytics and, eventually, AI.
That part mostly happened. What didn’t change much is the data itself. Duplicate customers, terms that mean different things to different teams, and records nobody owns all came along for the ride.
Analysts still go back to source systems when they need a number they can defend. It’s still common for finance and sales to show up to the same meeting with different revenue figures.
With dashboards, people learned to live with it. They knew which records to ignore and fixed the rest in their heads.
AI agents can’t do that. They act on whatever data they’re given.
Why this is getting more urgent
When a dashboard has a duplicate customer, someone sees a wrong number and usually catches it. When an agent has a duplicate customer, it might send an invoice to an old account or answer a support question using outdated contract terms. By the time anyone notices, the invoice has already gone out.
Adoption is moving fast. Deloitte’s 2026 State of AI in the Enterprise report found that nearly three in four companies expect to use agentic AI at least moderately within two years. Only 21% say they have a mature way to govern those agents (Deloitte).
Gartner has been pointing at data as the bottleneck for a while. It predicts that through 2026, organizations will abandon 60% of AI projects that aren’t supported by AI-ready data. It also found that 63% of organizations either lack the right data management practices for AI or aren’t sure whether they have them (Gartner).
When we talk with data leaders, we keep hearing the same story. A pilot runs on a clean sample of data someone prepared by hand. Then it moves to production, starts reading from the lake everyone actually uses, and stalls.
Five things to fix before agents touch the lake
You don’t need a new platform to start on any of these. Mostly, they come down to deciding where problems get fixed and who’s responsible for fixing them.
1. Clean up your most important data before it reaches the lake
Customers, products, suppliers, and locations cause most of the trouble. If duplicates and conflicts get resolved in the systems where the data is created, the lake receives one version of each record to work with.
- Start with whichever type of data your first AI use case leans on most.
- Count how many duplicate records it has now, so you have a baseline to show progress against.
2. Settle on one definition for each critical term
If sales, finance, and support each have their own idea of an “active customer,” an agent has no way of knowing which one you meant.
A good place to begin is the 10 terms that cause the most debate in your meetings. Write down one definition for each and give it an owner.
3. Give every area of data an owner
Customer data, product data, and supplier data each need someone responsible for them. When customer data belongs to nobody, each team patches it locally and the fixes don’t stick.
- The owner should sit in the business. IT can support, but the business knows what the data should look like.
- Give each owner a simple quality target they can track.
4. Limit what agents can access
An agent should only see the data its job requires. Personal information, financial data, and confidential information each need their own rules.
Before an agent goes live, label the data it will use by how sensitive it is, and keep a log of what each agent accessed and why.
5. Treat cleanup as ongoing work
A one-time cleanup starts to decay almost immediately. New records come in, people change jobs, and systems overwrite each other when they sync.
Add checks where data enters as well as where it’s reported. If the same issues keep showing up quarter after quarter, the problem usually starts at the source.
What we’re seeing from teams that get agents into production
The teams that make it past the pilot stage tend to have a few things in common.
Most of them start small: one use case and one or two types of data, like customers or products, cleaned all the way through. Deloitte’s research found the same thing. The most successful companies begin with lower risk use cases and expand from there.
Monotype is a good example. After an acquisition, its customer data was split across two Salesforce instances, Zuora, Marketo, and its own databases. Its new global data team skipped the multi-year overhaul. It connected the systems that mattered most, had its first three use cases live within weeks, and expanded from there. With that foundation in place, the team is now running multilingual customer feedback through AI workflows (read the Monotype story).
They also fix data where it’s created. Because issues get resolved before data reaches the lake, nobody has to rebuild the same cleanup logic in every report and model.
Governance gets settled during the build. Owners, definitions, and access rules are in place before an agent goes live, so a mistake in production isn’t what starts the conversation.
And they bring the business in early. The people who create and use the data help decide what “clean” means, which makes adoption much easier after launch.
Where to start this week
Pick the AI use case your team is most excited about. Write down the three types of data it depends on most, like customers, products, or suppliers, and find out who owns each one.
If nobody can tell you, start there.
Want this as a two-page brief to share with your team? Download the PDF.
