Imagine that you’re wrapping up your quarter, and at the quarterly planning meeting, somebody on the board asks you if you’ve ever considered using AI to analyze and optimize your spend data.
You haven’t yet, so you try it out. Your AI copilot takes your prompt: How much do we spend with our largest supplier? The answer appears in seconds, to the dollar, with what seems like total confidence, it even looks well structured.
Then the buyer at your second plant looks at the number AI came up with and says, “Wait, that’s not right. We alone spend much more than that.”
You try to find the right number, but nobody can find it. Someone on your team opens a spreadsheet, and another pulls vendor lists. And by the end of the day, you have a brand-new number, but there's a problem: the first number the AI gave you has already been repeated in a meeting.
Did the AI lie? Not necessarily. It simply worked with the data it had at its disposal. And that’s why knowing the full story of AI and spend data is becoming increasingly important. Because if AI doesn’t have the data it needs, and the data is not in a clean condition, you may create bigger problems faster than you anticipate.
The simple answer is that it reads your records as they are written. And if those records aren’t matching up, then it will give you what it has.
It follows logically. If the same exact part is described in three different ways, the AI will report three parts. Have a single supplier showing up under three names? It’s likely that the AI is going to report three suppliers.
The sneaky part is that nobody entered anything wrong. Each site just entered what made sense locally. Then, when prompted, the AI returned the sum of those decisions.
Now put this in a real moment. Use a report built on flawed data, and you may understate the true spend with a supplier. That could cause you to walk into a negotiation thinking you have less volume behind you than you really do.
When the report shows up, nothing is flagged. From the AI’s point of view, all of the numbers add up as they should.
A wrong answer can read and feel right when delivered confidently by AI. There are two issues working here:
Issue 1: AI is always making a best guess.
AI works with the data it has available and its programming. It’s not an exact science, it's pure math, and a guess that’s made on bad data will land much further off than the same guess made on good data.
Issue 2: The answers that AI gives can be convincing.
AI plays your own data back to you. That means you’ll often see data that seems familiar to you, even if it’s not quite right. Over time, you can build trust in the answers AI provides without having to double-check its work.
Part of that is because it’s so convincing. A machine wrote and delivered the answer, so it is well-structured and detailed. That builds trust faster than you might expect!
Here’s a case in point: a procurement specialist we work with built an engine that turns spend data into UNSPSC categories.
We tested the system more than 20 times, and each time the results came back seemingly perfect. It was well structured and included a lot of information.
But when we went back and checked the answers against the evidence, we found that every single result was wrong. Some of the codes weren’t even legitimate UNSPSC codes.
We ended up having to tell the specialist about the issue, because he was already selling the engine to clients.
A confident and well-formatted wrong answer is the hardest to catch, which is why the three questions below matter so much.
When you look at the statistics, you’ll find that you’re in good company. Nearly every company knows that they should be implementing AI in their operations and reporting, but are risking accuracy for the sake of speed.
Gartner predicted that by the end of 2026, organizations will abandon 60% of AI projects that are not supported by AI-ready data. And 63% either lack or are unsure that they have the right data management practices to work with AI.
A Deloitte study surveyed 100 chief procurement officers and found that while 92% planned to invest in generative AI, only 37% were truly piloting or deploying it at all. The problem seemed to revolve around concerns about data quality, with limited governance and disparate systems contributing to that lack of faith.
But the potential of AI to help companies accelerate their operations is still there.
AI-ready spend data is data that an AI can resolve, categorize, and trace. One of the best ways to look at it is with three questions. Think of it as questions you’d expect a new analyst to answer on day one:
1. Is this one supplier or five?
With AI-ready spend data, you want data to resolve. That means every real-world supplier maps to a single record. In any ERP system, that takes deliberate work.
Microsoft defines a company as the container for the business data of a single legal entity or business unit. It also describes synchronizing master data between companies as something you can set up so you don’t add vendors and items manually in each one.
If nobody sets that up, then each entity is likely to keep its own supplier list. A group with five plants can hold the same steel supplier five times, and an AI tasked with calculating total spend with that supplier will return five partial answers.
Business Central is the example here, but this problem shows up in nearly every ERP, and CostBits works with all of them. In Business Central specifically, you can set up CostBits yourself, because the platform is open and simple. Dynamics 365 Finance & Operations is nearly the same, but the setup requires a bit more customization but you will be live in less than a week.
For every other ERP, we provide our API setup and work with you to build a live data extract that fits our well documented data model, which is part of what makes CostBits ERP-agnostic.
You can try this right now: ask Accounts Payable to report how many suppliers you have. Then ask the same for each plant. Better yet ask the CFO how many suppliers are being created every month and then ask the AP team the same questions - they will converge as you have expected.
If the numbers don’t add up, you can already tell where the first test will take you.
2. Can you trust the category?
AI needs to resolve spend data and categorize it. That means every transaction line and not the supplier, sits in one scheme so that spend in a particular category is a single number, no matter how you ask about the data.
Using free-text item descriptions or not leveraging the built-in catalogue in the ERP system makes this type of categorization incredibly challenging, typically because each site describes an item differently. One way to fix this is to map to a consistent scheme.
A note on categorization: CostBits works with your own category tree or the one our other customers already use. A standard like UNSPSC is great for physical goods, but it can be broader than you need, making it hard to use in day-to-day operations. It also covers services rather poorly and has no spot for items like “finance cost”.
And the payoff here can show up quickly. When Dania Plast started clicking through its spend by category, it found that cleaning costs were much higher than expected. Dania Plast COO Mark Wolff called the result “proof that when data is cleaned and then presented in a clear and intuitive way, you can quickly turn insights into real value.”
And he’s right. A supplier conversation quickly followed the data reveal, leading to a renegotiated agreement and 100,000 DKK found within about a week. By now, Dania Plast is past 500,000 DKK in recurring savings.
The money always lived in the data, and the categorization based on the cleaned data helped make it visible. Now imagine being able to run that spend data search using AI to get accurate recommendations that saves you money.
You can read the entire story of Dania Plast on the CostBits blog. Learn how CostBits and the Clean Data module can help you consolidate supplier and item records, and map spend to categories for your organization.
3. Can you check the answer?
The third part is trace. This means every answer links to the ERP lines behind it, and spend links to the contract that governs it.
It’s important to note that AI is not foolproof here. Microsoft’s own documentation for Chat with Copilot in Business Central says that AI-generated content "might not be accurate," and it tells users to have a process for reviewing and verifying responses before making decisions based on the data.
AI isn’t removing the checking so much as it’s moving it. When you need that checking to still exist in the process, a tool like CostBits Live Issue Detection lets you jump from a flagged issue straight back into your finance system to check.
In Business Central and Dynamics 365 Finance & Operations, you can do this with a single click. For other ERPs, you copy a reference such as an invoice number and find it in Infor M3, SAP, or whatever you use.
And the Contract Repository ties each contract to a cleaned supplier record, resulting in live contract spend that no one else offers today.
This means you can see how much spend sits under a contract (and how much doesn’t).
There's a risk that referring to "AI-ready" implies the data is clean. But waiting for clean data is a failure mode in itself. The key is to start with the data you currently have and phase the work by use case. Move through by priority and importance.
Clean data does matter because AI’s “guesses” are only as correct as the data underneath them. But master data management (MDM) is a slow process, and few people enjoy it.
CostBits cleans all of your data for you and doesn’t write it back to your ERP (which could result in things breaking if you depend on data that is currently wrong but operationally fine).
A good solution is to export the cleaned data from CostBits and hand it directly to your MDM team. They can then update the ERP through their normal process, keeping humans in the loop. And when they amend the data, CostBits sees the changes and adjusts automatically.
Just remember that data is only one barrier in all of this. The Gartner study of chief procurement officers also found that only 36% said they were “very confident” in their ability to redesign roles and processes around AI moving forward.
So while fixing the data could remove one obstacle, the others are still there. But it’s a step well worth taking.
As you’ve read this blog, you’ve probably run a few tests in your head and have a good idea of which your data may struggle with.
But it’s worth the effort to run them for real on your ERP or AI tool and compare the answers to each one:
How many active suppliers do we currently have, and does the count agree across all of our entities? Get the real total, then each entity’s own count, and add those up. If there’s a gap, then you may have a single provider listed as several records.
What did we spend in one category, by supplier and then by item? Two different totals may mean that categories are assigned in more than one place.
For our largest supplier, what share of last year’s spend sat under a current contract? Can we reach the underlying lines in a single step?gartner If you can’t, then it’s likely that someone is checking by hand.
Make sure to put the same questions to any vendor who offers you AI on your spend data, including CostBits, and ask to see a clear path from a number back to its source lines.
You likely already know which one of those three failed. That’s the one you should fix first.
An AI will only ever be as sharp as the records it can read. It’s better to learn that in an afternoon sifting through your data than in front of the board at a quarterly or annual spend meeting.
Want to see where your own data stands? Start right now at costbits.com. You can also run a free scan on your ERP and AP data to get instant insights with no strings attached.
Don’t let missing data hold your business back. Explore our top 20 free tips for rapid supplier cost reduction or contact CostBits to learn how our platform can unlock the full potential of your invoice data.