Your AI will give you a confident answer about your spend. It may also be wrong. Run three questions on your ERP to find out where your data would fail.
Step 1: Do the free controls first
Most of the work is enforcing rules. Consultant Tommaso Maria Ricci lists six controls in a recent piece on the procure-to-pay process, and here’s the order he would put them in:
- Separate supplier creation from payment release.
- Verify bank detail changes through an independent channel.
- Detect duplicate payments before release also across similar supplier names.
- Suspend dormant suppliers automatically.
- Sample auto-approved tolerance invoices each quarter.
- Add a second review for first payments to new suppliers.
Ricci also notes that five of the six controls need no new software. So if you haven’t put the basic controls in place, this is a great place to start, because they cost almost nothing.
But even free controls take time. Someone still needs to step in to run the duplicate check and the dormant-supplier review. Plus, there’s the running of the tolerance sample on each month’s transactions.
Step 2: Count what sampling leaves unchecked
We recommend you start with two numbers:
- How many invoices do you receive each month?
- What share of those invoices are being checked each month against all your compliance rules?
Sampling is a sensible answer to the problem of limited hours, and the cost is what it leaves out. In a 10% sample, nine of every ten invoices go through unseen. Same goes for your compliance rules. Most companies have somewhere between 50 and 100 checks to make consistently across the Source to Pay process, but some never make it from the powerpoint into operations - they are simply too time-consuming for the staff.
The Association of Certified Fraud Examiners’ 2026 report, Occupational Fraud 2026: A Report to the Nations, found that across 2,402 occupational fraud cases in 143 different countries, the median loss was $104,000. Worse yet, the typical scheme ran 12 months before anyone found it.
The report also found that losses were 55% lower where management review was implemented. They were 53% lower where an organization used proactive data monitoring or analysis.
These are associations, not proof that control caused the drop, but the lesson is clear: the sooner and more widely you look, the less goes unnoticed.
Sampling can also cause you to miss potential patterns in the data. A duplicate payment may live in two invoices, and a sample only reads invoices one at a time. For instance, imagine that an $18,250 invoice is paid on the 3rd by one site under one supplier record and again on the 19th by another site under a second record for the same supplier but with a slightly different supplier name.
If each invoice has a 10% chance of being pulled, the chance that both payments land in the sample is about 1 in 100, and even if one does, nothing points you to the duplicate. The company pays $18,250 twice, and the money may never end up coming back.
Granted, not every loss is fraud. Duplicate payments and price errors can happen to any organization at any time, and they often hide in the same unchecked invoices. That’s part of the human cost in this gap.
Step 3: Run the review-hours math
Assume that you have 3,000 invoices a month. Checking one against a purchase order, contract price, and vendor history may take four minutes. Then consider that investigating a flagged invoice takes around 10 minutes.
Now run the numbers along with a median wage for bookkeeping, accounting, and auditing clerks: $50,670 a year in May 2025, or $24.36 an hour over 2,080 working hours. Below are some examples built on invented inputs, so treat the results as an illustration:
|
Approach |
Invoices checked |
Hours a month |
Base wage cost a month |
Left unchecked |
|---|---|---|---|---|
|
Full manual review |
3,000 |
200 |
$4,872 |
None |
|
10% sample |
300 |
20 |
$487 |
2,700 |
|
Exceptions only (3% flagged, invented rate) |
All, by software |
15 |
$365 |
None |
A full review takes 200 hours a month (2,400 hours per year). That’s a little more than one full-time person (2,080 hours) for a single control. And once you compare both, you’ll see why many hire someone as the default fix.
The sample takes 20 hours and leaves ~2,700 invoices unchecked and unseen. In the third row of the table, you’ll see that software checks all 3,000 and sends only the 90 that it flagged to a human for manual review: 15 hours.
The soft number is the 3% flag rate. At 10%, that row becomes 50 hours, which is still a quarter of a full review.
You can get a sense of the true hours for your organization by timing five real checks and running two equations:
- Full review hours = invoices × minutes to check ÷ 60
- Monitoring hours = invoices × flag rate × minutes to investigate ÷ 60
Step 4: Let software check every line
At this point, the work shifts from hunting for problems to working with a list of them. CostBits’ Live Issue Detection tool continuously mines your transaction data for any cases where employees or suppliers don’t follow the set business rules.
Results from that check land in one live list that you can sort and filter, then delegate to the right team member to resolve. Each issue links back to your finance system.
For Business Central and Dynamics 365 Finance & Operations, that’s done in a click. For other ERPs, you can copy a reference (such as an invoice number) and find it in Infor M3, SAP, or any other system you currently use. CostBits works with any ERP.
This changes how your controllers spend their time. Human involvement is less line-by-line analysis, more interpreting what the system flags and how to resolve it.
Step 5: Check that your data and rules are ready
Three things will decide whether monitoring works for you as intended. Two of them, resolve and trace, come from the tests that we discuss in our recent guide on AI and spend data. The third is the rules you give the software itself to run.
1. Resolve: clean supplier records
If you find that one supplier exists as two records, a duplicate payment between them may look like two different payments to two different companies. You only discover this once records are cleaned to resolve to a single supplier.
CostBits cleans supplier records without writing them back to your ERP, so you don’t have to wait on a master data project to move ahead.
2. Written rules
Software can only enforce the rules you give it. That means if your approval limits live in someone’s head, then it’s time to get it all down on paper officially. Set the limits by tier, starting with the free controls we mentioned in Step 1. Remember rules only work when enforced - nothing was ever saved by just showing good intent on a PowerPoint.
3. Trace
Ask any vendor for two things: the checks that it runs on your live data and a path from a flagged item back to the source lines.
If you find that a particular vendor can’t provide both of those, there’s a high chance that you’ll have to verify the data by hand, and then the hours come back.
When should you still hire a controller?
Remember, monitoring doesn’t close the books or prepare the reports. It can’t decide whether a flagged payment is a mistake or part of a policy exception.
These are all things that still require judgment by a trained professional. When you weigh the cost of hiring a controller against the risks you may run without that judgment, you’ll see that investing in a professional with judgment is vital many times.
Hire when the exceptions list is outgrowing those who are reading it and when the work is increasingly analysis rather than checking. Don’t simply hire to widen a sample.
What should you do this week?
- Write the rules down: list any and all approval limits and tolerance levels for price and quantity. Identify who can create or change a supplier. If you can’t write it down, software can’t enforce it.
- Put the free controls in place: go back to step 1 and consider what free controls you can put in place right now. Separate supplier creation from payment release. Verify bank detail changes through a second channel.
- Run the numbers: Count last month’s invoices and time a few checks of your own. This will give you your own version of the table above, and get you started on comparing the price of hiring vs using a system + human judgment method.
Price the checking before you hire
Ready to compare the options for yourself? The first step is to take your own invoice count and run the two formulas from Step 3 above.
If the results come back showing that a full review goes beyond one controller’s working hours, it may be time to consider automated monitoring as part of your process.
Want to learn more? See how CostBits helps organizations like yours enhance their spend control. See more about tools like Live Issue Detection, and run a free scan on your own data today.
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.