Can AI detect revenue leakage?

Yes, AI can detect revenue leakage, but only in data it can reach and only where a comparison can be made. It works best on leaks that arise between systems: a won deal without an invoice, a contract whose indexation was never applied, a customer who uses more than they pay for. AI finds those discrepancies faster and more widely than a person with Excel. It does not decide on its own whether something really is a leak: that takes context which often only your people have.

What does "AI" actually mean here?

In leak detection, the label AI covers three different techniques, each doing something different:

1. Rules and comparisons. "Every won deal must have an invoice within thirty days." Strictly speaking this is not AI, but it is the foundation of every detection. It finds exactly what you have defined, nothing more.
2. Statistical models and machine learning. These learn what is normal, per customer, per product, per month, and flag what deviates from it. This is called anomaly detection. It finds things you had not thought to write a rule for.
3. Language models. These can read unstructured text: contracts, quotes, emails, notes. They extract agreements, such as an indexation clause or a discount with an end date, and turn them into something you can compare with the invoice.

A good detection system uses all three. Rules for what you know for certain. Models for what you had not foreseen. Language models to extract the agreements from documents that would otherwise never make it into a comparison.

How does AI detect a revenue leak?

Step by step, it looks like this:

1. Read. The system pulls data from the connected sources: deals from the CRM, orders and invoices from the ERP or billing system, contracts from a document folder, tickets from support.
2. Match. It links records from different systems. This deal in HubSpot belongs to this customer in Exact or Xero, and to this contract. That is often the hardest part, because systems rarely share the same keys: names are spelt slightly differently, customer numbers are missing from the CRM.
3. Compare. It sets what should be there against what is there. Agreed price against invoiced price. Licences sold against licences used. Expected volume against actual volume.
4. Price. It calculates what a discrepancy costs in euros, per month or per year.
5. Assess. It estimates how certain it is that this is a real leak, and not a data error or a deliberate arrangement.
6. Propose. It prepares a finding with evidence, amount, confidence and a proposed next step.

The first two steps largely decide whether the rest works. A model that does not know "Baker Ltd" and "Baker Building Services Limited" are the same customer sees two half-invoiced customers instead of one correctly invoiced one. More on the mechanics in how to detect revenue leakage automatically.

What is AI good at?

• Looking at everything at once. A person checks twenty customers in an afternoon. A system compares every customer, every invoice and every deal every night.
• Spotting patterns nobody defined. A customer who drops from five product groups to two, a region where discounts slowly creep up, an account manager whose deals more often end up without an invoice.
• Extracting agreements from text. A language model reads a hundred contracts and produces a list of indexation clauses, end dates and discount arrangements. Done by hand, that work was often left undone because it took too long.
• Staying consistent. It does not forget in December and does not skip a customer because they are awkward.

Where does AI fall short?

• It only sees what exists. Work that was never recorded, extra work agreed verbally, a discount promised over the phone: there is no data for any of it, so no detection either.
• It does not know the reason. A customer on a 30 percent discount may be a leak, or a strategic account for which the board deliberately granted that discount. The model sees the deviation, not the decision.
• It can be wrong. A language model can misread a clause. A statistical model can mistake a seasonal dip for a customer winding down. That is why every finding should carry a confidence level, and why a person must check the important ones. See why AI output must be checked.
• It is only as good as its connections. With billing alone, you find price discrepancies. With CRM and billing, you also find won deals without an invoice. Add contracts and you also find missed indexation.

Worked example

Worked example: suppose a financial controller spends one day a month manually comparing won deals with invoices, and gets through 40 deals. You win 150 deals a month, so they see just over a quarter. If 2 percent of deals go wrong somewhere and an average deal is worth EUR 12,000:

• Errors per month: 3 deals, together EUR 36,000.
• What the controller finds on average: just over a quarter of that, around EUR 10,000.
• What remains: around EUR 26,000 a month.

A system that compares all 150 deals can in principle find all three. It may also produce false alerts that the controller has to dismiss. Their day shifts from searching to assessing. The percentages here are assumptions for the example, not averages.

What should you check in an AI finding?

Four questions to ask of every finding:

1. What is the evidence? Which records from which systems underpin it? A finding without traceable evidence is a suspicion.
2. How confident is the system? And what does that number mean: how often was a finding at that confidence level correct in the past?
3. What does it cost? An amount with the calculation behind it, so you can check it yourself.
4. Who decides? A proposal to correct an invoice or call a customer should pass a person before it is carried out. The principle is called human-in-the-loop.

Which data does the AI get to see?

To find leaks, a system needs access to your customer and invoice data. That raises fair questions about privacy. Names, email addresses and other personal data are usually not needed for leak detection: it is about amounts, dates, quantities and agreements. A good design removes them before a language model sees the data. How that works is explained in how to protect personal data in AI systems. RiOS is built on that rule: the security page explains how data is tokenised before an AI model sees it, and that a finding which does not reach the confidence threshold is not published.

Do you need AI?

Not to get started. Many leaks can be found with a good export and a few comparisons in Excel, as described in the wider approach to finding revenue leakage. AI becomes valuable when the volume grows too large for manual work, when agreements sit in text rather than in fields, or when you want the check to run continuously rather than once a year.

Frequently asked questions

Can AI find revenue leakage without connecting to my systems?

No. Without data there is no detection. A structured review with people who know your processes can, however, expose many leaks without any connection. That is a different kind of work with a different result: a snapshot rather than continuous monitoring.

How many false alerts does such a system produce?

That depends on data quality and thresholds. More at the start, because the system does not yet know which deviations are deliberate arrangements. Once those arrangements are recorded, the number falls.

Can AI also close the leak itself?

It can propose and prepare a fix, such as a correction to a subscription or a task for an account manager. Whether it also carries that out is a choice. For anything that affects a customer, approval by a person is sensible.

Is a language model like ChatGPT enough?

A standalone language model can read a contract or analyse an export you paste into it. It cannot continuously compare your systems, and you need to think carefully about which customer data you paste into it. For one-off analyses it is a useful tool; for structural detection it is not.