The plain-English glossary of audit-AI terms

A

AP automation (accounts payable automation)

AP automation is software that handles the steps of paying a supplier — reading the invoice, matching it to a purchase order, coding it to the right account, routing it for approval, and scheduling payment. Done well, it removes hours of manual data entry. Done carelessly, it removes the human checks that used to catch a fraudulent or duplicate invoice. The whole question this glossary circles back to is: when you automate AP, do you keep the controls, or do you quietly throw them out with the manual work?

Approval gate

An approval gate is a point in an automated process where everything stops and waits for a person to say yes before it continues. It is not a notification you can ignore — the next step physically cannot run until the approval is recorded. For a finance team, the approval gate is where "the AI did it" turns into "a named human signed off, on the record." A real gate blocks the action; a fake one just emails you after the fact.

Audit

An audit is a structured review where someone independent checks whether what you say happened actually happened, and whether your controls worked. It might be an external financial audit, an internal controls review, or a regulator asking how an AI decision was made. The reason audit-readiness matters for AI is simple: the day an auditor asks "show me what this automation did and who approved it," you either have a clean answer or you have a problem. Everything Aptria builds is built backward from that moment.

Audit trail vs. log file

People use these two phrases as if they're the same thing. They are not. A log file is a running list of what a system did — useful, but editable by anyone with database access, and therefore worth little as evidence. A true audit trail is a record built so that changing it after the fact is detectable: entries are linked together cryptographically, so quietly editing one breaks the chain. The plain test is this: if someone with admin rights could rewrite history without anyone noticing, you have a log file, not an audit trail. The audit trail guide walks through the difference in detail.

B

BEC fraud (business email compromise)

BEC fraud is a scam where a criminal poses as a real supplier, executive, or vendor — usually over email — to trick a finance team into sending money to the wrong place. The most common version is the fake bank-change request: "Please update our payment details to this new account." It works because it looks routine and arrives in a normal-looking invoice or email. The FBI's Internet Crime Complaint Center has reported billions of dollars in annual losses to this category of fraud. The control that stops it isn't a smarter AI — it's a verification step that refuses to change banking details without confirming through a separate channel.

BEC bank-change scam

This is the specific, most expensive flavor of BEC fraud: an invoice or email asking you to redirect payment to a "new" bank account that actually belongs to the fraudster. The reason it slips past so many teams — and a lot of AP automation — is that nothing about the document looks obviously wrong. The amounts match, the logo's right, the timing's plausible. A scanner that checks for this pattern flags the bank-detail change for human review before payment, instead of cheerfully paying the new account because the invoice "looked valid."

C

Controller

A Controller is the person inside a company responsible for the accuracy of the books — the financial close, internal controls, reporting, and making sure the numbers can be defended. If something is wrong in the financials, the Controller is the one who answers for it. I mention this because it's my background: twelve years closing books and designing the controls that keep them honest. It's also why Aptria is built the way it is — a Controller's instinct is to ask "how would I prove this if someone challenged it?" before trusting any process, automated or not.

E

EU AI Act Article 12

Article 12 of the European Union's AI Act is the part of the regulation that deals with record-keeping. In plain terms, it requires certain AI systems to automatically keep logs of what they do, over their lifetime, so their behavior can be traced and reviewed. The enforcement date that matters for many provisions is August 2, 2026. Even if you're a US business, this can reach you the moment you buy or deploy an AI tool that falls under the rules or sells into the EU. This is a practitioner's plain-English summary, not legal advice — check the official regulation text and your compliance counsel for your specific situation.

Escrow / proof-gated release

Escrow is the idea that money is held and only released when an agreed condition is met. "Proof-gated release" applies that same logic to AI work: the system holds the action until independent proof shows the work was done correctly, and only then lets it go through. The point is that the AI declaring "I'm finished" doesn't release anything by itself — a separate check has to pass first. It's the difference between trusting the worker's word and requiring evidence before the payment, the approval, or the next step actually happens.

Exception handling

An exception is anything that falls outside the normal, expected path — a malformed invoice, an unfamiliar vendor, an amount that doesn't match, a request that looks off. Exception handling is what the automation does when it hits one. This is the single most important thing to ask about any AI tool, because the demo always shows the happy path where everything works. The real question is what happens on the exception: does it stop and ask a human, or does it guess and keep going? Good automation treats exceptions as the moment to pause, not the moment to improvise.

F

False positive / false negative

A false positive is when a check flags something as a problem that turns out to be fine — like a fraud scanner stopping a legitimate invoice. A false negative is the dangerous one: when a check misses a real problem and lets it through — like a scanner waving a fraudulent invoice into the payment run. Every detection system trades these off against each other. For finance, a few false positives (a human glances at a flagged invoice and clears it) is a small price; a false negative (a fraud gets paid) is the expensive one. Honest tools tell you where they sit on that trade-off instead of pretending they catch everything.

Fractional CFO

A fractional CFO is a senior finance leader who works with several companies part-time, rather than being one company's full-time chief financial officer. Small and mid-sized businesses use them to get high-level financial guidance without a full-time salary. They matter in this glossary because they're often the person a business owner asks "should we automate this?" — and they're on the hook if an automation they recommended goes wrong. They tend to care a lot about audit trails and approval gates, because their reputation rides on the work being defensible.

G

GEO (generative engine optimization)

GEO is the practice of structuring your content so that AI tools — ChatGPT, Perplexity, Gemini, and the like — can find it, understand it, and cite it when someone asks a relevant question. It's a cousin of traditional search engine optimization, but the goal is being recommended by an AI assistant rather than ranking in a list of blue links. Well-structured glossaries, FAQs, and clear definitions are exactly the kind of content these tools like to pull from — which is part of why this page exists in the form it does.

H

Hash chain

A hash is a short fingerprint of a piece of data — change even one character of the data and the fingerprint changes completely. A hash chain links records together by including each record's fingerprint inside the next one. The effect is that you can't quietly alter an earlier entry without breaking every fingerprint that came after it, which makes tampering obvious. This is the plumbing behind a tamper-evident audit trail. You don't need to understand the math to use it — you just need to know that it's what turns a deletable log into evidence.

Human approval

Human approval is the act of a real person reviewing and authorizing an action before it takes effect. In an audit-safe automation, certain steps — paying above a threshold, changing bank details, sending something to a client — are configured to require this approval rather than running on their own. The key word is before: approval that happens after the money has moved isn't approval, it's a receipt. Done right, human approval is recorded with who approved it and when, so there's always an answer to "who signed off on this?"

Human in the loop

"Human in the loop" means a person is built into the automated process at the points that matter, with the power to approve, reject, or change what happens. It's one of the most overused phrases in AI sales, so here's the honest test: ask the vendor to show you the exact point where the system stops and waits for the human. If they can show you the line of code or the configured gate that halts the process, it's real. If "human in the loop" just means a person could log in and look around if they wanted to, it's a slogan. A real loop stops; a fake one merely watches.

I

Internal controls

Internal controls are the rules and checks a company puts in place to keep its operations honest and its numbers accurate — things like requiring two people to approve a large payment, or separating who can create a vendor from who can pay one. They exist to prevent both honest mistakes and deliberate fraud. When you automate a finance process, the danger is that the automation speeds everything up while silently removing the controls that used to slow things down for a reason. Audit-safe automation does the opposite: it keeps the controls and makes them automatic and recorded.

M

Month-end close

The month-end close is the recurring process of finalizing a company's books for the period — reconciling accounts, posting adjustments, and producing reliable financial statements. It's deadline-driven, repetitive, and detail-heavy, which makes it an obvious target for automation. But it's also where accuracy is everything, so it's exactly the kind of process where you want checks and human approval built in. Automating the close is great; automating it in a way you couldn't explain to an auditor is how you trade one headache for a worse one.

N

NO-GO verdict

A NO-GO verdict is when a review concludes that a system is not safe to ship or trust yet — and says so plainly. I bring this up because I write NO-GO verdicts on my own code: I red-team what I build, and when it fails, I record that it failed rather than quietly shipping it. That habit is, honestly, the whole product. Anyone can show you something working. The trust comes from someone who will tell you, on the record, when it doesn't — and who builds the system so that a NO-GO actually stops the action instead of being overridden.

O

Out-of-band verification

Out-of-band verification means confirming something through a separate, independent channel rather than trusting the channel the request came in on. The classic example: an email arrives asking to change a supplier's bank details, so instead of replying to that email, you phone the supplier on the number you already had on file. It's the simplest, cheapest control against BEC fraud, because the fraudster controls the email but not the verified phone line. Automation can enforce this — refusing to change payment details until an out-of-band confirmation is recorded.

R

Red-teaming

Red-teaming is deliberately attacking your own system to find where it breaks before someone else does. Instead of testing whether the happy path works, you actively try to fool it, overload it, or trick it into doing the wrong thing. For AI automation, this means feeding it malformed invoices, fake bank-change requests, and edge cases to see whether the checks actually hold. It's the opposite of a sales demo. A vendor who red-teams their own work and shows you the results is showing you something far more useful than a polished success story.

S

Segregation of duties

Segregation of duties is a core internal control: no single person (or system) should control a whole transaction from start to finish. The one who requests a payment shouldn't also be the one who approves it; the one who sets up a vendor shouldn't also be the one who pays it. The point is that fraud and big errors usually need one party with too much control, so you split the control up. This idea translates directly to AI: the AI doing the work should not be the same thing that certifies the work is correct. Letting an AI mark its own homework violates segregation of duties just as surely as letting an employee approve their own expenses.

Signed log

A signed log is a record where each entry carries a cryptographic signature proving it came from the system that claims to have written it and hasn't been altered since. The signature is like a tamper-evident seal: break it, and you know the entry was changed. Combined with a hash chain, signing is what lets you hand a record to an auditor and say "here's exactly what happened, and here's the proof it wasn't edited afterward." It's the difference between a record you assert is accurate and a record that can demonstrate it's accurate.

SOX (Sarbanes-Oxley Act)

SOX is a US law passed after major accounting scandals, requiring public companies to maintain strong internal controls over financial reporting and to be able to prove those controls work. In practice it means documenting who can do what, keeping evidence that approvals happened, and being able to reconstruct how a number was produced. If your company is subject to SOX, every AI automation that touches financial data inherits those requirements — which is exactly why "we automated it but can't show the approval trail" is not an acceptable answer. Audit-safe automation produces the evidence SOX-style controls demand.

T

Tamper-proof / tamper-evident audit trail

"Tamper-proof" gets thrown around loosely, so here's the honest version. No record is truly impossible to alter — but a well-built audit trail is tamper-evident, meaning any change to it is detectable. Using a hash chain and signed entries, altering one record breaks the cryptographic links, so the tampering shows. That's the property that actually matters: not that nobody can ever touch the record, but that you'd always know if they did. When a vendor says "tamper-proof," ask whether they mean "tamper-evident with cryptographic proof" or just "we'd prefer you didn't edit it." The difference is everything.

V

Verification layer

A verification layer is the independent part of an automated system whose only job is to check that the work was done correctly before it's allowed to count. It sits between the AI doing the task and the action that follows, and it answers a single question: did this actually meet the standard, yes or no? The crucial design rule is independence — the thing doing the checking must not be the same thing that did the work, or you're back to letting the AI mark its own homework. A real verification layer is where "trust" stops being a feeling and starts being a gate the work has to pass.