Who Sets Your Materiality Threshold, and Why It's Breaking Right Now

Ask a controller why a $4,200 variance in office supplies got flagged for investigation last month while a $180,000 swing in COGS did not, and you’ll usually get a shrug. “That’s just the threshold we use.” Ask who set it, and the trail often goes cold at a spreadsheet someone built three finance leaders ago, or a number an auditor mentioned once in passing that got treated as gospel.

Materiality threshold is one of the most consequential numbers in a finance function’s month-end process, and it is also one of the least examined. It decides, line by line, which variances get a human’s attention and which get waved through. Set it too high and a real problem hides in the noise until it’s expensive. Set it too low and the team burns its limited attention budget chasing rounding error, then starts ignoring flags altogether because too many of them turn out to be nothing. Either failure mode teaches people to stop trusting the review process, which is a problem long before you ever point an AI at it.

What a materiality threshold actually is

A materiality threshold is the rule that converts a variance (actual versus budget, actual versus prior period, actual versus forecast) into a decision: investigate, or don’t. It is almost never a single number. A functioning threshold policy usually combines a dollar floor and a percentage trigger, because either one alone breaks in a predictable way.

A pure dollar threshold ($5,000, say) is blind to scale. A $5,000 miss on a $20,000 marketing line is a 25 percent variance: something is probably wrong. A $5,000 miss on a $4 million revenue line is noise. A pure percentage threshold (5 percent, say) has the opposite problem: on a small account, a 5 percent swing might be $40, not worth anyone’s time, while on a large account it might be $200,000, very much worth someone’s time.

The standard fix is a two-sided rule: flag the variance if it exceeds both a percentage threshold and a dollar floor. That combination is what most competent variance review policies converge on, whether or not anyone wrote it down as a formal rule. The trouble is what happens after someone sets it once.

Why the number in use is usually wrong

Materiality thresholds don’t get reviewed on a schedule. They get set once, during a period when the company was a certain size, and then they sit, while revenue, headcount, and account structure move on without them.

What changed	What breaks
Revenue grew 3-5x	A fixed dollar floor that once caught meaningful swings now catches almost everything, or nothing, depending on which side it errs on
New account or cost center added	No threshold exists for it; it defaults to whatever the closest analogous line uses, which may not fit
Entity structure got more complex (new subsidiary, new currency)	A single global threshold stops making sense once entities are wildly different sizes
Team turned over	The rationale for the original number left with the person who set it; the number remains as unexamined convention

The growth-inflection case deserves its own line. A company at $2M ARR and a company at $20M ARR should not be using the same materiality threshold, but very few finance teams schedule a review of the number as part of scaling. It just quietly stops matching the business. The result is a slow drift: either the review queue fills with immaterial noise that erodes trust in the process, or worse, real problems clear the bar without tripping any flag because the bar hasn’t moved in two years while the business has moved a lot.

Materiality is not one number: it should vary by account type

Even within a single period, a flat threshold across every account type misprices risk. Some accounts warrant tighter scrutiny at smaller dollar amounts because errors there are more likely to be systemic rather than timing noise, or because they carry disproportionate downstream consequences.

A workable framework tiers accounts by risk profile rather than applying one rule everywhere:

• Revenue and COGS: tightest threshold. Errors here distort gross margin, a number that flows into almost every other decision and every external report.
• Payroll and headcount-driven costs: tight threshold, because variances are usually structural (a hire, a departure, a comp change) and worth understanding even at modest dollar amounts.
• Accruals and estimates: moderate threshold, but paired with a trend check, since a small variance that repeats every month for six months is a different finding than a one-off.
• Discretionary opex (travel, supplies, software subscriptions): loosest threshold. Timing noise is common and rarely material to the story the numbers tell.

This is a policy decision, not a math problem, which is exactly why it tends to get inherited rather than chosen. Someone has to decide that revenue deserves a 2 percent trigger while discretionary opex can run at 15 percent. That decision belongs to the controller or CFO, made explicitly, not left to whatever a template shipped with.

What this means once AI is doing the first pass

A materiality threshold used to live implicitly in a reviewer’s head: a controller with five years on the account instinctively knows a $3,000 travel variance is nothing and a $3,000 revenue variance in a small business unit deserves a look, even without consulting a written rule. That tacit judgment does not transfer to a system by default.

When an AI assistant is doing first-pass variance review or drafting variance commentary, it needs the threshold made explicit, as data, not as instinct. Feed it a flat percentage and it will either flood the review queue with immaterial flags or silently pass over things a human would have caught, because it has no institutional memory of which numbers matter and which don’t.

This is precisely the kind of policy an AI should retrieve and apply, not infer. The threshold table (by account type, by entity size, reviewed on a set cadence) needs to sit as structured, versioned data in the layer the AI queries, tied to the same reconciled model that produces the actuals and budget figures being compared. That way, when the assistant flags or doesn’t flag a variance, the reasoning is auditable: this account, this period, this threshold, this math, not a black-box judgment call baked into a prompt. It’s the same principle behind KPI variance analysis done right: the AI’s job is to apply a defined policy consistently, not to decide the policy itself.

Setting the threshold: a starting framework

If your team is working from an inherited or undocumented threshold, here’s a defensible starting point to build from, not a universal answer:

1. Set a percentage-and-dollar combined rule per account tier, using the risk-based tiers above as a starting structure.
2. Recalculate the dollar floors against trailing-twelve-month actuals, not against the budget set a year ago, since growth outpaces annual budget revisions.
3. Put a review of the threshold itself on the calendar, ideally tied to your annual planning cycle, so it moves when the business does instead of drifting for years.
4. Document the rationale, not just the number, so the next person to own the process can adjust it intelligently instead of treating it as fixed.
5. Separate “material” from “unusual.” A threshold catches size; it won’t catch a small variance that’s structurally odd (revenue recognized in the wrong entity, for instance). Pair the dollar rule with a handful of pattern checks for things that are wrong regardless of size.

The takeaway

A materiality threshold is a policy choice masquerading as a technical setting, and most finance teams are running on one nobody currently at the company actually chose. That was tolerable when a human reviewer’s judgment quietly filled the gaps. It stops being tolerable the moment an AI is doing triage, because the AI has no gaps-filling instinct: it will do exactly what the threshold tells it to do, which means the threshold has to actually be right.

Rexfin’s reconciled model is where a threshold policy like this belongs: versioned, tied to the same ledger-verified figures the variance is computed against, and applied consistently instead of reinvented per review. To see how variance review works against your own chart of accounts, book a demo.