A risk assessment answers three questions in order. How bad is this before we count our controls? How much do our controls actually take off? Is what is left something we are willing to carry? The first answer is inherent risk, the second is control effectiveness, the third is residual risk compared against risk appetite. Most registers have columns for all four. Far fewer can show how one number was turned into the next.
| Term | What it means | How it is measured |
|---|---|---|
| Inherent risk | The exposure of an activity, system or process before the controls you are assessing are credited | Likelihood × impact on a 5×5 scale, or a loss figure such as annualized loss expectancy (ALE) |
| Control effectiveness | How much a control lowers the likelihood or the impact. It has two halves: design effectiveness (is the control built so that, if it runs, it stops the risk?) and operating effectiveness (did it actually run, consistently, through the period?) | A walkthrough proves design; sampled test evidence proves operation. Expressed as a percentage or a rating |
| Residual risk | The exposure left once the controls operate in the real world. It is never zero | The same scale as inherent risk, so the two can be compared directly |
| Risk appetite | How much risk, of which kinds, the organisation is willing to carry in pursuit of its objectives | A board-approved statement, usually with a threshold per risk category |
COSO’s 2017 ERM framework splits residual risk in two: target residual risk (where management wants it) and actual residual risk (where it really is). The gap between them is the work still to do.
1 · The Formula
The common quantitative model treats control effectiveness as the share of the inherent risk the controls remove:
Worked through with money: a ransomware scenario with an inherent ALE of $100,000, and controls judged 70% effective, leaves
The same arithmetic works on a score: an inherent 20 out of 25 with 70% effectiveness becomes 6. The formula is only as good as the 70%, though, and that number should come from two separate judgements:
A control designed to stop 90% of attempts that ran in 80% of the sampled periods is worth about 0.90 × 0.80 = 72%, not 90%. That product is a modelling convention rather than a standard’s requirement, but it forces the right question: what did the testing show? Auditors test the two halves separately for the same reason — PCAOB AS 2201, the standard behind SOX 404 audits, asks for evidence of design effectiveness and of operating effectiveness.
Try it: residual risk calculator
Use any unit — dollars of ALE or a 1–25 score — as long as inherent risk and appetite share it. Controls combine as 1 − (1 − e₁)(1 − e₂)(1 − e₃), which assumes they fail independently (see the next section). Leave a control at 0 if you do not have it.
2 · Stacking Controls: Why 70% + 70% Is Not 140%
Real risks have several controls. Adding their percentages produces nonsense above 100%; averaging them says a second control adds nothing. The right model is that each control stops a share of what got past the ones before it:
| Controls | Gets through | Combined | Residual of 100 |
|---|---|---|---|
| One 70% control | 0.30 | 70% | 30 |
| Two 70% controls | 0.30 × 0.30 = 0.09 | 91% | 9 |
| Three 70% controls | 0.30³ = 0.027 | 97.3% | 2.7 |
| 70% + 50% control | 0.30 × 0.50 = 0.15 | 85% | 15 |
▸ The product rule assumes the controls fail independently. MFA, conditional access and privileged-access management all sitting on one identity provider are three controls with one shared failure: when the IdP is compromised, all three go at once. For correlated controls, credit the strongest one and treat the rest as partial, not multiplicative.
This is the arithmetic behind defense in depth: each independent layer shrinks what gets through by a factor, so the third layer is still worth having even when the first two look strong.
3 · Where the Formula Misleads
The single-percentage model is a useful teaching device and a poor audit trail. Five traps account for most of the bad residual numbers in real registers:
- A control works on likelihood or impact, rarely both. MFA makes account takeover less likely; it does nothing to the cost of a breach once it happens. Backups do the reverse. One “effectiveness %” applied to the whole risk hides which side moved — so apply it to the side it actually acts on (worked example in the next section).
- Ordinal scores are not quantities. On a 1–5 scale, a 4 is not twice a 2; the numbers are labels in order. Multiplying 20 by 0.30 to get 6 treats them as ratios. It gives a different band from re-scoring likelihood and impact after the controls, and neither is “the” right answer. Pick one method and use it for every risk — see Semi-Quantitative Risk Scoring for why.
- Double counting the baseline. If inherent risk was scored with existing controls already in mind — as NIST SP 800-30 and FAIR analyses normally do, because they assess the environment as it is — then applying those controls again overstates the reduction. Write down which controls the inherent score assumes.
- Design is not operation. A well-designed control that failed its operating test is worth 0% for that period, however good the design. Credit only controls whose operation was tested and passed; re-rate the moment one fails.
- Correlated controls are multiplied as if independent. See the shared-identity-provider example above. Stacking overstates protection whenever controls share a dependency.
4 · One Risk, Two Ways: 5×5 Scores and FAIR
Take one risk: ransomware via a compromised remote-access account, worked qualitatively and quantitatively. (The same risk is scored with other methods in Qualitative Risk Matrix, Semi-Quantitative Risk Scoring and OWASP Risk Rating.)
| Step | 5×5 score (re-scored) | FAIR-style ALE |
|---|---|---|
| Inherent | Likelihood 4 × Impact 5 = 20 (High) | 0.5 events/yr × $200,000 per event = $100,000 |
| Add MFA acts on likelihood, 70% | Likelihood falls to 2: 2 × 5 = 10 | 0.5 × 0.30 = 0.15 events/yr × $200,000 = $30,000 |
| Add tested offline backups acts on impact, 50% | Impact falls to 3: 2 × 3 = 6 (Low) | 0.15 × $100,000 = $15,000 |
| Combined | 20 → 6 | $100,000 → $15,000, i.e. 85% = 1 − 0.30 × 0.50 |
In FAIR, risk is the probable frequency and magnitude of future loss:
Controls do not enter FAIR as a single percentage. They lower threat event frequency (fewer attempts reach you), vulnerability (fewer attempts succeed, because resistance strength rises against threat capability) or loss magnitude (each event costs less). That is the same lesson as trap 1, built into the model. The 5×5 column reaches a similar end point here, but it cannot tell you that the residual is worth $15,000 — or whether a $40,000 control to remove it is worth buying.
▸ Use scores to rank many risks quickly; use a loss model like FAIR to decide on the few where the money for a control is real. The formula on this page is the bridge between them.
5 · Deciding Against Appetite
A residual number on its own decides nothing. Compared with the appetite for that category of risk, it leads to one of four treatments:
| Treatment | When | Example |
|---|---|---|
| Accept | Residual is within appetite | The risk owner signs and the risk is reviewed on schedule |
| Reduce | Above appetite, and a control costs less than the loss it removes | Add phishing-resistant MFA, segment the network |
| Share | Above appetite, and a third party can carry part of it | Cyber insurance, contractual liability on a vendor |
| Avoid | Above appetite, and no affordable control brings it down | Retire the internet-facing legacy system |
ISO 31000:2018 lists more options than four (it also names taking on risk to pursue an opportunity, removing the risk source, and changing the likelihood or the consequences); ISO/IEC 27005 groups them as modification, retention, avoidance and sharing. They collapse to the four above in practice.
ISO/IEC 27001:2022 clause 6.1.3 (f) makes the last step explicit: the organisation must obtain the risk owners’ approval of the risk treatment plan and their acceptance of the residual information security risks. Residual risk is not accepted by the security team on someone’s behalf; it is signed by the person who owns the risk. Then it is watched — a key risk indicator with a threshold tells you when the residual has drifted back above appetite.
Bottom line
- Residual = Inherent × (1 − Control effectiveness) — a useful model, as long as the effectiveness figure is earned.
- Effectiveness needs both tests — design × operation; a control that failed testing counts as 0%.
- Stack by multiplying what gets through — two 70% controls make 91%, if they fail independently.
- Apply each control to the side it moves — likelihood or impact, not the whole score.
- Write down the baseline — say which controls inherent risk already assumes, or you will count them twice.
- The owner signs what is left — ISO 27001 6.1.3 (f), then a KRI watches it.
Sources & method. The residual-risk formula as commonly stated in GRC practice (for example ServiceNow community GRC guidance and Aravo’s inherent-vs-residual explainer); the combination rule for independent controls is standard probability. Definitions and treatment options from ISO 31000:2018, ISO/IEC 27005:2022 and ISO/IEC 27001:2022 (clause 6.1.3); target and actual residual risk from COSO ERM (2017); likelihood and impact assessment from NIST SP 800-30 Rev. 1; the frequency × magnitude loss model from Open FAIR (The Open Group). Design and operating effectiveness from PCAOB AS 2201. All figures in the worked examples are illustrative, not benchmarks. Original analysis; no third-party graphics or text reproduced.
← Back to Cyber security