Blog/Inside CyberWatch AI

Understanding Human Cyber Risk: How to Measure It Instead of Assuming It

Most organizations guess at their human risk. The signals worth measuring, how to weight them, and why an honest measure must be able to say it does not know.

CyberWatch AI18 November 2026 · 7 min read
A laptop showing a bar chart next to a notebook on a desk

Ask a security team what their human risk is and you will usually get one of three answers. A completion percentage from last year's training. A click rate from one simulation, quoted without the difficulty of the lure. Or an adjective: we think our people are fairly good.

None of those is a measurement. The first counts attendance, the second is a single reading with no baseline, and the third is a feeling. Meanwhile the same organization can tell you how many messages its mail gateway quarantined last Tuesday.

Human risk can be measured. It takes more care than measuring a technical control, because the signals are noisier and the consequences of getting it wrong land on people rather than servers. Here is what the measurement actually involves.

Start with behavior, not opinion

The first rule is that self-report does not count. A survey asking staff whether they feel confident spotting phishing measures confidence, which is uncorrelated with accuracy in a way that should worry anybody relying on it. What counts is what people did when something arrived.

That gives you a small number of genuinely observable signals.

  • What happened under test. Who opened a realistic simulated attack, who clicked it, who reported it, and how long each took.
  • What happened for real. How many genuinely suspicious messages staff reported, and what they turned out to be.
  • Judgment in situations. How people answer realistic scenarios about payments, credentials and data, where the question describes a situation rather than asking for a definition.
  • Learning actually completed. Which short lessons people finished, and how they scored.
  • Repetition. Whether the same person misses the same category of attack more than once, which is a different and more serious finding than one bad morning.

Everything else is a proxy for one of these.

Click rate is necessary and nowhere near sufficient

Click rate dominates this field because it is easy to produce and easy to put on a slide. It deserves less authority than it gets.

The problem is that click rate moves with the difficulty of the test. NIST built the Phish Scale specifically to let teams rate how hard a given lure was, because without that rating a falling click rate is ambiguous: the organization improved, or the lures got easier. A program that quietly drifts toward gentler tests will show beautiful numbers and teach nothing.

Two companions fix most of this.

Report rate measures something click rate cannot. A person who notices a phishing email and deletes it has protected themselves. A person who reports it has protected everyone, because the security team can now warn the organization and block the sender. Those two behaviors look identical in a click-rate view and are worth very different amounts.

Time to first report is the metric most organizations are not watching and should be. It is the difference between a warning that reaches people before they open the same email and one that arrives after the payment cleared. It is also the number that improves fastest when you make reporting easy, which makes it a good early indicator that a program is working.

How to build a score that is worth believing

If you combine those signals into a single number, three design decisions determine whether the number means anything.

1. Weight by what the signal tells you

Not every signal is equally informative. In our own readiness model, phishing resilience carries the heaviest weight, because behavior under a realistic test is the closest proxy available for behavior under a real attack. Threat reporting and broader security awareness carry less, training completion and behavior patterns less again. The exact weights matter less than having them stated, so anyone reading the score can see what drove it.

2. Carry the evidence with the number

A dimension that produces a bare number is a dimension nobody can act on. Each one should be able to show its own working: what percentage of how many simulated emails were clicked, what percentage of attacks were reported rather than ignored, how many people have sat an assessment out of how many employees. When a manager asks why the figure is 54, the answer should be a list of facts rather than a shrug.

3. Let it say it does not know

This is the one most scoring systems get wrong, and it is the one that decides whether the output is usable.

If an organization has never run a simulation, the honest value for phishing resilience is not zero and it is not fifty. It is unmeasured. Our model excludes an unmeasured dimension from the total and redistributes the weight across the dimensions that do have evidence, and the dimension itself reports why it is empty: nobody has been sent a simulation or answered a phishing scenario yet. A score that silently invents a number for something it has not measured is worse than no score, because it will be used to make a decision.

The same applies to bands. We use five rather than four, because with four bands everything under fifty collapses into one shade of red, and an organization at 46 is in a genuinely different position from one at 10.

Measure the distribution, not just the average

An organization-wide average is the least interesting view of human risk, because risk is not evenly spread.

Finance, payroll, procurement and the people who support executives are targeted differently and more often, because that is where the money and the authority are. The FBI's reporting on business email compromise describes attacks aimed specifically at those functions. A company average of 15 percent click rate might be five percent in engineering and forty in operations, and only one of those numbers tells you what to do on Monday.

So the useful views are a breakdown by department, and a view of concentration: how many people have clicked more than one simulation, and how many have missed the same category of attack repeatedly. A handful of repeat patterns is a specific, fixable finding. An average is a mood.

It is also worth weighting a person's own risk by how exposed their role is. A click by someone who approves payments is not equivalent to a click by someone who cannot move money, and a model that treats them identically will point your attention at the wrong place.

Let old evidence fade

People change, and a measurement system that never forgets will misrepresent them.

Someone who clicked a convincing lure eighteen months ago, took the lesson, and has reported every test since is not the same risk they were. Our per-person model decays the weight of each piece of evidence over time, with a half-life of around four months, so recent behavior dominates and old mistakes stop following people around. The practical effect is that the model rewards improvement, which is the behavior you actually want to encourage.

It also means the score is a trend instrument rather than a snapshot. One round of measurement tells you where you are. Only a sequence tells you whether anything is getting better, which is the question leadership is really asking when it asks for a number.

A caution on comparison. Human risk scores are not comparable between organizations, because they depend on test difficulty, role mix and how reporting is set up. The only comparison worth making is against your own previous months.

Where measurement becomes surveillance

Every system described here could be turned into something harmful, and that is worth saying plainly rather than hoping nobody notices.

Per-person risk modeling is useful because it decides what someone practices next. It becomes corrosive the moment it becomes a ranking that affects how people are treated. The distinction is not subtle, and it shows up in design choices: whether leaderboards are anonymous by default, whether a single click generates a lesson or an email to a manager, whether an administrator can read a check an employee chose not to report.

The practical test is simple. If being measured makes people less likely to report a real attack, the measurement has made the organization less safe, whatever the dashboard says. The NCSC's board guidance on security culture is blunt about this: a culture where people fear admitting a mistake produces silence, and silence is what attackers rely on.

Starting from nothing

If none of this exists in your organization yet, the sequence is short. Run one realistic simulation on one department, so you have a reading rather than an assumption. Make reporting one step and answer every report, so the reporting signal starts existing. Put a verification rule in writing for the processes attackers target. Then measure the same things again in a month.

That is enough to replace an adjective with a number you can defend. For the practice this sits inside, see human risk management; for the metrics in more detail, measuring security progress; for how the tests themselves should be built, phishing simulations explained. If you are doing this for a smaller organization, the risk review walkthrough and the small business cybersecurity guide cover where the human layer fits alongside everything else. And while the program is being built, anyone in your organization can check a suspicious message for free with CyberWatch AI Scan.

Frequently asked questions

What is human cyber risk?

The security risk arising from how people behave: what they click, what they report, how fast they report it, and how they respond to pressure in processes like payments and password resets. It is observable, which means it can be measured rather than assumed.

Is click rate a good measure?

It is a useful signal and a poor single measure. It moves with the difficulty of the lure, so a falling click rate can mean improvement or easier tests. It needs to sit alongside report rate and time to report.

What is the most underrated metric?

Time to first report. It decides whether a security team can warn everyone and block a sender before the rest of the organization opens the same email.

How do you measure something you have no data for?

You do not. An honest measure reports a dimension as unmeasured rather than filling it in. A score built partly from guesses cannot be used to make a decision.

Should individual employees have risk scores?

Only if they are used to assign help rather than consequences. A per-person model is valuable for choosing what someone practices next and damaging if it becomes a ranking people are judged by.

Why weight evidence by how old it is?

Behavior changes. A click eighteen months ago says much less about someone today than a click last month, so older evidence should count for less rather than following a person indefinitely.

Why measure by department?

Risk is not evenly distributed. Finance, payroll, procurement and executive assistants are targeted differently from engineering, and an organization-wide average hides exactly the concentration you need to act on.

How often should this be measured?

Continuously rather than annually. The useful view is a trend across months, because a single round tells you where you are and only a sequence tells you whether anything is improving.

Can a human risk score be compared between organizations?

Not reliably. Scores depend on the difficulty of the tests, the mix of roles and how reporting is configured. The comparison worth making is against your own previous months.

Sources

  1. SP 800-50 Rev. 1, Building a Cybersecurity and Privacy Learning Program, NIST Computer Security Resource Center
  2. The Phish Scale: NIST-Developed Method Helps IT Staff See Why Users Click on Fraudulent Emails, NIST
  3. Cyber Security Toolkit for Boards: Understanding the cyber security threat, UK National Cyber Security Centre
  4. 10 Steps to Cyber Security: Risk management, UK National Cyber Security Centre
  5. Cross-Sector Cybersecurity Performance Goals, US Cybersecurity and Infrastructure Security Agency
  6. Cybersecurity Framework, NIST
  7. Awareness and Cyber Hygiene, European Union Agency for Cybersecurity
  8. Annual Reports, FBI Internet Crime Complaint Center
Share this article
LinkedIn WhatsApp X
For organizations

Your people get these messages at work too.

CyberWatch AI sends your staff realistic practice attacks, trains the gaps it finds, and shows management exactly where the organization stands.