Automation handles volume, people handle judgment. See the difference between humans in, on, and out of the loop, and where each belongs in email security.
“Humans on the loop” means people watch over automated decisions rather than approving each one. In email security, automation handles the volume. A person then reviews what it flags, judges the unclear cases, and reports what it missed. The split from “humans in the loop” matters. Approving every decision does not scale. But removing people entirely leaves the attacks built to look real unchallenged.
The three models describe who holds the decision. Each fits a different kind of problem.
| Model | Who decides | Fits |
|---|---|---|
| Human in the loop | Person approves every action | Rare, high-consequence decisions |
| Human on the loop | System acts, person supervises and intervenes | High volume with judgment calls |
| Human out of the loop | System acts alone | Settled, low-ambiguity decisions |
Email security needs all three at once, applied to different categories. Regular simulations are what keep the human half capable of the judgment calls.
Blocking known malware belongs out of the loop, because nobody should review a signature match. Releasing a held message belongs on the loop, since the system makes a call and a person can override it. Approving a payment from an emailed request belongs firmly in the loop, with a human signing off every time.
The common failure is using one model everywhere. Teams either drown in approvals, or they automate calls that really need judgment.
Still, automated email security is good at scale and bad at context.
A filter can check a billion messages against known patterns. It cannot know that your CFO never emails accounts payable directly. It does not know this supplier always bills on the fifteenth. And it has no idea the project named in the message ended last quarter.
That context gap is where the costly attacks live. A message from a truly hacked supplier account passes every technical check, because every technical part is real. No amount of tuning catches it, since there is nothing wrong to find.
Discover how Threatcop protects your workforce from modern cyber threats.
Meanwhile, full manual review of email security fails for simpler reasons.
First, volume makes it impossible. Nobody reviews every message, and a team that tries will miss things through fatigue rather than ignorance. False positives make it worse, since roughly half of security alerts create no security value. That half still eats attention.
There is a quieter cost too. A person who spends the day clearing obvious noise stops reading closely. So the one real threat lands when their attention is already spent.
Putting people on the loop is a design choice, not a staffing one. Three things have to be true.
People need to see the decision. A system that quietly holds mail gives nobody anything to watch. Visible warnings, clear summaries, and a reason for each flag are what make oversight possible.
Stepping in has to be fast. If releasing a safe message takes two days, people stop using the process. Then they start demanding weaker filtering instead.
Reports have to go somewhere. A person who spots something wrong needs a route that reaches a responder in minutes. Without that, oversight just produces notes nobody acts on.
So this is where the model stops being theoretical.
Verizon’s Data Breach Investigations Report found a median phishing reporting rate of 21% among people trained in the past 30 days. For those not trained in the past month, it was 5%. Training roughly quadrupled reporting.
The report’s own conclusion is worth noting. Verizon’s analysts argued that stopping every click is unrealistic. So the long-term play is to reward reporting, then automate those reports into the workflow, which blocks the offending emails and links faster.
That is humans on the loop, described exactly. Not people catching everything, but people supplying the signal automation cannot create on its own.
Automation and people each cover what the other cannot. So the link between them is what matters.
Threatcop Phishing Incident Response (TPIR) exists to close that link. One-click reporting from email and WhatsApp means oversight takes seconds, not a forwarded message and an explanation. Behind the report, it runs header analysis, link and file scanning, an IP reputation check, and fake domain checks in a sandbox. So the responder gets an assessment, not raw suspicion.
Its “Who Else” insight then does something one reviewer cannot. It shows which other mailboxes got the same message, turning one person’s judgment into the scope of a campaign. Real mail returns to the inbox. Bad mail gets removed everywhere. And the loop closes without anyone working through a queue by hand.
Regular practice keeps the human half of the loop working, because oversight depends on people who still recognize what wrong looks like.
The argument is not whether to trust automation or people. It is which decisions belong to each. Let the system handle volume and settled patterns on its own. Put people on the loop where judgment matters, which means making its calls visible and its overrides fast. Keep people in the loop only for the few actions where being wrong is costly, such as approving a payment. Then connect the two halves. Oversight that produces reports nobody acts on is not oversight. It is just watching. Fast reporting is what turns it into a working loop.
Humans on the loop describes a model where automated systems act on their own, while people watch over them, review flagged calls, and step in when needed. It sits between humans in the loop, where a person approves every action, and full automation with no oversight at all.
Humans in the loop means a person approves each decision before it happens. Humans on the loop means the system acts, and a person watches the outcome, stepping in on exceptions. The first suits rare high-stakes calls. The second suits high volume with judgment involved.
Not for the attacks that matter most. AI handles volume and known patterns well. But a message from a truly hacked account has nothing technically wrong to find. Judging whether a real-looking request actually fits needs context the system does not hold.
Yes, measurably. Verizon’s breach research found a median reporting rate of 21% among people trained in the past 30 days, against 5% for those not trained recently. The effect on reporting is far stronger than the effect on click rates.
In an email security program, automate the calls that are genuinely settled. Make flagged ones visible and quick to override. Route reports straight to a responder, not a review queue. The bottleneck appears when people are asked to approve things that never needed approval.
Sushant Kumar is the AVP – Technology at Threatcop, bringing over a decade of experience in technology leadership and product development. He has worked across technology-driven organizations, including Paytm, and focuses on building scalable solutions that address evolving business and cybersecurity challenges. His areas of interest include cybersecurity technology, AI-driven security, product innovation, and enterprise technology. He is passionate about using technology to solve complex security challenges.
Attackers follow the news because crisis makes people act fast and check less. See the lure patterns and the...
The most common data breach is an employee emailing the wrong person. See the four parts of outbound email...
Every filter needs data that does not exist when a campaign launches. See why the detection gap is structural,...
