What is the ROI of doing nothing, and when is it the right answer?

Driftnet reads and scores every email Corscaden Security's customers receive before it reaches a real inbox. Amadis Ecclestone owns its model roadmap. Crowmarsh Regional Health, a three-campus hospital network and one of Driftnet's nine highest-risk customers, had a near miss last month: an email written to sound exactly like its own CFO nearly moved $184,000 to a fake vendor. Tovah Sedgeworth, Corscaden's VP of Engineering, wants a real number before she funds a quarter of engineering time to fix it.

How to answer this, stage by stage

Nobody is grading whether you can say "it depends." They are grading whether you can turn a mood, doing nothing feels risky, into an actual number, and tell them exactly what would have to be true for that number to flip.

Let's learn

Driftnet reads and scores every email that reaches a customer's inbox, in under a second, and decides whether it is safe, worth a second look, or dangerous enough to block outright.

A rules layer catches known bad senders. A learned model reads the rest for language and context, and turns it into one score.

Two years ago, before Driftnet, Crowmarsh's own security team hand-reviewed anything a basic spam filter flagged as unsure, which meant slower delivery for real mail and a small team triaging a growing pile of maybes by feel. Driftnet now auto-blocks about 9,400 emails a month across Crowmarsh's 260,000 external emails, and routes the ones it is not sure about to a security-team review queue. A year ago that queue held about 480 emails a month. It holds about 640 now.

Here is the turn. The extra emails in the review queue were never the real problem, since a person still reads every one of those. The real danger is the email that never reaches the queue at all, because Driftnet scored it low enough to just deliver.

A classic phishing email scores 70 to 95. The near-miss email, written to sound like Crowmarsh's own CFO, scored 12, well under even the review line at 40.

At its worst, this costs a real transfer. The email impersonating Crowmarsh's CFO, Adaora Thornhollow, asked Crowmarsh's accounts-payable lead to rush $184,000 to a "new vendor" ahead of a system migration. It was written using Adaora's own earnings-call transcript and a LinkedIn post, so it read exactly like her. Driftnet's dashboard never flagged it. Nobody at Corscaden even knew the model was blind to this pattern until a person happened to notice, about an hour before the payment would have gone out.

What I would leave alone: Driftnet's twenty-five lower-risk customers, small businesses with no large wire-transfer workflow. Even a fully successful attack there is unlikely to cost enough to justify moving engineering time away from the higher-risk segment first.

The lesson: a schedule that was right when the world it was watching held still stops being a schedule and starts being a blind spot the moment that world starts moving on its own.

Now here is the same thing as a story

Read the short version above when you are in the room. Read this one when you want to feel why an hour, not a headline, was the whole margin Crowmarsh actually had.

Philippa Duskin has run accounts payable at Crowmarsh Regional Health for eleven years. She knows the rhythm of a real vendor request the way a nurse knows a normal heartbeat. A real one has a purchase order number attached. A real one never asks her to skip the second signature. She has caught three fake invoices by feel in the last two years alone, before Driftnet ever flagged a single one of them.

Before Driftnet, she checked everything by hand. Once it arrived, she leaned on the review queue to catch what needed a second look. The near miss never reached the queue at all.

When Driftnet arrived, it took the worst part of her week off her desk. Instead of squinting at every external request, she trusted the ones Driftnet delivered clean and gave a proper second look to the handful it routed to review. For a year and a half, that trade held. The routine got smaller. She stopped assuming every unusual request was hiding something, because the model had been right often enough to earn that.

Then, on a Tuesday afternoon, a request landed from "Adaora Thornhollow," Crowmarsh's CFO. It asked her to push $184,000 to a new vendor before a system migration locked the payment queue for the week. No typos. No generic urgency. It even referenced a real detail from the migration timeline. Driftnet had already delivered it clean, score of 12, nothing to review.

Philippa read it twice. Something about the rhythm was wrong, not the words themselves, more like a sentence Adaora would never actually build that way. She called Adaora's office directly instead of replying to the email. Adaora had sent nothing. The payment queue closed forty minutes later. The transfer never went anywhere.

Amadis had a real story and no clean number. Tovah needed a number before she would hand over a quarter of engineering time.

The story reached Amadis Ecclestone two days later, secondhand, through Crowmarsh's account team. Amadis brought it straight to Tovah Sedgeworth, Corscaden's VP of Engineering, expecting the near miss alone to be enough. It was not. Tovah had heard near misses before. What she wanted was a number she could actually defend to her own leadership, not a story about one very alert accounts-payable lead.

The decision that opened this gap traced back to a planning meeting almost two years earlier, the week Driftnet first shipped. Someone asked how often the model should retrain. Quarterly was the answer, because back then phishing kits barely changed for months at a time, and quarterly retraining kept the model current without burning engineering time nobody had spare. Nobody wrote down a condition for revisiting that choice. It was just the setting, and settings that work quietly stop getting looked at.

Run that same near miss again, with the faster pipeline already live. This time Driftnet compares the email's writing style against ninety days of Adaora's real messages, not just its surface content, and the mismatch alongside an unusual payment request pushes the score to 61, well past the review line at 40. It never reaches Philippa's inbox marked "clean." Crowmarsh's on-call security analyst opens it, checks it, and kills it in twelve minutes. Not an hour of one person's gut standing between Crowmarsh and $184,000. Twelve minutes, by design.

What Amadis would tell that planning-meeting version of themselves: quarterly retraining was not a careless call. It was the right one for a world where attackers were still reusing kits. It was never built to survive attackers switching from a kit to a writing tool, and nobody ever asked it to.

BOUND, or the price tag on doing nothing

Not a way to dress up "I have a bad feeling about this" in five letters. BOUND is what forces a real number out of a feeling, and forces you to say exactly what would make that number wrong.

Even the low estimate clears the breakeven line. The near miss, the one real number in the whole model, sits close to it too.

Three things worth stating directly, since this is where the real judgement sits. The alternative Corscaden considered, and rejected, was cheaper and faster: just lower Driftnet's score threshold so more borderline mail lands in review. It does not work, because the near-miss email scored 12, far under even the 40-point review line; tightening a cutoff catches classic phishing harder, it does nothing for an email that never looked like classic phishing in the first place, and pushed aggressively enough it would delay real mail, referrals, lab results, vendor invoices, into a queue nobody asked to grow. The AI-specific failure worth naming by name is a form of drift: the model's learned features, typos, generic urgency, reused phrasing, no longer discriminate once attackers write with a language model instead of a kit, so accuracy looks fine overall while a whole slice of real risk goes quietly unscored. The guardrail is a held-out eval set built only from generative-style phishing, tracked separately from overall accuracy, with a standing rule that a false-negative rate crossing 25 percent on that slice forces an out-of-cycle retrain, calendar or no calendar. And the trade-off being accepted plainly: the new pipeline will flag more legitimate mail into review while it tunes, a real cost in a hospital's inbox, where a late referral is not nothing. Corscaden is accepting that friction, and the $60,000 a year it costs to run, in exchange for closing a gap a threshold change was never going to close.

And if you want to be sure it really works, try it somewhere else

Same five letters, a grain co-op instead of a hospital network, and this time the thing evolving mid-cycle wasn't an attacker with a language model, it was a fungus.

Blightscope is Tarnworth Cooperative's own crop-disease detector, built in-house by its agronomy-data team to flag early blight and mildew across 2,200 acres and 46 grower fields from drone and scout photos. It retrains once a season, a default set three years ago when the strains it watches for barely changed within a five-month growing window.

A field scout caught a spreading patch in Field 14 by eye, days before Blightscope's own score would have flagged it at its current trend. The catch saved an estimated $46,000 in crop value in that field alone. Coenraad Farrowfield, who owns Blightscope, rebuilds the case the same way Amadis did: somewhere between one and three patches a season are likely going undetected past the point where they can still be saved, at $35,000 to $50,000 each, putting $35,000 to $150,000 a season on the table. Moving to biweekly retraining during the growing season, using scout-confirmed photos, costs about $18,000 once and $9,000 a season to run. Even the low end clears that with room to spare, and the one real save already on the books, $46,000 in a single field, is bigger than the entire season's running cost on its own.

Same shape of problem as Driftnet's. A model trained once against a target that keeps changing shape underneath it.

Same rank, different swing assumption: validating against a real save comes first here too, but the number that would flip the call is different. At Crowmarsh, it was a hit probability across nine accounts. At Tarnworth, Isambeau Vosswinkel, the co-op's general manager, would only reverse the call if it turns out the new strain never spreads past Field 14, since a single contained patch would not have justified the spend, and the co-op will not actually know that until the season is already over, which is itself the argument for not waiting to find out.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to the range: the low case for doing nothing already clears the cost of building, doing nothing is only right if the hit rate is under about 6 percent.
Cost: there's no budget this quarter for the full build. Instrument the eval slice first, cheap and immediate, and treat the probability estimate as provisional until a second real data point lands.
The model got better, for real: say Driftnet's classifier already halves its false-negative rate on generative-style phishing on its own. The case still holds, since the breakeven math does not change, only the assumed probability shrinks, and it is still worth checking whether the new number lands above or below 6 percent.

Where people run it wrong.
They treat "no incident yet" as proof nothing is wrong, instead of checking whether the model's confidence has quietly drifted underneath.
They try to fix a feature problem by turning a threshold dial, and either still miss the same emails or drown real mail in false positives.
They build the range once and never re-check the one assumption it actually swings on, so an estimate calcifies into a fact nobody re-tests.

How to use it live. Ask yourself, out loud if you have to: "which single number in this estimate, if wrong, changes the answer, and how would I find out?" Naming the swing assumption is what shows real judgement instead of a rehearsed calculation.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0