Why The Most Convenient Feature Is Often Wrong

What convenience quietly removes from your product, what changed when AI made effort free, and how to design and measure the right amount of friction 

An edtech client running live online classes sent us a feature request that looked like an easy win: 

"Integrate Zoom so attendance is marked automatically when a student joins." 

No roll-call, no spreadsheets, no complaints. Ishita read it and asked: "What is attendance for?" 

It isn't for the spreadsheet. The institute promises parents that students learn in the live session, and attendance is supposed to be evidence of that. Auto-marking on join only records that a device connected. A student can join, mute, switch off the camera and walk away, and still be marked "present." 

Srikanth added the part that made it urgent: in 2026, joining doesn't even prove a human was there. AI note-taking bots now join video calls on people's behalf, record them and send back a summary. Universities including Cornell and Tufts have started blocking them. 

So we proposed something slightly less convenient: the teacher drops a short code into the session at a moment of their choosing, and students enter it on the platform within 90 seconds. Nobody knows when it's coming, so the reliable way to be marked present is to stay for the whole class. 

It costs students five seconds and teachers one tap, and it changes what the attendance number means. 

That's the idea behind this post. Convenience is a tool, not the goal. Optimise for it blindly and it quietly removes the things that matter most: the meaning of an action, the feeling behind it, the human connection in it and the security around it. AI has made this far more urgent. 

Four things convenience can quietly remove 

Removing friction is usually right for low-stakes, frequent, reversible actions. Nobody wants a password prompt to like a post. The mistake is applying the same logic to actions where the friction is doing a job. 

Before removing a step, ask which of these four it was protecting. 

Before removing any step, ask four questions: 

1. Meaning: does the action still prove what we actually care about? 
2. Feeling: can the person still feel what they're doing? 
3. Connection: does the other person still feel the effort behind it? 
4. Security: is there still a moment to notice that something is wrong? 

The pattern shows up in every industry: 

• Meaning, in education: auto-attendance measures access, not presence. 
• Feeling, in payments: on another project, we tested approving UPI payments with face recognition alone. Users loved it for small amounts. But around ₹1,000 they started hesitating and asking for the PIN back. One said, "For big amounts, I want to feel that I'm paying." When India's payments body, NPCI, launched biometric UPI approval in October 2025, it capped it at ₹5,000. The industry drew the same line our users did. 
• Connection, in communication: a 2023 Ohio State University study found that people who received a message a friend had written with AI help felt less satisfied with the friendship and less sure where they stood. 
• Security, everywhere: in 2024, a finance employee in Hong Kong joined a video call with what looked like the company's CFO and colleagues. All of them were deepfakes. He transferred about US$25.6 million. Familiar faces are the most convincing kind of trust there is. 

What AI changed 

Most research on friction comes from the 1980s to the 2010s, when effort was expensive. A long message meant someone cared. Joining a call meant a person was there. Seeing a face meant it was really them. Effort and presence were reliable signals because they were hard to fake. 

AI has broken that almost all at once. 

In 2010, friction was a cost to remove. In 2026, the right friction is proof. 

• Presence is fakeable: bots attend meetings for you. 
• Effort is fakeable: a heartfelt message or cover letter takes two seconds. LinkedIn has reported around 11,000 job applications a minute, up about 45% in a year, driven partly by AI auto-apply tools. 
• Identity is fakeable: deepfakes can pass a live video call. 
• Learning is skippable: in a 2025 study of nearly 1,000 high-school students (published in PNAS), students with unrestricted access to GPT-4 scored 48% higher in practice but 17% lower on the exam once access was removed. A version designed to give hints instead of answers largely avoided the harm. 

The shift in one line: when generating anything is free, proof of genuine human intent becomes the valuable thing. Products that remove all friction also remove the only signals that separate a person from a bot, care from autopilot and learning from copying. 

The research behind it 

These nine ideas explain why the right friction works. Most are decades old, and all of them matter more in an AI world. 

Nine ideas, from 1983 to 2025. 

1. Goodhart's law: when a measure becomes a target, it stops being a good measure. Make "attendance marked" the target and people (and now bots) optimise for being marked, not for learning. 
2. Desirable difficulties (Bjork): some struggle is how learning sticks. The 2025 AI study is this idea in action. 
3. The pain of paying (Prelec and Loewenstein, 1998): the small discomfort of paying is how people feel and control their spending. In a related study, people bid up to twice as much for the same tickets when told they'd pay by card instead of cash. 
4. The effort heuristic (Kruger and colleagues, 2004): we judge quality and care by perceived effort, which is why AI-written messages can feel less sincere. 
5. The labour illusion (Buell and Norton, 2011): showing the work behind a service makes people value it more. But a 2020 study found that slow responses make people trust algorithms less. The lesson: show what the system checked; don't add delay. 
6. Fluency feels safe (Song and Schwarz, 2009): things that are easy to process feel familiar, and familiar feels safe. That's exactly what deepfakes and frictionless scams exploit. 
7. The ironies of automation (Bainbridge, 1983): automate most of a job and people are left with the hardest exceptions, with less practice at them. 
8. The moral crumple zone (Elish, 2019): when automation fails, the nearest human often takes the blame, even if they had little control. 
9. The IKEA effect (Norton, Mochon and Ariely, 2012): people value what they help build, so let users shape part of the system themselves. 

Why convenience-first deployments stall in large companies 

This explains a pattern we see often. A forward-deployed engineer (an engineer embedded with a client to make a product work in their environment) walks into a large company and does what they do best: removes steps. Nine clicks become one. The demo is impressive, and then the room goes quiet. 

They optimised for convenience. The enterprise was buying control: who signs off, what gets checked, how fast a mistake can be undone and who is accountable when it isn't. Remove a human checkpoint while leaving someone's name on the decision and you've created a moral crumple zone, and they will rightly refuse it. 

MIT's GenAI Divide report (2025) found that around 95% of enterprise generative-AI pilots stalled before delivering measurable returns. It blamed not model quality but a "learning gap": tools that didn't fit real workflows or earn the trust needed to scale. 

Convenience is a consumer metric. Control is an enterprise one. 

The teams that succeed keep the checkpoints and make them faster, and let the client configure the thresholds and rules themselves. The pitch changes from "less work" to "you keep control, and it's faster." 

Four approaches to designing friction on purpose 

1. Match friction to the stakes 

Friction should rise with what's at stake and fall as trust is earned. Payment regulators worked this out early, because their stakes come in rupees and euros: 

Friction set by the stakes, in payments. Limits as of late 2026. 

India allows small UPI Lite payments up to ₹1,000 without a PIN and caps biometric approval, bringing back the PIN above it. Europe lets payments up to €30 skip strong authentication, and allows larger exemptions only for providers whose fraud rates stay very low: low risk earns convenience. Since October 2024, UK banks can hold a payment for up to four business days when they suspect fraud. 

Apply the same ladder to any product: 

• Low stakes (a like, a draft, a small payment): remove the friction. 
• Medium stakes (attendance, a mid-size payment, an AI agent editing a document): add light friction that proves intent, like a code, a biometric or a clear confirmation. 
• High stakes (large transfers, grades, hiring decisions, irreversible changes): add deliberate friction with context, a human checkpoint and an undo window. 

2. Combine the convenient signal with an intentional one 

You rarely have to choose. Use the convenient data as a baseline and add one small intentional step. That's how we built attendance: 

Option C keeps Zoom's free data and adds a small proof of presence. 

• Baseline (free): Zoom's join and leave data gives minutes present for every student. 
• Intent (cheap): the teacher drops two codes per session at moments they choose, each valid for 90 seconds, so sharing one in a group chat rarely arrives in time. 
• Fairness: students who lose connection can request a review, which the teacher approves with one tap. 
• A better report: the dashboard no longer says "attended." It says "presence verified: 2 of 2 codes, 54 of 60 minutes," and parents trust it more because it means more. 

The same pattern works elsewhere: face recognition plus a confirmation screen for large payments; one-click applications plus one question only a genuine reader of the job description can answer; an AI agent that drafts plus a human who approves above a threshold. 

3. Run the friction budget 

Every step has a cost and a benefit, but most teams only measure the cost, because it shows up in tomorrow's conversion numbers. The benefit shows up later, in someone else's report. 

Keep the step if the value it protects is greater than the value it costs. 

Protects: fraud and bad outcomes avoided, trust earned. Costs: completions lost, time, support load. 

Here's one verification step tested across three payment sizes, with illustrative numbers: 

The same step, three payment sizes, opposite answers. 

Below ₹1,000, the step costs about twice what it saves, so remove it. Between ₹1,000 and ₹10,000, it pays back about 79 times, so keep it light. Above ₹10,000, it pays back nearly 2,000 times, so make it deliberate. For attendance, the code costs each student about ten seconds per class and protects the one number the institute sells to parents. 

4. Make the step meaningful 

Bad friction is a wall. Good friction tells you something: 

Each element of this screen does a job. The same principles apply to any confirmation step. 

1. Say what's about to happen, in plain words. "You're sending ₹18,500 to Sharma Electronics." 
2. Give context, not warnings. "This is 4× your usual payment." 
3. Show the checks the system ran: a verified name, a known device. 
4. Let people set their own threshold, such as "ask for my PIN above ₹___." 
5. Offer a way back: an undo window or a review request. 
6. Make intent easy for people and hard for bots. 

Measure the outcome, not the shortcut 

The deepest mistake convenience makes is measuring the shortcut ("marked present," "payment completed," "application sent") instead of the outcome (learned, trusted, hired well). Track both, and the gap between them. 

Two numbers matter everywhere: 

• The gap between the shortcut and the outcome. If 96% are "marked present" but only 71% are "presence verified," that 25-point gap is the illusion convenience was selling. 
• How often users choose more friction than you require. When they pick "use my PIN instead," they're telling you where you went too far. 

What to do this week 

Convenience is still a superpower, but it isn't the definition of quality. The best products are effortless where the stakes are low, deliberate where they matter, and clear about which is which. In a world where AI can generate, fake or automate almost anything, that clarity is becoming the product. 

1. Pick your most convenient feature and ask the four questions: meaning, feeling, connection, security. 
2. Find one shortcut metric you report and start measuring the real outcome beside it. 
3. Add one moment of intent that's easy for a person and hard for a bot. 
4. Run the friction budget on your highest-stakes workflow. 
5. Replace one loading spinner with a list of what the system checked. 
6. Before automating in a large organisation, map who signs off on each decision and give them thresholds, overrides and an undo. 

The client's attendance figure dropped the week we shipped the codes. The institute's director called Srikanth, and we expected a complaint. Instead she said, "For the first time, I believe this number." 

Design AI products people trust 

Removing every step isn't the same as building a better product. AtliQ helps businesses design AI-powered products and workflows with the right checkpoints, guardrails and human oversight, so they're fast where it's safe and deliberate where it matters. 

Rolling out automation or AI agents in your organisation? Talk to our experts about where friction should stay. 

Studies and limits as of October 2026; check current RBI, NPCI and EU texts before relying on them. Client stories are anonymised, and the friction-budget figures are illustrative. 

Sources 

•Bastani et al., PNAS 2025 (Wharton summary) 

•Ohio State: AI-assisted messages to friends 

•Hong Kong deepfake video-call fraud (AI Incident Database) 

•Cornell IT: AI note-takers in meetings 

•LinkedIn's 11,000 applications a minute (eWeek) 

•MIT report on GenAI pilots (Fortune) 

•Ironies of Automation, Bainbridge 1983 (PDF) 

•Song & Schwarz 2009 (PubMed) 

•Efendić et al. 2020 (Tilburg) 

•NPCI biometric UPI launch · UPI Lite ₹1,000 limit (Business Today) 

•PSD2 SCA exemptions (Checkout.com) · UK payment delays (Grant Thornton)