UX Roundup: AI by Design Phase | Product Managers vs. Designers | Individualized Tutoring | Text Labels | Stopping AI | Unsubscribe | OODA Loop | Timeouts | Buzzwords | Forced Registration
Summary: Designers use AI most for ideation and least for finishing | Product managers benefit more from AI than designers | AI tutors as good as human experts but 918 times cheaper | Icons need text labels | Users must be able to halt an AI run | Easy-in, easy-out | Observe–Orient–Decide–Act (OODA) | Recovering from a timeout | Buzzwords make writers seem less intelligent | If you want customers’ money, don’t force them to jump through hoops to do business with you
UX Roundup for October 19, 2026 (GPT Image 2.5)
Designers Use AI Where Ideas Are Cheap: 72% in Discover, 41% in Deliver
A survey of 443 designers found that AI use peaks in the first phase of the Double Diamond and bottoms out in the last. Designers have figured out what I’ve said since 2023: ideation is free with AI, but finishing still takes human effort. What sets holdouts apart is doubt about AI’s utility.
This study used the famous “Double Diamond” model of the design process to classify AI use. (GPT Image 2.5)
Designers have voted with their prompts: AI belongs at the start of the design process. Sepideh Tajarmakan (an independent researcher) and Khashayar Hojjati Emami from the Iran University of Art asked 443 designers in 43 countries about their AI use in each phase of the Design Council’s Double Diamond (arXiv preprint). Most were product, industrial, or UX/UI designers.
AI traffic has worn the Discover step hollow, while cobwebs gather on Deliver. (GPT Image 2.5)
Later Phase = Fewer AI Users
Overall, 80% of respondents used AI in at least one phase, but only 35% of these adopters used it in all 4. Adoption accumulated in workflow order: Discover alone was the most common single phase, Discover plus Define the most common pair, and so on. Nobody started with Deliver. (Well, 2 did.)
Right now, AI use in design starts strong but finishes weak. I expect AI to run the full course through all 4 phases next year. (GPT Image 2.5)
The decline is a matter of headcount: fewer designers bring AI along, but those who do pick nearly as many tools in Deliver (1.89 on average) as in Discover (2.09).
Ideation Is Free, So AI Rules Divergence
In 2023, I reviewed 12 creativity studies and concluded that ideation is free with AI: more shots on goal score more goals, and AI supplies the shots at almost no cost. I also predicted that AI would contribute most during divergence, while humans would drive convergence. Score that prediction mostly RIGHT.
Ideation used to be expensive and required a snowstorm of sticky notes. With AI, it’s cheap, so do more of it. (GPT Image 2.5)
Once AI has generated a mountain of design ideas, someone still has to judge them, and that someone is senior UX staff. (GPT Image 2.5)
Discover, the first divergent phase, has the most AI users. In Develop, the second divergent phase, standalone generators (mostly for images) took 58% of all tool picks, vs. 7% or less in the other phases. The exception is Define, a convergent phase that still ranks second in AI users. But defining a problem is largely synthesis, another AI forte.
Defining a problem means distilling a vat of research into a few drops of insight, and AI runs a fine still. (GPT Image 2.5)
Why does AI use decline steadily across the phases? My explanation: early errors are free; late errors ship. A weird AI idea in Discover costs nothing: you discard it, or it sparks a better one. A wrong AI edit in Deliver reaches customers.
Shipping the wrong idea is expensive, so the human curator still holds the veto. (GPT Image 2.5)
General Chat Is the Accidental Design Tool
Only 9 of the 73 AI tools named were general-purpose assistants, yet they won most tool picks in 3 of the 4 phases. ChatGPT alone took a whopping 52% of picks in Define. So most early-stage AI design work happens in tools that weren’t built for design. Ironically, as the authors note, the AI design tools that researchers build concentrate on Develop and Deliver, the two phases where designers use AI least.
Most AI design tools are specialized gadgets, yet designers mostly reach for a general chatbot. (GPT Image 2.5)
Finishing has its own mode: AI built into familiar software (such as Photoshop’s Generative Fill) peaked in Deliver at 38% of picks. Near the finish line, designers want AI inside trusted tools, working on a selection they control.
AI built into familiar software peaked in the Deliver phase, and it may not even feel like AI to the designers using it. (GPT Image 2.5)
Holdouts Doubt Utility, Not Usability
Of 10 barriers, only 1 separated the 84 holdouts from the adopters: 40% of holdouts found AI insufficiently useful, vs. 16% of adopters. Ease of use didn’t distinguish the groups (12% vs. 10%). In my 1993 book Usability Engineering, I split usefulness into utility and usability. The holdouts question utility.
Adopters griped more about cost (37% vs. 24%, though the difference didn’t survive statistical correction). Ideas are free; subscriptions aren’t.
AI cost is a worry, but prices are dropping fast. (GPT Image 2.5)
The data was collected from August 2025 to February 2026, an eon in AI years, and since current AI is the worst we’ll ever experience, the holdouts judged what is now an obsolete product. If you’re one of them, retest AI on a real project whenever a new frontier model ships.
Swearing off AI after a 2025 test is like swearing off bananas after biting into a green one. AI ripens fast, so give today’s frontier models a fresh taste, and plan to rethink where and how to use AI again next year. (GPT Image 2.5)
Seen through the wrong end of the telescope, AI looks small. Try it on a real project instead. (GPT Image 2)
Weak Sample, Strong Pattern
The authors recruited respondents through social media and their own networks: 47% came from Iran, and 60% were in their twenties. AI enthusiasts are likelier to answer an AI survey, so the 80% adoption rate is probably inflated. ChatGPT was also a preset checkbox, whereas Gemini and Claude had to be typed in, which tilts the tool rankings.
Self-reports are slippery, too: in a general question, only 107 respondents said they used AI in Deliver, but 182 confirmed it when asked about that phase specifically. (My guess: AI built into familiar software doesn’t feel like AI.) Still, both questions show the same decline, so I trust the shape of the curve, if not every percentage point.
If you just ask general questions, respondents may not remember everything they use AI for. (GPT Image 2.5)
AI Is Startware, for Now
Today’s AI is startware: designers trust it to open a project far more than to close one. So exploit free ideation: ask for 20 alternatives, not 3. Then reinvest the saved hours in convergence, which remains the human bailiwick: winnowing ideas, testing with real users, and finishing the design.
As AI floods the field with options, the designer’s job shifts from producing ideas to winnowing them. (GPT Image 2.5)
AI vendors, take note: general chatbots already own the first diamond, so the untapped market is finishware that designers trust to deliver. My prediction: if this survey is repeated in 2028, Deliver will have caught up with Discover. Designers voted with their prompts for the first diamond. AI must earn their vote for the second.
When ideation is free, the bottleneck moves downstream to finishing. (GPT Image 2.5)
My prediction: Deliver will catch up with Discover by 2028. (GPT Image 2.5)
AI Design Tools Help Product Managers Twice as Much as Designers
In a randomized experiment, an AI prompt-to-design tool cut product managers’ time on design tasks by 35% but designers’ time by only 17%, shrinking the gap between the two roles by about 70%. That echoes the skill-leveling effect found in many earlier studies. But participants got a picture of the finished design, so AI closed the execution gap while leaving the judgment gap untested.
Product managers using an AI design tool completed standard design tasks in 35% less time than PMs working by hand. Designers gained half as much, and only on the hardest task. So the group with less design skill benefited most. The forklift for the mind has entered the design studio.
AI lets product managers crash straight through the wall into design work instead of filing a request and waiting for the design team. (GPT Image 2.5)
AI is a forklift for the mind, which lifts less-skilled staff the most (in this study, PMs doing design work). The “no pain, no gain” slogan holds for AI in education, but not for production tasks. (GPT Image 2.5)
The Experiment: 100 Professionals, 3 Tasks, 1 Answer Key
Remy Stewart and colleagues at Figma ran a randomized controlled trial of Figma Make, the company’s prompt-to-design tool (arXiv, 2026). The independent research firm MeasuringU recruited 50 product designers and 50 product managers and moderated the sessions, which lends credibility to a vendor’s study of its own product. Half of each group got Figma Make (after a 5-minute primer), while the rest designed by hand in regular Figma.
A 5-minute primer on the AI tool erased much of the advantage that designers had built through years of practice. Traditional skills are becoming museum pieces as AI takes over those tasks. (GPT Image 2.5)
Everybody did the same 3 tasks on a social media app: convert a screen to dark mode, add a “Help & Support” item to a settings menu, and prototype a comment flyout. Each task came with a picture of the finished design, and the clock stopped when a moderator confirmed a match.
Prompting is like briefing a police sketch artist: you describe what’s in your mind and hope the AI draws it right. (GPT Image 2.5)
PM + AI ≈ Designer
Averaged across tasks, AI cut completion times by 20%, but the gains were lopsided:
AI also made PMs 38 percentage points more likely to finish the flyout. By my arithmetic from the paper’s regression model, when both groups worked by hand, PMs needed 26 minutes more than designers to complete all 3 tasks; when both had AI, only 8 minutes more. Roughly 70% of the skill gap disappeared.
Usability ratings converged, too. Overall, designers gave Figma a System Usability Scale (SUS) score of 70, while PMs gave it only 51. (73% of designers use it several times a day, vs. 37% of PMs.) AI raised PMs’ scores by about 15 points and did nothing measurable for designers. Prompting made a complex tool friendlier to its occasional users.
Same Pattern, New Profession
AI narrowing skill gaps is one of the best-replicated findings in AI productivity research. I reported it in 2023 for support agents, business writers, and programmers, then for elite BCG consultants and for users of different education levels. This summer, 2.26 million freelance contracts showed the labor market pricing it in.
Studies in different domains keep showing that AI narrows skill gaps. A finding that replicates this reliably deserves its own museum wing. (GPT Image 2.5)
The Figma study adds a between-roles twist, also seen at Procter & Gamble, where AI let R&D staff write commercially sound product proposals and commercial staff write technically sound ones. So AI narrows gaps across specialties, supporting my prediction of a rebirth of the UX unicorn.
The true UX unicorn may still be a work in progress, but we’re getting closer to one person doing everything, and we’re already close enough for many projects to work this way. (GPT Image 2.5)
Delegatency: Why Designers Gained Nothing on Easy Edits
Why didn’t designers speed up on the easy edits? Every AI request carries delegatency, the fixed time cost of handing off work: writing the prompt, waiting for the generation, and checking the result. (Figma Make added occasional broken renderings that needed extra prompts.) An expert can add a menu item faster than he or she can describe it. Only on the flyout, which took designers 24 minutes on average, did AI’s speed outrun its overhead.
History rhymes, but backward. In 1984, the Macintosh made direct manipulation the beginner’s interface. In 2026, the beginner’s interface is prompting, while experts make small edits faster by hand.
Reverse history: typing a prompt is now the beginner’s way to design, while experts still reach for the mouse. (GPT Image 2.5)
The study measured how fast people reproduce a known design, which is where AI is strongest and designers’ edge is thinnest. In real projects, nobody hands you the reference image. Deciding what the design should be is the design job.
The study was design by numbers: the reference image came in the box, and participants filled in the colors. That’s the smallest part of the real job. (GPT Image 2.5)
AI generates endless variants, which makes deciding what to build the scarce skill, worthy of a senior designer. (GPT Image 2.5)
What to Do
PMs: Show your ideas as prototypes instead of describing them in specs, then let a designer judge them.
“Show, don’t tell” now applies to design ideas: PMs can prototype a design faster than they can write a spec, which nobody wanted to read anyway. (GPT Image 2.5)
Designers: Stop selling tool speed. Sell judgment: user research, deciding what to build, and picking the best AI variants. Delegate the long, fiddly prototyping to AI.
Managers: Expect more design work to originate outside the design team, and make designers its quality gate.
A forklift lets anybody move a one-ton pallet, bodybuilder or not. AI gave PMs a forklift for design work. But in this study, somebody had already decided which pallet to lift and where to put it: every task came with a picture of the finished design. Take away that picture, and I predict that the gap between PMs and designers will reopen exactly where it matters most.
AI Tutors Now Match Human Experts at 1/918 the Price
AI tutors now teach as well as human experts. Curtis Northcutt and colleagues at Handshake AI gave 2,383 students one hour of GRE prep with one of 13 AI tutors, an expert human tutor, or nature documentaries (the control group).
In this study, the control group watched nature documentaries. Tutoring was better. (GPT Image 2.5)
Both kinds of tutoring roughly doubled the control group’s score gain, and pooled AI tutoring was statistically equivalent to the human experts (p=0.015). In 5 of the 7 GRE topics, the best AI tutor beat the humans. (Caveat: learning was measured only right after the session.)
This is AI in education done right: the AI tutors were told to teach, never to hand over answers. As I’ve written before, AI hurts learning when it does students’ homework and helps when it tutors them.
AI improves learning when it acts as a tutor and harms learning when it does the homework, because the AI can’t follow the student into the exam room. (GPT Image 2.5)
Aristotle for Everyone
Private tutoring has been wealthgated ever since Philip II of Macedon hired Aristotle to teach young Alexander in 343 B.C. A pity, because the average tutored student outperforms 98% of classroom students, as Benjamin Bloom found in 1984. Now everybody can afford a tutor. Even the cheapest AI tutor to match the humans (Gemma 4 31B) did it for 7 cents per session, versus $75 for a human hour: 1/918 the cost per point of learning.
Alexander the Great received personal tutoring from Aristotle. This cost his dad a pretty penny, but kings can afford such things. (GPT Image 2.5)
With AI, even a peasant boy gets a private tutor, which equalizes educational opportunity. (GPT Image 2.5)
The three best AI tutors were premium models: Opus 4.8 at extra-high reasoning (16.0 points), GPT-5.5 Pro (15.7), and Gemini 3.1 Pro at high reasoning (15.4), vs. 14.9 for the humans. Gemini 3.1 Pro beat all three Gemini Flash models, GPT-5.5 Pro beat regular GPT-5.5, and Opus 4.8 fell to 13.6 with reasoning off. (Differences between AI tutors weren’t statistically significant, so read this as a trend.) The newest models weren’t even tested, so the best AI tutors available today probably beat these scores.
For best learning, use frontier models at a high reasoning setting, and padlock the dial if you must. (GPT Image 2.5)
Cliché but true: today’s AI is the worst we’ll ever have. When superintelligence arrives in 2030, AI tutoring will be better still, and education will never be the same. (GPT Image 2.5)
Pay for the Better Tutor
If you’re a student or the parent of one, you must pay for an AI subscription. Free plans ration the premium models and deep reasoning that won this study, stranding users on the wrong side of the two-tier AI world I predicted. A standard plan costs less per month than one hour with a human tutor.
Advice to students and parents: pay for good AI. It’s the cheapest way to improve education. (GPT Image 2.5)
Alexander had Aristotle. Now every student can have one. Just don’t settle for the free-tier Aristotle.
Cutting corners with free AI is tempting, but the free tier is inferior. (GPT Image 2.5)
Every Icon Needs a Text Label
My old friend Vincent Flanders coined the perfect name for unlabeled icons in his classic Web Pages That Suck: mystery meat navigation. You can’t tell what’s under the silver dome until you lift it, and users refuse to lift 8 domes to find dinner. Almost 3 decades later, sadly, many sites still suck just as badly as when Vincent wrote his prescriptions, because minimalist fashion keeps stripping the labels off.
Words carry the meaning that icons only hint at. Except for a handful of universals (the magnifying glass, the shopping cart), an icon without a label is a riddle, and users don’t click riddles. Label everything. Vincent was right.
Tonight’s menu: 8 covered dishes, 0 descriptions. Bon appétit. (GPT Image 2)
Every AI Agent Needs a Real Stop Button
A visible, reliable Stop control makes delegation easier to trust. Users can let AI run the machinery and still halt it the moment circumstances change.
Clear stopping controls let users halt AI actions, preserve useful progress, and see exactly what still needs attention. (GPT Image 2)
The woman beside the giant approval stamp has the right idea. Once she spots an error, making her hunt through a menu for “Manage ongoing activity” would turn one fixable mistake into a whole production run of them.
Imagine an agent updating supplier records when the purchasing manager notices that two companies have been mixed up. Stop must block new actions instantly. The interface should then identify completed updates, untouched records, and anything still in flight. A reassuring “Stopped” message is worthless if the agent keeps changing data behind the user’s back.
Stopping should preserve both control and useful progress. Keep completed work available for inspection. Provide Pause for temporary interruptions, Stop for ending the task, and Undo wherever completed changes can be reversed. Stopping a stamp doesn’t erase its previous impressions.
In my article on user control for long AI tasks, I explain how preserving partial results makes course correction cheaper.
Test this interaction while a real task is running. Can users find the control, understand what has stopped, and determine what needs repair? A button that passes a visual inspection can still fail this test.
I’d judge an agent partly by how gracefully it accepts dismissal. Competent assistance includes knowing when to put the stamp down.
Make Leaving Easy
A subscriber who can’t find the exit is a hostage, and hostages don’t buy. Companies that bury the unsubscribe link under guilt trips, retention offers, and 6-step confirmation flows are optimizing a vanity metric (list size) at the expense of the metric that pays (goodwill). The trapped reader still leaves, only angrier: marking messages as spam, blocking the sender, and badmouthing the brand.
The platforms have run out of patience. In 2024, Gmail and Yahoo began requiring one-click unsubscribe from bulk senders, turning a courtesy into a condition of delivery. Treat the rule as a floor: put the link where eyes expect it, make it work in one click, and skip the goodbye interrogation.
The payoff comes from a paradox: an easy exit lowers the perceived risk of entering. People subscribe more willingly to lists they know they can escape.
A list that’s hard to leave gets left anyway, with a spam report as the parting gift. (GPT Image 2)
OODA Loop: Faster Decision Cycles Beat Smarter Competitors
The Observe–Orient–Decide–Act model holds that victory goes to whoever cycles through decisions faster on sound mental models. AI now compresses every stage of the loop, so executives should clock their loop lag this quarter and start shrinking it.
The OODA Loop rendered as architecture: an eye for observing, a compass rose for orienting, converging marbles for deciding, and climbing stairs for acting. In business, you walk the whole building before your competitor does, then walk it again. (GPT Image 2)
Definition: The OODA loop models decision-making under competition as a repeating cycle of 4 stages: Observe (collect signals), Orient (turn signals into an updated picture of reality), Decide (choose a response), and Act (execute, which changes the world and feeds the next Observe).
The competitive claim follows: operate inside an opponent’s loop, cycling faster on a better-matched picture of reality, and the opponent’s responses arrive too late to matter.
Born in a Cockpit Over Korea
Colonel John Boyd (1927–1997) was a US Air Force fighter pilot and later the Pentagon’s most productive irritant. His starting puzzle came from the Korean War: the MiG-15 could outclimb and outturn the American F-86 on paper, yet F-86 pilots won lopsidedly. Boyd’s explanation: the F-86’s bubble canopy let its pilot observe better, and its full hydraulic controls let him switch maneuvers faster, so the American ran more decision cycles per dogfight while the MiG pilot fell ever further behind events. As an instructor at the Fighter Weapons School, Boyd kept a standing bet that he could defeat any challenger in simulated air combat within 40 seconds; per Robert Coram’s 2002 biography, “Forty-Second Boyd” never paid out.
He published almost nothing. Apart from a 1960 air-combat manual, his written legacy is one dense 1976 essay, “Destruction and Creation” (7-page PDF), on why minds must keep dismantling stale mental models to build better-fitting ones. The rest lived in marathon briefings, notably Patterns of Conflict: roughly 196 slides delivered over 6 hours. (The Pentagon endured slide-deck warfare decades before the rest of us.) His only drawing of the loop appears in his final 1995 briefing, The Essence of Winning and Losing, and it surprises first-time viewers: Orient sits at the hub, feeding everything, with a direct wire from Orient to Act for well-trained responses. Boyd’s big O was always Orientation, the stockpile of models, experience, and culture through which observations acquire meaning.
Your Company Runs OODA Loops All Day
Strip away the gunsmoke, and the OODA loop describes knowledge work. An executive spots a competitor’s price cut (Observe), works out what it means for the quarter (Orient), picks a response (Decide), and ships it (Act): one full revolution. The elapsed time from signal to executed decision is your loop lag. Give two companies identical strategies but different loop lags, and the faster one pulls ahead: it runs more pricing experiments per year, kills bad projects sooner, and compounds small informational edges the way interest compounds money.
Usability is loop-lag engineering at the micro scale. My three response-time limits (0.1 seconds for perceived instantaneity, 1 second for uninterrupted flow of thought, 10 seconds before attention wanders) govern the thousands of miniature Observe–Orient cycles an analyst runs each day. A dashboard that takes 30 seconds to refresh costs far more than those 30 seconds: the wait evicts the very orientation the analyst was building. Bad information scent, cluttered reports, and buried alerts all bill the same account.
AI Compresses Every Stage, With One Landmine
Now map 2026-era AI onto the 4 stages:
Observe: Monitoring agents watch feeds, tickets, and telemetry around the clock and surface anomalies, so signals no longer wait for a human to go looking.
Orient: The largest win and the landmine. AI synthesis compresses a week of reading into 10 minutes. But a hallucinated synthesis corrupts orientation, and corrupted orientation makes the remaining stages fast and wrong. Demand source-linked summaries, and spot-check them.
Decide: Ask AI for explicit options with tradeoffs and confidence levels, and keep a decision log so future loops learn from past ones.
Act: Agentic AI executes the decision, and its activity ledger becomes tomorrow’s Observe.
Of course, the classic misuse is speed worship: accelerating Act while starving Orient. Boyd never preached raw speed; “Destruction and Creation” is entirely about keeping mental models matched to a changing reality. An organization that acts instantly on last year’s model of the market is a race car navigating by an outdated map. Dashboard glut is the same disease at the Observe stage: I’ve seen executives with 20 dashboards and no picture. Vanity metrics jam your own loop more reliably than any competitor could. Speed without orientation is just faster failure.
7 Guidelines for Tightening the Executive OODA Loop
Clock your loop lag. Pick one recurring decision class (pricing responses, incident calls, hiring approvals) and measure signal-to-execution time. Baseline it this quarter; shrink it next.
Consolidate Observe. Give each role a single briefing surface, with anomalies pushed to it. Prune vanity metrics on sight; each one taxes orientation.
Spend AI on Orient first. Buy synthesis with linked sources before buying autonomous action. A faster wrong picture is the worst purchase in enterprise software.
Keep decision tools inside the 1-second band. Analysts abandon lines of thought that stall, so sub-second response times are a strategic asset, even if the budget files them under infrastructure.
Sort decisions by reversibility. Reversible calls deserve a fast, delegated loop; irreversible ones deserve the full ceremony. Most organizations run one speed for both and get the worst of each.
Automate Act, and log it. Let agents execute approved decisions, with every action written to a ledger that feeds the next Observe stage. A loop without feedback is just a line.
Schedule destruction and creation. Once a quarter, name the mental model that served last year and test whether it still matches reality. Boyd’s 7 pages will outlast every strategy deck in your archive.
Alice and Zimo run the OODA Loop in Cel-Shaded Graphic-Novel Key Art style. (GPT Image 2)
A Session Timeout Shouldn’t Erase the Task
Users often stop touching your application while they keep working on the task: they consult a document, answer a colleague, or mull over a difficult answer. An application that treats this pause as permission to discard their work is making an expensive mistake about how people work.
Your answers were profound. Your session was brief. (GPT Image 2)
The mountain pilgrim has devoted his life to acquiring wisdom and a considerable afternoon to entering it. The timeout converts both investments into an empty form. Software has invented a new productivity metric: work successfully erased per session.
In Top 10 UI Annoyances, I criticize premature signouts and the loss of form data when users navigate away and return. These failures punish normal behavior. People shouldn’t have to perform a task in one perfect, uninterrupted take merely to keep the information they’ve already supplied.
Preserve the work across the interruption. Save drafts automatically, make their status visible, and restore the appropriate task state after users sign in again. Test three interruptions: consulting another page, losing the connection, and returning after a timeout. Check the actual content on return, including attachments. Security may require users to sign in again. Nothing requires them to reconstruct an afternoon’s work from memory.
Corporate Buzzwords Make You Look Less Intelligent
Corporate-ese is a dialect in which nothing happens. Synergies get leveraged, paradigms get disrupted, stakeholders get empowered, and the reader still doesn’t know what the product does or what will change on Tuesday. Daniel Oppenheimer at Princeton tested the popular belief (that big words make the writer sound smart) in a 2006 paper whose title, “Consequences of Erudite Vernacular Utilized Irrespective of Necessity,” is itself the joke (it won an Ig Nobel Prize). In three experiments, readers judged the authors of needlessly complex text as less intelligent. The effect ran through processing fluency: in a fourth experiment, merely printing a text in a hard-to-read font made its author seem dimmer.
Buzzwords add a second cost: they carry no information scent, so readers can’t tell whether the page is about their problem. The cure is concrete. Replace each abstraction with the thing it stands for: “optimize workflows” becomes “cut approval from 3 days to 1”; “empower stakeholders” becomes “let managers approve from their phones.” If nothing concrete survives the replacement, the phrase was decoration and can go.
Then read the result aloud, imagining a customer across the table: if you’d blush saying it, your customer will cringe reading it. Plain words are the only claims readers can check.
Every buzzword is a place where a fact should have been. (GPT Image 2)
Cut the Buyrocracy at Checkout
Government offices make you prove who you are before they’ll serve you. Many checkouts work the same way, smothering the sale in buyrocracy: paperwork the site demands before it will accept your money. The customer is holding a credit card; the merchant hands over a form. Baymard Institute’s latest survey finds that 18% of US online shoppers have abandoned an order because the site wanted an account, and its benchmark finds 44% of sites hiding the guest option well enough that users assume registration is required.
At ASOS, James Hart, then the ecommerce director, reported that removing every mention of registering from the first checkout page halved abandonment there, even though customers still set a password one page later. Users will happily type. What they flee is the announcement of a relationship they didn’t ask for. The customer proposed a transaction; the form proposed marriage. Each account is also a liability on both sides of the counter: the customer carries one more of the 120 personal passwords NordPass counts per person, and the merchant carries one more record to protect.
ASOS merely disguised the account. The better fix postpones it: let anyone buy as a guest, then offer account creation on the confirmation page, where the details are already in hand and the ask is one click. Baymard found 42% of sites still make the pitch before or during checkout. The AI shopping agents now buying on customers’ behalf take the shortest path: the store without a form. Sell first, register later; anything else is buyrocracy, and buyrocracy costs sales.
Paperwork first, payment second, customer gone. If you want people’s money, let them give it to you! (GPT Image 2)
Final Thought of the Day
