Work · Surinder Thakur

Dutch B2B commerce · Blueprint prototype, 2 weeks
Agents take the order. People keep the decisions.
An urgent order in minutes, not an hour, Modeled, from the prototype
Try the prototype (opens in a new tab)
One urgent order, from call to ERP.
The buyer calls and says the order
The agent asks 1 question, never a form
It checks live stock and the contract price
Over €300, the workshop owner approves
Written to the ERP, with 10 minutes to undo
From the prototype, with mock ERP data.
Delta Parts order line
Answered by an AI ordering agent, day and night
Voice line
WhatsApp
Email
Delta PartsOrdering agent, on your call
The water pump on the Golf in bay 2 has gone. I need one today, and the timing belt kit while it is open.
Golf, plate K‑482‑GH
Water pump1
Timing belt kit1
2 water pumps fit this engine. Is the impeller metal or plastic?
MetalPlastic
Metal
Checking the main warehouse, the nearest branch and equivalent parts.
Main warehouseStock and contract price
Nearest branchStock and van times
Equivalent partsCross-reference, same catalogue
It’s over your limit, so your agent is asking the workshop owner.
Sent to the workshop owner for approval, €325.50
The workshop owner approved €325.50.No agent ever approves.
Water pump and timing belt kit, €325.50, at the workshop by 10:45.
Written to the ERP
10:00 left to undo9:59 left to undo9:58 left to undo
Undo
ListeningThinkingAgent speakingListeningChecking the ERPWaiting for approvalOrdered
Buyer’s agentKeeps your €300 limit
Workshop lead needs 2 parts for the Golf in bay 2. €325.50, over the €300 limit.
€325.50
Over the €300 limit your company set. Stock stays held for 2 hours.
Approve ApprovedDecline
Approved €325.50
An urgent order in minutes, not an hour.
The buyer orders by voice, WhatsApp or email
Sales reps keep the hard calls and the relationship
Brakes, steering and airbags always go to a person
The urgent order today
About an hour
First call to order placed, mostly waiting.
The buyer phones the merchant
A sales rep reads out part numbers
The sales rep checks stock by hand
The workshop owner signs off from somewhere else
The sales rep types the order in by email
With the agents
Minutes
First message to order placed.
The ordering agent answers at once
The ordering agent asks 1 short question
The ordering agent reads live stock and price
The workshop owner approves in 1 tap, only over the limit
The fulfilment agent writes the order to the ERP
Both times are assumptions. We will time 30 to 50 phone orders before day 1.
Each agent has 1 job. People approve rules and exceptions.
An agent orders alone only when all 3 checks pass.
The buyer’s 8 steps. People stay in control at 3.
The merchant’s side: every gap found before the first order.
“Your agentic prototype was strong: the buyer and merchant flows were coherent… your human-in-the-loop approach to trust stood out.”
Product teamDutch B2B commerce platform
Buy SaaS when
The job is common, like payroll or email
Your process can bend to the tool
A monthly fee beats paying for a build
Build when
The work is how you are different
People bridge gaps by hand every week
Approvals and exceptions are yours alone
2 questions decide it: is the job common, and who fills the gaps today.
Building an AI toolShould we build a custom AI tool or buy SaaS?Buy SaaS when your process matches how the tool works. Build when the work is yours alone: your approvals, your exceptions, your mix of systems. Then watch the gaps. If people copy data between tools every week, those hours are the budget. A custom tool earns its cost when it removes them.
Built
Stalls when it fits no real job
Used every week
Stalls when it lives in its own tab
Trusted to act
Stalls when nobody sees what it did
Trusted with money
Stalls on the first wrong number
In real use
Each stall designed out
The 4 places an AI pilot stalls, in the order a tool meets them.
Getting your team to use itWhy do AI pilots stall before real use?AI pilots stall because a demo only proves the tool can work. Real use needs a team to choose it every week. Pilots stall in 4 places. The build never fit a real job. The team keeps the old spreadsheet. Nobody trusts the agent to act. 1 wrong number in a money flow ends it.
Weekly use
Week 1Week 8
In its own tab
Inside the daily task
Novelty fades after week 1. A feature inside the work keeps its users.An illustration of the shape, not measured data.
Getting your team to use itWhy do users stop using AI features after the first week?Because the first week runs on novelty, and novelty is not a job. The demo moment amazes, then the blank prompt box starts asking the user to work. A feature survives when it attaches to something the user already does every week. Most AI features live in their own tab, waiting. Habits do not form in tabs nobody opens.
In the path
Starts where the task starts
First step filled
The agent drafts, a person checks
Work shown
Every step and its rule
Easy to correct
1 click, and it remembers
Weekly use
Tracked after launch
5 design moves that turn a launch into weekly use.
Getting your team to use itHow do you get a team to actually use a new AI tool?Design the first week of use as carefully as the tool. Put it where the task already starts, so nobody has to remember it. Fill the first step in for them. Show what the agent did and let people correct it. Then track weekly use after launch, and fix what people skip.
AI SaaS · Built solo, live at proux.design
Agents do the research. Every answer cites its source.
Every answer cites its source, across 350+ UX principles
Try ProUX (opens in a new tab)
Not a chatbot. You set a task and ProUX runs it.
Experts with curated skillsEach expert works from its own skill file
Your project contextGoals, users and rules, read on every task
Guideline libraryVerified UX guidelines, retrieved by RAG
Your screenMeasured from pixels, not described
The ProUX methodThe steps every task runs through
ProUX engine
The right expertAutoChosen for the question, with its curated skills
The right modelAutoPicked for the job, yours to overrule
Informed design decisionsEach one cited, with the reasoning behind it
Your product gets the same engine. Your rules, your data, decisions it can defend.
The ProUX engine reads 5 sources:
Experts with curated skills.
Your project context.
Guideline library.
Your screen.
The ProUX method.
It returns the right expert, the right model and cited design decisions.
You set the task. It picks the expert and the model.
Every choice named on screen
One click changes it
The task: Audit this product page. Buyers drop before Add to cart. What do I fix first?
Expert: Design audit. You asked for an audit.
Model: Auto · Sonnet 5. Deep, picked by the job.
Grounding: 3 project documents. Your project first.
Then it runs: Analysing the design, Searching ProUX guidelines, Mapping findings.
It used 16 guideline sources.
Every claim links to its source.
1 tap opens the guideline and its evidence
When no source covers it, it says so
Thinking partner · Auto
Darken the price text. It measures 2.43:1 against 4.5:1. S1The Invisible Conversion Killer: Text Legibility in eCommerceGuideline #411Why it applies, real examples, and how to check it on your page.Click to open in side panel ↗Collapse long specs, so Add to cart stays in view. S2Assumption: no project document says who buys.Thinking partner · Auto · 2 sourcesGrounded in 3 documents
Darken the price text: it measures 2.43:1 against 4.5:1.
Source 1: guideline 411, text legibility in eCommerce.
Collapse long specs, so Add to cart stays in view.
Source 2: guideline 32, collapse sections.
Assumption: no project document says who buys.
It measures the screen. It does not guess.
ProUX measures a product page from its pixels.
Price text: 2.43:1, needs 4.5:1.
Add to cart: 36pt, needs 44pt.
Body text: 16px, passes.
The score is 64, at risk: 100 less 24, 8 and 4.
Price text is fixed at 5.20:1. Add to cart is fixed at 48pt.
With 2 criticals fixed, the score moves from 64 to 80.
Same task. A different kind of result.
The same task, answered by a general LLM and by ProUX.
LLM: “Looks light”. ProUX: 2.43:1 against 4.5:1, Measured from pixels.
LLM: “Best practice”. ProUX: Guideline #411, 1 tap away.
LLM: 8 equal tips. ProUX: Ranked fixes, Effort on each.
LLM: A new answer. ProUX: Same rules, Every page.
LLM: Start over. ProUX: Re-measured, Fixed. Now 5.20:1.
- AI apps, annual plans21.1%
- Other apps, annual plans30.7%
- AI apps, monthly plans6.1%
- Other apps, monthly plans9.5%
Subscribers still paying after 12 months, AI apps against the rest.Source: RevenueCat, State of Subscription Apps 2026.
Getting your team to use itWhy do AI apps lose subscribers faster than normal apps?Because the wow moment is not a habit. RevenueCat's 2026 data puts AI app annual retention at 21.1%, against 30.7% for everything else. The demo converts, then daily life asks a harder question: is this still worth opening. Most AI apps never design for that question. They monetize the novelty and lose the month 2 user.
Invoice arrives
The agent reads it
Rules check
Payee, limit, budget
Outside the rules
A person approves
Sees what, why, and the rule that stopped it
10 minute wait
Anyone can undo it
Paid
Logged with its rule
The money rule. Inside your rules it pays, with a 10 minute undo. Outside them, a person decides.
Trusting AI with moneyHow do you stop an AI agent making a mistake with money?You limit what it can do alone. The agent pays only inside the rules you set: approved payees, a limit, a monthly budget. Each payment waits 10 minutes before it goes out, so you can undo it. Anything outside the rules waits for a person. Every step is logged, with the rule it followed.
- How much a gain weighs1x
- How much an equal loss weighsAbout 2x
People weigh a loss about twice as heavily as an equal gain.Source: Tversky and Kahneman, 1992. Their estimate is 2.25.
Trusting AI with moneyWhy does 1 wrong AI answer break trust in my product?Because trust in software is asymmetric. 10 right answers build it slowly. 1 confident wrong answer spends it all at once. Loss aversion runs the math. A wrong answer is a loss, and losses weigh about twice an equal gain. The fix is not a better model. It is designing the recovery, the moment after the mistake.
ID uploaded
Checked for glare and blur
Agent reads it
Name, number, expiry filled
A mismatch or an expired ID
Your team reviews
With what the agent found, and why
Customer confirms
No typing
First transaction
Where onboarding ends
The agent does the typing. People only review the flags.
Trusting AI with moneyHow do you fix drop-off in fintech onboarding and KYC?People quit KYC when a step asks for trust it has not earned. Compliance sets the rules, so the fix is the order, the reasons and the effort. Say why before you ask. Let an agent read each document and fill the form. The customer only checks what it flags. Let people stop and resume.
Puffy · design manager, 2 years
Repositioned as attainable luxury.
+67% YoY revenue, Measured year on year, validated by A/B test
Now it looks worth the price.
Emotion: dark, warm, muted gold reads as luxury.
Social proof: 4 named awards, then 5 stars from 12,000+ sleepers.
Zero price: $240 of gifts, free.
+67% YoY revenue. Measured year on year, validated by A/B test
Checkout cut from 3 steps to 2.
Checkout before
3 steps
A big coupon box mid‑flow sent shoppers off to hunt for codes.
Step 1
Coupon?
Step 2
Step 3
Checkout after
2 steps
Express wallets first. The summary stays in view. The coupon waits for review.
SummaryExpress
Step 1
SummaryCoupon
Step 2
+21%checkout conversion on mobile
Measured by A/B test
−14%cart abandonment
Measured on the live checkout
Mobile checkout, step 1 of 2.
Anchoring: $3,438 struck through.
Trust: the total stays pinned, so nothing surprises at pay.
Zero price: each gift shows its full price, then free.
Every model, side by side.
FullStory and user tests: shoppers left after the hero, unable to compare models.
Decision fatigue: the facts line up.
Effort: 1 size picker for all 4.
Ads that look like the store.
Message match: the ad's dark, warm look is the store's.
Benefit first: each ad leads with what the sleeper gets.
2 pillars: materials and comfort.
On the store
On the store
PuffySponsored
PuffySponsored
PuffySponsored
PuffySponsored
PuffySponsored
The first fold, tested every month.
1M+ sessions, a new direction each month
AEscape to luxury sleep
BBuilt for self-care
CAward-winning comfort
Live behaviour shows what works. Each month, the learnings shape the next first fold.
Find the friction, test the fix, keep the winner.
Find the friction
Competitor scan
Luxury rivals: dark and story‑led. Puffy: loud.
FullStory and user tests
Shoppers left after the hero, unable to compare models.
Funnel audit
3 checkout steps, an oversized coupon box: last‑mile doubt.
Write a hypothesis
1 hypothesis, from all 3
Add a premium look, side‑by‑side models and a 2‑step checkout. Shoppers will see more value and buy more.
Design the test
Dark‑luxe product page
Comparison cards
2‑step checkout
Run the A/B test
On the checkout
A · 3 steps
B · 2 steps
+21%checkout conversion on mobile
It becomes the control
The 2‑step checkout goes live for everyone.
Live: −14% cart abandonment.
Then again. 30+ A/B tests in 2 years, mostly on product pages.
“Designs implemented have had positive impact on user experience and bottom line growth.”
Arthur AndreasyanFounder, Puffy
Google reads
The rendered page
Structured data, for rich results
Headings, links and speed
AI assistants read
The visible text, quoted as written
The first 2 or 3 sentences under a question
llms.txt, the plain summary
2 kinds of reader. Both reward a clear answer in plain words.
Being found by AIHow do ChatGPT, Claude, Perplexity and Google read a product site?They read the words people can see. Google crawls, renders and ranks, and uses structured data for rich results. AI assistants like ChatGPT, Claude and Perplexity fetch the page text and quote it. Most never see code hidden in scripts. So the answer has to live in the visible copy, worded the same on every page.
Agents need structure
Clean product data
Clear prices and stock
Policies a machine can read
People need trust
Proof near the price
A way back from a mistake
A checkout built for thumbs
2 funnels, 1 store. The same discipline runs both.
Being found by AIWill AI shopping agents kill my store's conversion?Not kill. Reroute. AI agents already comparison-shop, fill carts and complete checkouts. The buyers behind them still decide who earns the sale. 2 funnels now matter: people who need trust and agents who need structure. Stores designed for both convert both. Stores designed for neither lose quietly to whichever competitor answered the agent first.

Work and insights

The board

Agents take the order. People keep the decisions.

One urgent order, from call to ERP.

An urgent order in minutes, not an hour.

Each agent has 1 job. People approve rules and exceptions.

An agent orders alone only when all 3 checks pass.

The buyer’s 8 steps. People stay in control at 3.

The merchant’s side: every gap found before the first order.

Agents do the research. Every answer cites its source.

Not a chatbot. You set a task and ProUX runs it.

You set the task. It picks the expert and the model.

Every claim links to its source.

It measures the screen. It does not guess.

Same task. A different kind of result.

Repositioned as attainable luxury.

Now it looks worth the price.

Checkout cut from 3 steps to 2.

Mobile checkout, step 1 of 2.

Every model, side by side.

Ads that look like the store.

The first fold, tested every month.

Find the friction, test the fix, keep the winner.