Insights

Notes from the studio

Essays and field notes on design, engineering and commerce, written between shipped projects and grounded in the work itself.

Subscribe by RSS

Hiring a Studio in Turkey from Germany: How the Working Model Actually Goes

Izmir and Berlin can work in the same daily rhythm. The real cross-border friction is in contracts, tax handling, review ownership and implicit language.

Read the note
Hiring a Studio in Turkey from Germany: How the Working Model Actually Goes

Where to Put the Human in the Loop, Precisely

A person who approves every action eventually approves nothing. Human review belongs at irreversible, ambiguous and high-consequence transitions.

Read the note
Where to Put the Human in the Loop, Precisely

RAG Is Not a Product

Retrieval makes a demo possible. Permissions, freshness, answer boundaries and operational ownership make an internal assistant useful after launch.

Read the note
RAG Is Not a Product

Evals Before Features: How to Know Your Agent Got Worse

A plausible answer is not a passing test. Golden sets, pairwise review and failure slices turn subjective agent quality into a release decision.

Read the note
Evals Before Features: How to Know Your Agent Got Worse

Agent Access Without Handing Over the Keys

An agent needs capabilities, not an administrator account. Read-only defaults, narrow tools, scoped tokens and an audit trail contain inevitable mistakes.

Read the note
Agent Access Without Handing Over the Keys

Returns, Subscriptions, and the Parts of Commerce Nobody Demos

The launch is only the visible edge of commerce. Returns, payment recovery, support tools and catalogue maintenance determine the real running cost.

Read the note
Returns, Subscriptions, and the Parts of Commerce Nobody Demos

Product Data Is the Real Moat in AI Commerce

AI cannot recommend what a catalogue has not described. Complete attributes, stable identities and explicit relationships matter more than polished copy.

Read the note
Product Data Is the Real Moat in AI Commerce

Checkout Is Where Trust Is Won: A Friction Audit for This Week

Checkout friction is usually visible without a rebuild. Follow real orders, record every surprise, and fix the point where the promise changes.

Read the note
Checkout Is Where Trust Is Won: A Friction Audit for This Week

We need an app usually means something else

Native is justified by background work, deep hardware access and operating-system extensions. Login, forms, camera and offline use are not enough by themselves.

Read the note
We need an app usually means something else

How to brief an AI automation so it can be quoted

A vague automation brief does not produce a flexible quote. It produces a padded one, because every missing decision becomes supplier risk.

Read the note
How to brief an AI automation so it can be quoted

When a chatbot is the wrong answer

If the system already knows the task, asking someone to explain it in chat adds work. Better search, remembered forms and silent workflows often win.

Read the note
When a chatbot is the wrong answer

The two-week AI pilot that produces a real decision

A useful pilot ends with measured evidence, known failure modes and a go or no-go decision. A polished demo is not enough.

Read the note
The two-week AI pilot that produces a real decision

Shopify or custom: four questions that decide the stack

Revenue and catalogue size do not decide the stack by themselves. The real thresholds are product shape, checkout rules, differentiation and ownership.

Read the note
Shopify or custom: four questions that decide the stack

What AI automation actually costs in 2026

The model bill is rarely the expensive part. Integration depth, exceptions, permissions and ongoing ownership determine the real budget.

Read the note
What AI automation actually costs in 2026

You can rank first and still be invisible to AI

83% of AI Overview citations point to pages outside the organic top 10, and ChatGPT and Perplexity agree on sources only 11% of the time. Ranking and citation have become two different games.

Read the note
You can rank first and still be invisible to AI

Gartner expects 40% of agent projects to be cancelled. The models are not the reason

More than 40% of agentic AI projects will be cancelled by the end of 2027, and every cause Gartner names is managerial: cost, unclear value, weak controls. Here is how to scope an agent that survives.

Read the note
Gartner expects 40% of agent projects to be cancelled. The models are not the reason

MCP just became boring, and that is the whole point

The 2026-07-28 specification is the largest revision since launch: a stateless core, server rendered interfaces and long running tasks. A protocol becoming unremarkable is what makes it safe to build on.

Read the note
MCP just became boring, and that is the whole point

Meta's WhatsApp agent is live everywhere. Here is when it is enough, and when it isn't

Over one million businesses already let Meta's Business Agent answer their customers on WhatsApp. The real question for a small business is not whether to use an AI agent, but which parts of the conversation to hand over.

Read the note
Meta's WhatsApp agent is live everywhere. Here is when it is enough, and when it isn't

Agentic commerce is winning discovery and losing checkout

OpenAI launched in-chat checkout in February and walked it back within weeks, while AI-referred traffic to retail sites grew 393%. The lesson: let AI handle discovery, keep the purchase on your own site.

Read the note
Agentic commerce is winning discovery and losing checkout

The EU AI Act in August 2026: what actually applies now

August 2, 2026 turned out to be a transparency deadline, not the day the whole AI Act switched on. Here is the revised calendar, the fines, and a practical checklist.

Read the note
The EU AI Act in August 2026: what actually applies now

SEO after AI Overviews: the playbook is citability

68% of searches now end without a click, yet branded clicks are up where AI cites you. What the numbers really say, and what to change on your site.

Read the note
SEO after AI Overviews: the playbook is citability

AI content now carries a watermark. Is your pipeline ready?

Anthropic will watermark text from its models, and the EU expects marking solutions to comply by December 2026. What this means for anyone running an AI content pipeline.

Read the note
AI content now carries a watermark. Is your pipeline ready?

An AI agent hacked a gym. What that teaches everyone deploying agents

Agents in sandbox tests reached real systems, and one deleted a stranger's gym booking. The guardrail patterns that separate useful agents from liabilities.

Read the note
An AI agent hacked a gym. What that teaches everyone deploying agents

When your next visitor is an agent

Cloudflare just shipped a browser that has no use for themes, tabs, or extensions, because its user is an AI agent. What that means for how you build your website.

Read the note
When your next visitor is an agent

Chat is becoming the storefront

A billion business conversations a day, retailers moving into ChatGPT and Gemini, and Meta charging by the token. The storefront is splitting into channels, and your site is only one of them.

Read the note
Chat is becoming the storefront

Your best visitors may be invisible: how to measure AI-referred traffic

AI-referred traffic to US retail sites grew 393% in a single year, and these visitors convert better than any other segment. Most analytics setups still file them under noise. Here is how to fix that.

Read the note
Your best visitors may be invisible: how to measure AI-referred traffic

Buy the tool or build the automation? A decision framework for small businesses

82% of small business employers have already invested in AI tools. The real question is no longer whether to adopt, but which problems deserve an off-the-shelf agent and which deserve automation built around your own data.

Read the note
Buy the tool or build the automation? A decision framework for small businesses

Headless Shopify in 2026

Headless commerce has left the hype cycle and entered the spreadsheet. A practical look at when decoupling Shopify pays off in 2026, and when it absolutely doesn't.

Read the note
Headless Shopify in 2026

Why motion design sells

Animation isn't decoration. Used with intent, motion is the most persuasive tool on a website. Here's the evidence, and the discipline behind it.

Read the note