The chapter I didn't plan

At the start of 2026 I found myself somewhere I never expected: inside a small, premium accessories brand. Not as an advisor. As the person who built its data and AI infrastructure from scratch. This series is the story of that build - the architecture, the mistakes, the lessons - and why I now believe small businesses, not enterprises, are where the real AI transformation is happening. This is the first of four posts.

At the end of 2025 I stepped down as COO of Sedna. Twenty-plus years, four company lifecycles - built, scaled, sold, and everything in between. The plan, insofar as there was one, was the conventional next move: take a breath, then find the next scale-up, the next exec seat, the next board deck.

Then life did what it does. Rachel - my wife, and the founder of String Ting, the accessories brand she built in London - needed support. The business was doing what successful small businesses do: growing faster than its operations. Wristlet phone straps and bag charms, handmade, sold direct-to-consumer through Shopify and through wholesale accounts from Tokyo to New York - and behind it, one person carrying an operational load built for ten.

So I stepped in. Not a sabbatical project, not a strategic career move I can dress up in hindsight - a husband with suddenly-free hands and a relevant skill set, showing up for the family business. No operations team. No engineering team. No data team. No IT department.

I know how it reads on a CV - which is partly why I’m writing this series. Because what started as stepping in to help turned into the most concentrated product education of my career, and what I built there is, I’d argue, a working prototype of where every company under fifty people is heading. The unplanned chapters, it turns out, are where the real learning hides.

The thesis: small business is the frontier, not the afterthought

In enterprise SaaS, we talk endlessly about AI transformation. We run pilots, we form steering committees, we write governance frameworks. Meanwhile, the actual transformation - the kind where AI changes what a single person can operate - is happening somewhere else entirely.

Here’s the thing about a premium DTC brand doing global volume with two people on the business side: every operational gap is existential. There is no analyst to pull the numbers. There is no ops team to reconcile the wholesale ledger against the DTC storefront. There is no marketing department to segment customers. Either the founders do it - at night, badly - or it doesn’t happen.

That constraint makes a small business the purest possible test environment for a question I’d been circling for years at Sedna, and which I wrote about in From apps to infrastructure: when the models are commoditised, where does the value actually live?

The answer I found: in the context layer. In the plumbing. In the unglamorous, deeply specific work of making a business legible to a machine.

What String Ting actually looked like from inside

From the outside: a beautiful brand with a global customer base - the US, UK, Japan, Hong Kong, Singapore, South Korea and beyond - and product in some of the best boutiques in the world.

From the inside, the operational picture was the same one I’ve seen in every SMB I’ve ever encountered, just compressed: Shopify held the DTC truth. Zedonk (a fashion-industry ERP) held the wholesale truth. Xero held the financial truth. Klaviyo held the customer-communication truth. GA4 held the traffic truth (as much as you can trust GA4). None of them agreed with each other, and none of them talked to each other. The same physical product existed as a Shopify variant, a Zedonk style, a barcode, and a line on a supplier invoice - with no reliable join between any of them. They all have their own definitions and formulas. None of them have the context that makes them reliable.

Every question that mattered - what’s actually selling, where, to whom, at what margin, and what should we make more of? - required a human to swivel-chair between five systems and hold the join logic in their head.

Sound familiar? It should. Strip away the scale and this is exactly the fragmentation problem enterprises pay millions to solve. At Sedna we called the enterprise version of this problem ShippingOS. The SMB version has no name because nobody builds for it - the deal sizes are too small for enterprise vendors, and the problem is too gnarly for plug-and-play tools.

The role I actually played

My brief, formally, was operations. My actual job, as it turned out, was to be a one-person platform team: I designed and built a data warehouse on Google Cloud - BigQuery as the spine, with every system in the business flowing into it. Real-time order data via webhooks, reconciled nightly against the source of truth. Product, inventory, marketing, traffic and wholesale data alongside it.

On top of that foundation, I built an AI layer using Claude - not a chatbot bolted onto the website, but an operating assistant with the business’s full context baked in: its data model, its conventions, its product taxonomy, its known data quirks. The founders can ask it the questions they used to lose evenings to, and it answers from the warehouse, correctly, in seconds. It also serves to automate and streamline workflows that push updates to the ecosystem of services the business relies on.

The next two posts go deep on each layer - the data foundation and the AI system respectively. The final post is about what it all means, for SMBs and for the SaaS companies that sell to them.

Why this was harder - and more instructive - than enterprise

A confession: I assumed this would be a scaled-down version of work I’d done before. It wasn’t. It was a different sport.

No abstraction survives contact with a real small business. In enterprise you can hide behind the integration partner, the data team, the six-month rollout. Here, if a price in a market catalogue didn’t match the canonical price list on Shopify, that was my problem, that afternoon, with no one to escalate to. It’s been a while since I was at the coal-face in this way.

The feedback loop is brutal and wonderful. Ship something useful and the founders use it that same day. Ship something confusing and it’s abandoned by lunchtime. There is no adoption curve to hide behind, no change-management programme. The product either earns its place in the workflow or it dies. Every product leader should experience this new world.

Trust is the entire product. When the person asking “how did we do last month?” is going to make a real decision - a reorder, a price change, a campaign - off the answer, a plausible-but-wrong number is worse than no number. Most of my engineering effort went not into capability but into correctness: reconciliation jobs, data-quality flags, the assistant knowing what it doesn’t know. This, more than anything, is what I think the AI industry still gets wrong.

The most valuable thing I built wasn’t the AI. It was the version of the business the AI could understand.

The chapter closes

That build is now done and running. Rachel and the business don’t need a platform engineer full-time - which was always the point. Infrastructure, done properly, is the thing you stop noticing.

So this chapter is closing, and I’m starting the search for my next role - back in SaaS, building products and teams, but carrying something I didn’t have before: I’ve now personally lived the full loop of the thing we all sell. I’ve been the vendor, the builder, the buyer, and the end user of an AI-native operating stack - in the same six months.

The next post gets into the concrete architecture: what I built on Google Cloud and BigQuery, the real-time-versus-batch decisions, and why the least glamorous part of the stack - matching products across systems - taught me the most.

If you’re building AI for real businesses, or you’re a leader wondering what an AI-native operating model actually looks like below the enterprise tier, this series is for you.