Blog
Things we've shipped — and how.
Practical AI engineering. Architecture, guardrails, eval loops, and the production gotchas no one writes about. Filter by tag below.
- 10 min readengineering · strategy
Your Data Isn’t Ready for AI Yet, and That’s the First Thing to Fix
Most AI projects that stall don't stall on the model. They stall in week three, when the team discovers the CRM has four versions of the same customer, the policy documents contradict each other, half the product catalogue is missing the fields the agent needs, and the real answers live in someone's inbox. AI doesn't fix messy data; it repeats it confidently, at scale. Here are the six data problems that break AI projects, what "AI-ready" actually means, a two-week audit you can run before you build, and which cleanup AI can do for you.
Read → - 12 min readagents · engineering
MCP Explained for Business Owners: How AI Agents Finally Plug Into Your CRM, ERP, and Helpdesk
For two years, every useful AI agent came with a hidden line item: custom integration work for each system it had to touch. Model Context Protocol, or MCP, is changing that. It's an open standard that lets an AI agent connect to your CRM, helpdesk, files, and internal tools through one common interface, the way USB-C replaced a drawer full of chargers. Here's what MCP actually is in plain terms, what it changes for a business buying AI, what it doesn't fix, the security questions it raises, and how to decide when an off-the-shelf connector is enough and when you still need custom work.
Read → - 11 min readagents · strategy
When Your AI Agent Should Hand Off to a Human, and Why Most Handoffs Lose the Customer
Most AI support projects are judged by how many conversations the agent handles alone. Customers judge them by something else: what happened the moment the agent couldn't help. A loop that won't let them out, a transfer that makes them repeat everything, a human who starts from zero — that's the part they remember and the part they review. Here's why handoffs fail mechanically, the five signals that should trigger one, what a good handoff carries across, and the metrics that tell you whether your agent is actually resolving problems or just containing them.
Read → - 11 min readagents · engineering
Your AI Agent Can Read Email and Issue Refunds. Who Else Can Give It Instructions?
A chatbot that says something wrong is an embarrassment. An agent that does something wrong is an incident. The moment your AI can read inbound email, open attachments, browse the web, and call tools that issue refunds or update records, anyone who can put text in front of it can try to give it orders — and the model can't reliably tell your instructions from theirs. Here's how prompt injection actually works against business agents, the design test every agent should pass, and the six defence layers that keep a stranger's email from running your operations.
Read → - 10 min readagents · engineering
Why Your AI Bill Tripled After Launch: The Cost Levers Nobody Prices Into the Proposal
The proof of concept cost pennies a conversation. Ninety days after launch, the invoice is three times the forecast — at the same traffic. Nothing broke; the system simply started behaving like production: longer prompts, fatter context, agents that take extra steps, and a flagship model answering questions a small one could handle. Here are the six leaks behind almost every AI cost overrun, the six levers that close them, and the pre-launch review that keeps the first invoice from starting an argument.
Read → - 10 min readaeo · ecommerce
Review Management in the AI Era: Why Your Trustpilot Profile Is Now an AI Training Input
New research across 800,000 AI responses found brands with no active review profile get cited in ~1% of AI answers. Brands with 80+ reviews and regular responses: 75%. Your review pages stopped being a reputation afterthought — they're now the raw material AI uses to describe, recommend, or ignore you. Here's the mechanics and the playbook.
Read → - 11 min readagents · engineering
The Staging Environment for AI: Testing Model Upgrades Before They Surprise Production
Model providers ship updates faster than your release cycle — and every upgrade quietly shifts behavior your users depend on. 'Just bump the model string' is the AI era's 'just deploy to prod on Friday.' Here's the staging discipline that makes upgrades boring: pinned versions, golden-set regression, shadow traffic, canary rollout, and the rollback plan you write before you need it.
Read → - 10 min readecommerce · agents
Shopify's Latest AI Push: What Merchants Should Adopt Now vs Watch
The Spring '26 Edition shipped 150+ updates with one unmistakable center: AI on every admin screen. Sidekick everywhere (usage up 4× YoY), Pulse turning the Admin home into a recommendation feed, Campaign Autopilot in early access, an AI sales associate in Inbox, and Agentic Storefronts pointing at where this all goes. Here's the working merchant's sort: adopt now, adopt with guardrails, and watch — with the honest limits the launch posts skip.
Read → - 10 min readaeo · seo
Voice Search Didn't Die, It Moved Into AI Assistants: Optimizing for Spoken Queries in 2026
The 2018 voice-search hype died because assistants were dumb. Then the assistants got LLM brains: ChatGPT voice crossed 100M monthly users, Gemini replaced Google Assistant, Alexa+ went generative — and 8.4 billion voice devices now route queries through the same answer engines you're already optimizing for. Voice SEO and AEO just collapsed into one discipline. Here's what's genuinely different about spoken queries, and the checklist.
Read →
What we write about, and what we don't
Everything here comes out of engagement work rather than a content calendar, which shapes both the subjects and the tone.
- Numbers with baselines attached
- Where a post quotes a result, the before and after are both there. A percentage improvement with no starting point is unfalsifiable, and the AI industry currently runs on a great deal of it.
- The parts that were harder than expected
- Posts about builds include what went wrong — the retrieval that worked on curated documents and fell over on the real knowledge base, the eval case nobody thought to write. Those details are the reusable part; the happy path rarely is.
- Written for the person who has to fund it
- Most of these sit between a purely technical write-up and a business case, because that's where the decisions actually get made. Enough architecture to judge whether the approach is sound, enough cost and timeline to judge whether it's worth doing.
- No model-release commentary
- We don't publish reactions to every model launch. That writing ages badly and it's not what we know better than anyone else — what we know is what happens when these systems meet a real business process for six months.
