Articles
Articles from the workbench.
Longer pieces on AI infrastructure, custom software, and keeping what you build honest.
Field notes
56 articles · Updated Sep 9, 2026Nine Shapes Do Ninety Percent of the Work
Business Process Model and Notation is a standard, has an ISO number, and contains 116 elements, which is roughly 107 more than anyone needs. The working subset fits on one page and takes about ten minutes to learn. It is the difference between a drawing your accountant can read and a photo of a whiteboard.
Read 055You Cannot Automate a Process You Cannot Draw
Every business runs two versions of every process: the one in the handbook and the one that actually happens. The second one contains all the money. Drawing it takes an afternoon, costs nothing, and reliably finds things that no software purchase would have fixed.
Read 053Give It the Two Paragraphs, Not the Fifty Pages
Three separate arguments on this site rest on the same idea and never name it together. It is the reason one system costs ten times another, the reason a step that worked in testing degrades in production, and the reason the answer changes when nothing about the question did. What the model is handed decides more than which model it is.
Read 054The Map the Agent Runs Inside
Agentic systems are sold on their flexibility, which is also the thing that makes them hard to trust. The fix is older than the technology: a diagram of the process that exists separately from the agent, marking which steps are fixed and which ones are the agent's to decide. The notation for drawing it has an ISO number and turns thirty this decade.
Read 052What Keeping the Model Swappable Actually Takes
Three articles on this site end on the same advice: put the model behind your own interface so switching costs a day rather than a quarter. It is right, and stated that briefly it is misleading, because the model is the one part that ports easily. Everything built around it is where the lock-in actually lives, and this is the piece that says what that costs.
Read 051What the Agent Is Allowed to Touch
Giving an AI agent one credential that reaches everything is faster, it works, and it is how most of them get built. It also concentrates more access in one place than any employee has ever held, granted to something that takes instructions from whatever text it happens to read. This is the buyer's version of scoping that access, written without assuming you run the servers.
Read 050How Long Are You Keeping That, and Why
Where your data physically sits gets all the attention. The obligations that actually bite are quieter: what you told people you would use their information for, whether that covers what you are now doing, and how long you are keeping it. Every one of those is a design decision in an AI build, and every one of them is cheaper to make at the start.
Read 049The System Got Worse and Nobody Told You
Software either works or throws an error. An AI system has a third state: still running, still confident, quietly worse than it was. Nothing broke, nobody changed anything, and the accuracy that was nineteen in twenty in March is closer to seventeen in twenty by September. This is the piece about what actually drifts, and how you find out before a customer does.
Read 048The Software You Already Run Decides What Is Possible
The largest cost in most AI builds is neither the model nor the AI work. It is connecting the thing to the software the business already runs on, which was never designed to be connected to anything. This is the piece about what that actually involves, what to do when the vendor has no interface, and how to tell a real integration from a promised one before you sign.
Read 047Thirty Cases and an Honest Grade
Two articles on this site say the same thing in two sentences and move on: take thirty real examples, work out what the right answer is, and check the system against them. It is the single highest-value hour in an AI project and almost nobody spends it. This is the piece that says how, including the awkward parts, like what you do when there is no single right answer.
Read 046How an AI Answer Earns a Citation
The single most common question a business owner has about AI is how to stop it making things up. There is a real answer, it has a name, and it is the architecture underneath every system on this site that promises a citation you can click. It is also where the failure modes live, because a system that retrieves the wrong document answers wrongly with perfect confidence.
Read 045Nobody Owns the Exceptions
A system that handles ninety percent of the work hands back ten percent. That ten percent is not a rounding error, it is a new job, and it is the one nobody costs, staffs or designs. Projects rarely fail because the AI was bad. They fail in month three because the pile of things it could not do had no owner and no route back in.
Read 044Where the Human Checkpoint Actually Goes
Every honest account of business AI says the same thing: keep a person on anything consequential. It is correct and it is not yet a design. Put the checkpoint everywhere and it becomes a rubber stamp within a month. Put it nowhere and the first expensive mistake ends the project. This is the piece about where it goes, and what it costs when it is in the wrong place.
Read 043Judging the Decision Before You Know the Outcome
Every significant business decision gets judged twice: once when you make it, and once when everyone finds out how it turned out. Those are different judgements, and confusing them is the most reliable way to learn the wrong lesson from a good year. AI can genuinely help with the first one. The evidence also says it makes the second one worse under specific, predictable conditions. Here is where the line falls.
Read 042The Depth of Agentic Orchestration
One agent finishes a task. Orchestration runs the sequence of them, and decides what needs a person. That sounds like a small step up from a single agent, and it is the point where most of the difficulty in applied AI actually lives. Error compounds, state has to survive between steps, and the system has to know what it did when something fails halfway through. Here is the honest depth of it.
Read 041Twelve Rooms, Four Seasons
Vancouver Island accommodation runs 88% occupancy in August and 41% in January. The same owner does the same jobs in both months, with a full crew in one and almost nobody in the other. The useful question is which of those jobs still has to be done by a person, and the answer has changed in the last two years.
Read 040Four in Ten Canadian Businesses Say AI Is Not Relevant to Them
Ask Canadian businesses about AI and the most common obstacle is neither cost nor risk. Four in ten say it is not relevant to the business, and among the smallest firms it is higher still. That is a reasonable conclusion from what most people have been shown, and it is wrong for a specific reason worth understanding, because the version of AI being pictured is not the version that pays.
Read 039The AI Rules a BC Business Actually Has to Follow
Canada has no artificial intelligence statute. The bill that would have created one died in 2025 and has not come back. That is a genuinely useful fact, and it is routinely misread as meaning a business can do whatever it likes, which is the opposite of true. The rules that bind a BC company using AI arrive through privacy law, and in May 2026 four commissioners including this province's showed they intend to use them. Verified as of 18 August 2026.
Read 038Document Intelligence and Spreadsheet Intelligence: The Business AI That Pays Off First
Every business already runs on two things: a pile of documents and a stack of spreadsheets. That is exactly where the most useful, most checkable AI lives. Document intelligence reads the paperwork and turns it into data you can trust. Spreadsheet intelligence makes the sheets you run on readable, answerable, and safe. This is the business AI we focus on, and here is what it actually means.
Read 037You're Recording Every Call Now. Here's What That's Worth.
The recording and the transcript are close to free now. Most businesses let them pile up unread. The value was never the transcript itself, it is what you do with the text, and the range of what you can do is a good deal wider than a summary in your inbox.
Read 036The Three Kinds of AI, and Which One Moves Your Numbers
You have used the chat window. It is the smallest part of what AI can do. Here is a way to think about the rest that holds up after the novelty wears off.
Read 035They Loved the Demo. They Went Back to the Spreadsheet.
The most common way an AI project fails is not a technical failure. The software works, the training happened, everyone nodded. Six weeks later the old spreadsheet is quietly running alongside it, and nobody will tell you why. The reasons are well documented, and none of them are about the technology.
Read 034The Model Costs Pennies. The Rest Costs a Salary.
Ask what an AI system costs and you will get a number about tokens. That number is real, and it is the smallest one on the page. The expensive parts are the integration nobody scoped, the data work nobody wanted, the people who have to change how they work, and the quiet annual cost of keeping the thing alive.
Read 033Garbage In, *Gospel* Out
The old warning was garbage in, garbage out, and it was easy to obey because garbage looked like garbage. The new problem is worse. Feed a machine a flawed record of how your business has always decided things, and what comes back is not obvious rubbish. It is fluent, confident, consistent, and wrong in exactly the way you have always been wrong.
Read 032There Is an Unapproved Employee at Your Company. It Costs $20 a Month.
Nobody hired it, nobody vetted it, and it has read your client files. Half of Canadian employees now use AI at work, most of them on tools their employer never approved, and the difference between the twenty-dollar seat and the business seat is not the model. It is who is on the hook.
Read 031Everybody Is Talking About AI. Almost Nobody Is Building It.
There has never been more conversation about AI and there has never been a wider gap between the conversation and anything actually running in a business. The talking is not a prelude to the building. For most companies it has quietly become a substitute for it.
Read 030Vertex, Bedrock, Microsoft, or Build Your Own: Where Should Your AI Actually Live?
Every AI build starts with a question that sounds like a technology decision and is not. Google, Amazon and Microsoft will all sell you a platform, and a Canadian region to run it in. What almost nobody tells you is that on two of the three, choosing a Canadian region does not keep your inference in Canada.
Read 029What's Going On in the World of AI: An Unbiased View
The AI story is told either as a miracle or a bubble, and both versions are being sold to you by someone with a position. Here is the plain account: what actually happened between the major players, what the numbers really say, and, just as importantly, where the honest answer is that nobody knows.
Read 028AI Doesn't Pay for Itself, Unless the Job You Point It At Does
Two businesses can buy the same AI, from the same competent firm, built to the same standard, and get opposite results. One is paid back inside a year. The other quietly writes it off. The difference was settled before a line of code was written, and it has almost nothing to do with the technology.
Read 027How to Start Collecting the Data That Becomes Your Edge
Everyone now rents the same AI models, so the model is not the advantage. Your data is, the record of how your business actually runs, that exists nowhere else. Most companies generate that record every day and let almost all of it evaporate. Here is how to start keeping it, and why it is the one asset a competitor cannot buy.
Read 026Calling All Personas: What a Well-Built AI Agent Is Really Worth
An AI agent is not a chatbot. It is a persona given one job, a set of tools, and clear limits, a specialist you add to the team without adding a chair. Built well, it is one of the most useful things AI can do for a business. Built badly, it is one of the most expensive. The gap between those two is the whole point.
Read 025Meaning, Measured: How AI Turns Your Words Into Numbers
Under every AI that seems to understand you is one quiet trick: it turns meaning into numbers, and lays those numbers out so that similar ideas sit close together. It is called an embedding, it is the most underrated idea in AI, and once you see it you will understand exactly why the way you say things changes what the machine gives back.
Read 024Harnessing the Light: How AI Actually Works, If You're Extra Curious
Ask what AI is really doing under the hood and you get either a shrug or a wall of jargon. Here is the honest answer, built up from a light switch, told the way you would explain it to a curious ten-year-old. The surprising part is how much of the modern AI story is literally made of light.
Read 023AI Development for Campbell River Operators
Campbell River businesses move information between the field, the office and specialist systems. The gap is rarely the model, it is the handoff that never reaches a structured record.
Read 022AI Development from Comox, for the Comox Valley
The useful AI opportunities in Comox, Courtenay and Cumberland are rarely a chatbot on a website. They are inside the operation, intake, document prep, field handoffs and the incomplete transfer that becomes tomorrow's problem.
Read 021When Retail and Wholesale Share One Messy Product Record
A specialty retailer and a regional producer do not run the same operation. Both lose time when product information, stock levels and orders live in separate places, and staff rebuild the same picture every week.
Read 020Why Project Information Breaks Between the Site and the Office
The crew knows what happened on site. The office knows what the spreadsheet says. The gap between them is where change orders stall, costing goes stale and customers call for status nobody can answer from a single source.
Read 019When a Notary Practice Outgrows the Inbox
The professional work is fine. The administration around it, intake, document assembly, follow-up, status, is where BC notary practices and small law firms lose the week. The fix is rarely a bigger platform. It is usually a controlled layer around the systems already carrying the record.
Read 018How to Solve Your Business Admin Pipeline
Every business runs a quiet pipeline of admin, the forms, the follow-ups, the copying of the same thing from one place to another. There are two ways to fix it, and most people only try one. Automating the repetitive is the obvious half. Rethinking what should exist at all is the half that actually moves the business.
Read 017Why We Don't Believe in Vibe Coding
You would never let someone build your house by feel, no architect, no blueprint, no permits, just vibes and a nail gun. Vibe coding is exactly that, applied to the software your business will run on. It's wonderful for a toy and reckless for anything real. Here's the difference, and why we start with a plan.
Read 016What You Think AI Can Do, and What It Can Actually Do, Are Further Apart Than You Think
Two gaps run in opposite directions. What's actually possible with AI is far bigger than most people imagine. And whether you ever reach it depends almost entirely on how the system underneath is architected. The imagination and the engineering have to meet.
Read 015How I Burned Through a Week's Budget in a Day
I left an agent running and came back to a torched budget. The lesson wasn't that AI is expensive, it's that the cost lives in the structure you wrap around the model, not the model itself. Get the structure wrong and the meter doesn't add up, it multiplies.
Read 014Building the Agentic Layer on Vancouver Island
A new layer is forming in how businesses run, where AI agents do real work, not just answer questions. That layer is going to get built everywhere. The only question is whether the Island's gets built for here, by people who are of here. We intend for it to be ours.
Read 013Why AI Is Not the One-Click Tool It's Made to Look
You watched someone type one sentence and get a finished thing in seconds. What you didn't see were the weeks that made those seconds possible. Here's the work under the waterline, and why it's the part that actually matters.
Read 012Quantum Security and the Agentic World
Quantum computers threaten the math that secures almost everything digital. The strange part is that the threat is already here, because data stolen today can be decrypted later. Here is what that means, what's hype, and why a world full of AI agents raises the stakes.
Read 011Just Hype and Noise, or Real ROI?
Every tool claims to change everything, every week is a breakthrough, and most of it is noise. Underneath the noise there is real, measurable ROI, but it hides in unglamorous places the hype skips. Here is a plain test to tell the two apart.
Read 010A Chat Subscription Is Not the Same Thing as AI Development
A chat subscription makes a person faster at a task. AI development makes the business run differently. One is rented and identical to your competitor's; the other is built around your data, wired into your operations, and owned by you. Here is the whole difference, honestly.
Read 009Prompt Security and Tool Security: The Two Doors Into Your AI
There are two doors into any AI system: the words going in, and the tools doing the work. Most of the real risk, and most of the real engineering, lives at those two doors. Here is what is behind each, in plain language, and how we build so an attack can't turn into a loss.
Read 008The Most Valuable Thing in AI Is the Data Nobody Has Yet
Everyone has the same models now, so the model is not the edge. The edge is information the model has never seen, data that does not exist anywhere yet because nobody went out and made it. There is a whole island of it, and that is what we are building toward.
Read 007Everything Your AI Makes Is Already Old
Ask the best AI in the world for something brand new and it hands you something made entirely of old. Everything a generative model produces is a derivative, derived completely from the past, with no source of its own. Once you see it, it changes what these tools are for and what they can never be.
Read 006The Smartest AI You Can Buy Still Cannot Answer the One Question That Matters
Judea Pearl won the Turing Award for teaching machines to handle uncertainty. Then he spent the rest of his career arguing that handling uncertainty was never going to be enough. His Ladder of Causation explains the ceiling, and it changes how you should think about every AI decision you are about to make.
Read 005Where Does Your AI Actually Run?
A field guide to hosted versus localized AI for business owners: the hardware, the money, the law, and how to tell which side of the line you are on. Most people are on the same side.
Read 004The Business Operating System
When intelligence becomes your company's primary input. A guide to what a business operating system actually is, what specifically changes about the value a company can produce, and how one gets built without betting the company to do it.
Read 003Why Applications Matter More to Businesses in 2026
The companies that grow faster are not just buying more tools. They are turning important workflows into applications that fit their customers, teams, data, and operating model.
Read 002Building With AI Without Handing the Product to the Tool
AI accelerates how we build and powers what we deliver, but it does not replace product judgment, architecture, security review, or operational responsibility.
Read 001CI/CD Is How Software Development Stays Honest After Launch
A build is not finished when the first version ships. Continuous integration and deployment create the release discipline that keeps changes reviewable, reversible, and safer to maintain.
Read