Key ideas: the most useful AI roadmap barely mentions AI. Smaller firms win on the absence of legacy, not the presence of tools. Start with the ecosystem you already pay for. Vibe coding is a genuine gift right up until it meets production. And the junior career ladder is breaking quietly while everyone looks somewhere else.
For the last episode of Season One of Beside Ourselves, we stopped choosing our own topics and asked listeners to choose them for us. What came back — from LinkedIn, from clients, from people we work with — grouped fairly neatly into six questions. Some strategic, some firmly in IT, and one that nobody in the industry has a good answer to yet.
The speedboat and the supertanker
The standard picture of a large organisation is a supertanker. Enormous momentum, a very long turning circle, and a decision-making process that runs through several decks before anything happens. Smaller businesses are the speedboats: quick, responsive, able to change direction in the time it takes a bigger competitor to schedule the meeting about changing direction.
That’s usually presented as pure upside, and mostly it is. But it’s a great deal easier to capsize a speedboat than a supertanker.
Speed is only an advantage when it’s pointed somewhere. Changing tools every quarter because a competitor mentioned one, or because a vendor put “AI” on a slide, isn’t agility — it’s chaos with a faster refresh rate. It makes everybody seasick, and it ends badly. The firms getting real value from AI aren’t the ones moving the fastest. They’re the ones who worked out where the friction actually was, picked the two or three points that mattered most, and went at those deliberately.
Which brings us to the roadmap.
Not so frequently answered
1. The best roadmap barely mentions AI. When clients ask what a realistic roadmap looks like, the honest answer is that it looks a lot like it did before the hype. Start with the problems, not the technology — and certainly not with AI. Nine times out of ten, arguably ninety-nine times out of a hundred, AI isn’t the answer. The friction is a training gap, an undefined process, or a data set nobody has curated in four years. Those come first. AI is a train that runs on data rails: no amount of magic dust fixes an unmaintained data layer.
2. Your advantage is the legacy you don’t have. The question was how SMEs use agility to take market share with off-the-shelf tools. The most reliable answer is to prove a negative. Where are you unencumbered compared to your larger competitors? Fewer layers, fewer approval gates, less regulatory drag, and — the big one — far less legacy technology. AI has also genuinely reset parts of the playing field. Search and discovery are being rebuilt in public right now, which means the content and positioning work you do this quarter counts for more than it did two years ago. Find where you’re light and push there.
3. Start native. Leave native on purpose. On ecosystem lock-in: if you’re a Microsoft house or a Google house, you’re already paying for a substantial amount of AI capability, and a good deal of it is better than its reputation. Our current read is that the Gemini suite is more sophisticated than Copilot — though Microsoft has opened up considerably, to the point of using Anthropic’s models inside PowerPoint, and a year ago we’d have been far ruder about Copilot than we are now. Start there. When the native tools genuinely can’t do the specific thing you need, go outside — but as a piloted, intentional decision with controls attached, not as a shrug in the direction of whatever’s trending.
4. Bespoke is worth the debt — if you can afford the funeral. Building custom tools has become astonishingly cheap, and Jevons’ paradox is doing exactly what it always does: the price of building a prototype has fallen through the floor, so we’re building far more of them. That’s fine. It’s worth owning a bespoke stack if you have the people to maintain it and, just as importantly, the processes to kill things that don’t work. Vibe coding done with a purpose is not as airy-fairy as the name suggests. But the engineering discipline doesn’t disappear because the typing got easier — information security, data protection and the question of what happens when it breaks all still apply. Expect this to sharpen: as more SMEs move towards accreditations and ISO standards, you’ll need to demonstrate a clear line between “prototype we experimented with” and “system we run the business on”.
5. Ask whether you’d miss it. Every platform and its dog now has AI stitched into it, and some of those implementations are excellent. Others are aspiration presented as reality. You don’t need to be technical to tell the difference. Two tests do most of the work. First: if the feature vanished tomorrow, would you notice? Second: is it actually AI, or is it an algorithm in a new jumper? Taking A and B and producing C is computation. AI is reasoning, filling in gaps, handling the ambiguous middle. Then pressure-test the breadth and depth — a surprising number of these features sort of work, in a first-iteration way, right up until you use them properly.
One criterion we apply constantly: use AI where the accuracy of the outcome is elastic. Where the output must be testable, verifiable and correct every time, a probabilistic system is the wrong place to put it. A great many bolted-on AI features fail precisely this test.
6. Somebody has to train the juniors. This is the one we can’t answer cleanly. The old mechanism was straightforward: hire a junior, break the work into manageable pieces, hand them over, review, correct, watch the capability compound. Those pieces now come ready-made, so the mechanism has quietly stopped working. The tempting response is to stop hiring juniors at all — a decision that will look extraordinarily expensive in a decade. The better one is deliberate mentoring: keep hiring, redefine the roles, and carve out teaching time on purpose rather than hoping it happens in the gaps. And accept that the exchange runs both ways. They will use the tools better than you do. Trust them on that, and take the lesson.
The encouraging part
Almost everything above is within reach of a business with no AI team, no data science function and no eight-figure transformation budget. The hard parts of AI adoption in 2026 are not technical. They’re about knowing where your friction actually sits, being honest about what your data looks like, and having the nerve to run a proper pilot instead of buying whatever your incumbent vendor put on the roadmap slide.
That’s a set of problems that SMEs are structurally well equipped to solve. It just requires doing them in order.
What to do next
- Find the friction before you find the tool. List your top five operational pain points. Note how many are really about people, process or data. Start there.
- Audit what you already pay for. Most firms are running native AI capability they’ve never switched on.
- Apply the elasticity test. For every proposed AI use case, ask whether the output can tolerate being approximately right. If it can’t, use something deterministic.
- Run pilots, not rollouts. Time-boxed, measured, with an agreed way to kill it.
- Draw the line between prototype and production. Write it down before an auditor asks you to.
- Book the mentoring time. If junior development isn’t in someone’s calendar, it isn’t happening.
Season Two arrives in a few weeks in a new, deeper-dive format — if there’s a topic you’d like taken apart properly, tell us.
