There’s a specific moment, when you’ve been using AI for long enough, where the output stops feeling like yours. The grammar’s fine. The structure’s fine. It’s also, unmistakably, written by anyone except you.
That moment is when most founders stop pretending the problem is the prompt.
I’ve been running Digital Magic Solutions for years. Twenty-plus in IT, fifteen-plus running my own thing, hundreds of hours inside ChatGPT, Claude, Gemini, and every AI tool a UK SME owner could plausibly try. I knew how to write a prompt. The problem wasn’t the prompt. It was the context.
So I did what any reasonable, slightly stubborn business software guy would do. I built an AI Context Layer for my own business first, before I’d dare sell one to anyone else’s. I wanted to install it on me. See what worked. See what broke. See what I was actually solving.
Here’s what I found was missing. Not from the AI. From me.
Gap one: my voice didn’t live anywhere
I assumed I had a voice. Of course I had a voice. Twenty years of writing emails, proposals, and short videos. It had to be somewhere.
It was somewhere. It was in my head.
The first time I tried to extract it onto a page, I got generic SME marketing copy back. Friendly. Professional. Forgettable. The phrases I actually use in conversation, the analogies I reach for when a prospect goes quiet on a call, the throwaway lines that make a Tech Support Clinic feel like a Tech Support Clinic and not a webinar. None of that was written down. None of it was reachable.
Most founders can’t articulate their own voice without a skilled interviewer drawing it out. I am a founder. I had no interviewer. I had a folder full of half-written brand documents that all sounded like a different company’s about page.
That gap, on its own, would have been enough.
Gap two: my client stories existed, but only as feelings
Ask me about a long-running coaching client and I’ll tell you the story. The slow CRM build, the course launches, the moment the open rate finally lifted. Ask me to put that in a proposal and I’ll get the dates wrong, the numbers fuzzy, and the punchline hazy.
When I let the AI write proposals without that information loaded properly, something worse happened. It made stuff up.
It invented case studies. It quoted percentages I’d never said. It described services I don’t offer. Not because the model was malicious. Because AI abhors a vacuum, and the model couldn’t tell me when it was filling it.
Air Canada had to honour a chatbot’s invented refund policy in court. I’m not Air Canada. I run a small UK service business. A made-up testimonial in a proposal isn’t a tone problem. It’s a trust-destroying incident.
The 54 percent of UK firms who say they use AI right now (British Chambers of Commerce and Paragon Bank, September 2025) almost all have this gap. Around 11 percent have actually integrated it. The 43 percent in the middle are running on stories that live in someone’s head and a model that doesn’t know it’s guessing.
Gap three: my pricing, products, and funnel state changed every week
I run lean. I iterate fast. The Firecore beta price moved twice in three weeks. The ebook funnel went from paid Meta ads to organic only in a single Friday review. The MRM webinar cycle landed on the kill list and stayed there.
If I asked Claude to write something about my offers without a current source of truth loaded, it would draft confidently from whatever it last picked up. Stale beta numbers. Retired packages. A funnel I’d shut down a fortnight earlier. None of it triggered an error. It just quietly went out wrong.
This is the part of the Flamekeeper problem that keeps founders awake. An ungoverned context layer doesn’t fail dramatically. It degrades gracefully. Proposals go out with last quarter’s pricing. Content references services you’ve retired. Bios cite roles you’ve moved on from. None of it pings a warning. It just slowly erodes trust.
The fix isn’t a smarter model. It’s a place where today’s pricing, today’s offers, and today’s strategic priorities live as the single source of truth, and where the AI is structurally pointed at that place before it does anything else.
Gap four: my frameworks kept getting mangled
I have named methodologies. The Flamekeeper Problem. The Kindling Method. The four-layer, fifteen-module Firecore architecture. Big Domino reframes. Each one is a specific thing with a specific structure I’ve refined across hundreds of conversations.
Without them loaded, the AI would give me a generic “marketing framework” or invent a fourth pillar to balance my three. It would describe the Kindling Method as a “strategy session.” Close enough to sound right, wrong enough to undo years of careful naming.
Frameworks are the part of an offer that lets a buyer hold an idea in their head. Get the framework wrong and the buyer reaches for the nearest competitor’s framework instead. I’d been losing that fight without noticing.
The pattern underneath all four
The voice, the stories, the offers, the frameworks. Four different gaps. One underlying problem.
Everything that mattered about Digital Magic Solutions lived in my head. The brand voice. The client stories. The product positioning. The way things get explained. The answers to the tricky questions. The tone that makes an email feel right instead of robotic. All of it, in one slightly tired skull, with no structured way for an AI tool to read any of it.
I call that person the Flamekeeper. In most SMEs, it’s the founder. Everything works fine while the Flamekeeper is available. The moment you try to scale, hire someone to write content, delegate proposals, or use AI to speed things up, everything comes out wrong. Because nobody else, human or otherwise, has access to what’s in the Flamekeeper’s head.
I am, exhibit A, the Flamekeeper of my own business. My AI tools didn’t fail. They were never given anything to succeed with.
What changed when I built the layer
I structured everything I’d been carrying into a Firecore: four layers (Identity, Business, Evidence, Operations) and fifteen modules underneath them. Voice and Tone. Brand Assets. Products and Services. Audience and Ideal Clients. Frameworks and IP. Strategic Priorities. The lot.
I governed it. Every change is logged, versioned, and approved. Nothing edits itself. Stale entries get flagged and updated, not buried.
I pointed every AI tool I use at it. Claude reads from it. The marketing skill reads from it. The Friday strategy review reads from it. When a tool drafts a proposal now, it pulls from the canon, not from a guess.
The before-and-after is honestly stark. I wrote about it in detail in my recent experiment giving the same brief twice, once with no context and once with the full Firecore loaded. Same model. Same prompt. Two different businesses, apparently.
Why this is becoming standard infrastructure
Harvard Business Review put it cleanly in February 2026: when every company can use the same AI models, context becomes the competitive advantage. Gartner has flagged context engineering as the breakout AI capability of 2026. McKinsey reckons only 1 in 10 companies are scaling AI successfully. The other 9 aren’t missing the technology. They’re missing the structured context to point it at.
The Firecore is, simply, context engineering for SMEs. The same discipline the enterprise tier is paying Glean and Writer per-seat licences for, built for a UK service business of five to fifty staff, owned by the client, run on their accounts, governed weekly.
You could build your own version. WordPress is free, too. But you came to a professional because you wanted it done right, done properly, and done in a way that actually works for your business. Same principle applied to your AI infrastructure.
The honest takeaway
I didn’t set out to find four gaps. I set out to install a system on myself and prove I’d use it before I sold it. The gaps surfaced because I was paying attention.
If you’ve subscribed to AI tools and you can feel the output isn’t quite landing, the model isn’t the problem. The context isn’t there yet. That’s a fixable thing. It’s also, in my experience as someone who has now built one for himself and is mid-build for the first beta clients, not as big a job as it looks. It just needs structure, governance, and someone to draw the knowledge out of your head and onto the page.
If you want to know whether your business has the same gaps mine did, the AI Context Layer Risk Quiz takes about six minutes and shows you exactly where the holes are. Take the quiz here. You’ll get a personalised score, a short report, and an honest read on whether this is a problem worth solving for you yet.
Either way, write down where your business knowledge actually lives this week. If the answer is “in my head, mostly,” you’ve already found the first gap.