Henry DentonFounder, CEO
Get a demo. Watch a live capture, then an AI agent query the result.
Ask anything. Pricing, security or integrating with your stack.
Free demo, no commitment
I'm sure you have heard countless times "the bottleneck to automation is no longer model intelligence, it's access to the tribal knowledge that actually runs your organisation". I'm even guilty of being one of the perpetrators myself. No, I wasn't reading off a 2026 AI bingo card, I do fully believe this. My only issue, in hindsight, with the statement is that it fails to pinpoint exactly why this is the bottleneck. Why can't our current systems capture this oh-so-famous "tribal knowledge"? Surely it's as simple as writing it down. Well, that in itself has just outlined the problem. Documentation.
When someone explains how to do something in person, they usually tell you why you should take each step. Or if you're nosey like me, you ask them why before they've even finished explaining it.
Written instructions don't do this. Anyone who has stared at an IKEA manual knows the feeling. The entire sum of communication reduced to a picture of a man with an Allen key and the letters A + B + C.
So why do we give that extra piece of information in person that we leave out when we write things down? Surely it can't only be so a nosey individual knows what you're up to. It's primarily for two reasons. When we aren't there in the moment undertaking the process ourselves, we can't remember the reasoning behind the action. And even when we can, we forget how important it is to provide the reasoning.
It allows us to know the underlying intent of what is supposed to happen in each step. Say, for example, you are following a recipe that says "sear the steak in a cast iron pan". You don't have a cast iron pan. No steak for you. But if the recipe said "sear the steak in a cast iron pan because cast iron retains heat evenly, giving you a nice sear", you'd know what you're actually trying to achieve. So you'd grab your heavy-bottomed stainless steel pan and get cooking.
This exact same principle applies to AI agents. Well, maybe not cooking steak, yet. But the idea is that if you give your reasoning, they can think from an intent perspective, which lets them know what to do when your situation isn't ideal.
If your organisation needs to document every process, preference and scenario, they're likely going to be using markdown. I can see why people describe it as the "lingua franca". One half of that is because humans are lazy, the other half is because LLMs will parse just about anything. Whether they intentionally parse that information is another question.
When your AI agent is reading your markdown docs, it's inferring everything. If you were a customer support agent, let alone a fighter pilot, I would rather you didn't guess.
Aside from the terrifying image of fighter pilots running on markdown, the other issue with this format is that you can't even tell if critical details exist, are valid or up to date.
Now, you could argue that the world runs on plain text: laws, emails, text messages, so what's wrong with markdown? Plain text causes miscommunication constantly. One of the great things about computers is that their true or false nature removes that ambiguity.
If I asked you to sit down and list how, when and why you do every task on a day-to-day basis at work, you probably wouldn't know where to start. And even if you did, it would take weeks. This is productive time that most employers simply don't have to spare.
But let's imagine your employer can afford to lose you for a few weeks. You sit down and write up every process you can think of. Then, after all those weeks, you've only captured 60% of what you actually do. The other 40% wasn't included because to you, you wouldn't even think to write it down, it just comes naturally to you. Just like how you would remember to pull a door whilst locking it to get the lock to turn. You never think of it unprompted.
That 60% you wrote down has a shelf life too. In six months' time half of the processes will have changed, but nobody is going back to update them. So now you've got documentation that's incomplete and out of date. At that point you might as well not have bothered.
Currently, there's no shortage of startups working on retrieval, memory and connectors for AI agents. The issue is that every one of these assumes the tribal knowledge is already stored somewhere and just needs to be found and accessed more efficiently. As we've established, this isn't the case.
So what would it actually take? If we work backwards from each of the outlined problems, the requirements are actually rather simple. We need something that can observe our actions, understand them and ask a question in the moment when something isn't clear or contradicts what we have done before. Then this needs to be provided in a consistent structure.
Just like an apprentice always watching over our shoulder and asking questions when something doesn't make sense. However, unlike any apprentice you've ever met, this one actually writes up everything it learns in a structured way.
In early 2026, I was going through annual reports from some of the UK's largest construction companies. I was aware the industry had significant issues and wanted to find out more. I found that only 8% of the UK construction workforce is aged 18-25 and over a third is over 50. Balfour Beatty, one of the UK's biggest contractors, described it as "unprecedented demand and a widening skills gap". I thought to myself, this means one of two things: either there is going to be an extreme transfer of knowledge from experienced individuals to a new workforce, or there is a requirement for automation.
Around the same time, I was building an agent for a client at the agency I ran. The agent was performing well and they were impressed, but after a while it was clear it was far from perfect. It was treating every order with the same level of priority, dispatching them in the order they came in. The problem was that retail customers were being put behind trade customers in the queue. You'd have Jane in Bournemouth expecting Amazon-level next-day delivery waiting a week, while a trade shop in Wolverhampton gets their order a month ahead of schedule. Anyone who'd worked there knew you always prioritise retail over trade because trade customers are locked into a contract and retail customers aren't. Between the construction industry losing its most experienced workers and an AI agent that was far from being a self-sufficient team mate, the problem became rather clear to me. We need a better way of capturing experience.
After working back from the problems, I ended up developing an always-on software application that passively observes your actions and understands what you are doing in the moment. If the reasoning behind an action isn't obvious, or if you contradict something you would usually do, it asks you questions in the moment to find out your reasoning. Then, it distils everything it captures into patterns, essentially a rulebook for AI agents to query. I called it FusedFrames.
Had FusedFrames been running in that warehouse, it would have observed them fulfilling retail orders before trade orders on their desktops, asked them why and never let an AI agent make the mistake. Jane would certainly have got her order on time with the Amazon-level delivery she expected. As for the sixty-year-old construction engineer who's been counting down the days to retirement, thirty years of experience won't be walking out the door when they do.
The bottleneck to automation is no longer model intelligence. That certainly hasn't changed. What I hope this article has made clear is why this bottleneck still exists, what it will take to capture that knowledge and the wrong assumptions people have made in solving the problem.
FusedFrames is my attempt at fixing that bottleneck. Capturing what has never been captured before: the unwritten reasoning and the edge cases you wouldn't even think to document. Then providing it in a format optimised for agents to act on.