// Field notes from the founders
AI has made your silos worse
AI was meant to be the fuel that turbocharged companies and productivity, automating the manual tasks, streamlining the workflows, shortening the time from ideation to action and ultimately unlocking value for businesses and customers alike. While there is no doubt AI has already changed how we work, and is having an impact in places (most notably in software engineering), in most cases the real world benefits are isolated, and in some aspects it has actively made things worse.Like everyone else, what started out as casual use a few years ago, turned into "how do we really put AI to work" for us over the last year as frontier model capability improved rapidly, and seemingly crossed an invisible Rubicon. What we observed is not what was promised. AI and automation didn't streamline workflows and the silos didn't disappear, they moved! Out of the tools and teams and into the chatbot threads.Look around a typical £10m consumer brand today. The growth lead has her ChatGPT. The founder "prefers" Claude. Finance uses a copilot living in a spreadsheet (we already pay for it as part of M365). Each of these completely mind blowing just a few years ago, yet each completely blind to the other two. The AI we were promised would dissolve the silos between tools has quietly rebuilt them between people — and made them invisible, because they live (and die) inside private chat histories nobody else can see.This fragmentation shows up in three specific places. Each is the subject of its own post in this series.One: the inputs are wrong, and "asymmetric". Your AI is only as good as the numbers you paste into it — and in most D2C brands, the numbers are quietly fiction. In-platform CAC from Meta is $30 USD but finance think it's $300. LTV is a back of envelope calculation based on the assumption that the average customer places "about four orders". Retention is in a spreadsheet refreshed when someone remembers. The growth lead has connected up Google and Meta to ChatGPT, the Finance team connect to Xero, P&Ls and spreadsheets with data dumps from the backend, and the founder (sometimes) adds the business plan and board packs as context. Each member starts with a different and incomplete "slice", and none of them seeing the full truth. Advances in model performance don't solve for this, and, if anything, actively make it worse as people progressively place more trust in the LLMs as they hallucinate less and reasoning improves (more on this: Why the next wave of AI is about structured data).Two: it's single-player in a multiplayer company. Everyone works in their own AI tool, and their own threads. Context lives, ages (an intrinsic part of automatic compacting of context by LLMs is to clear, condense or down-weight old context) and then dies in the chat thread. The output either stays in the thread, or a dense wall of text is pasted into Slack which nobody reads — the current day equivalent of the fifty-slide powerpoint which makes a small part of your soul depart you, never to return. (more on this: AI is single-player but your D2C business is multiplayer)Three: it can't act — and when it does, nobody can see what it did. Claude tells you exactly what to do, but the next step is normally to copy the recommendation and paste it into your Jira board where it rots for three days, or bounces around in a Slack channel. Did it get actioned? Who knows 🤷. With more MCPs and connectors coming by the day, it is increasingly possible to take action, but this opens up a new Pandora's Box. When was a change made? By whom? And why? And was it a good idea? Do you want the finance team acting on your Google Ads account? Do you want your growth lead making changes to your finance system? Who is leaking PII data out of the backend into publicly accessible gsheets? (more on this to follow)We're not the only ones noticing. On YC's recent Lightcone podcast, Pete Koomen put words to it: we're still in "the single-player era of agents — the harnesses that have got really popular are all designed to be used by a single human." Garry Tan described the consequence of scaling that single-player pattern across an organisation without fixing it: "without a persistent, shared source of truth, you don't get a cohesive virtual organisation — you get a thousand disconnected, hallucinating micro-teams stepping on each other's work."Until recently AI use was often "at the margins" and this effect wasn't visible, but now it's increasingly embedded into how people work, we can see the whites of its eyes, and we wonder how we didn't spot this problem creeping up on us sooner! If this doesn't immediately resonate, ask yourself, are we collaborating better or worse than we were a year ago? Does cross functional execution feel more or less effective than it did a year ago?Plenty of companies are now solving one of those three fragments. Some fix the metrics, some bolt agentic execution onto an ad account, some are working on the multiplayer AI and some frankly don't serve much use at all. The most ambitious are going after the whole "company brain", but they're going horizontal, of which Glean is the best example: search across whatever tools any company already has. It reads what already exists, for everyone, anywhere. Despite being seemingly slow to react to this, Google and Microsoft will almost certainly dominate the generic company brain alongside one or two well funded new entrants like Glean since they have everyone already locked into their business platforms. If you pay attention, you can watch as week by week they incrementally ship features that move them in this direction.The trouble with horizontal is that it has to stay shallow. A brain built for any business can't build in specific responses to changes in your M1 retention, or that your CAC is double-counted across three ad platforms, or that pausing a creative is reversible but increasing your RAF discount is very hard to row back. It can search your fragments but it can't learn. It can't fix them, connect them, and act on them.The loop (correct inputs → shared memory → permissioned execution) is only possible to close if you go narrow, so we picked our (sectoral) slice and are going deep: high-retention consumer/D2C (Subscription, repeat-purchase, mixed Subscribe-&-Save). Deep enough to compute the right metrics, hold shared context the whole team can act on, combine agentic execution with deep expertise, finely tuned skills and machine learning models to take real actions across the company stack with a seatbelt on.Horizontal company brains will keep getting better at search. Vertical tools will keep getting better at whatever it is they are doing today. We're not building an enterprise search tool, we're closing the loop for one kind of business. We're not focussed on being another productivity tool, we're focussed on helping our clients build a better business and move the metrics that matter.The posts in this series take each fragment in turn: the wrong inputs, the single-player problem, and the last mile AI still can't complete.Check out Signal over Noise to apply for the Pilot!