The UK is putting serious weight behind AI. London’s new AI and Jobs Taskforce signals that AI has moved from a technology debate into a workforce and productivity debate. It is the right conversation, but most businesses do not have an AI problem first. They have a collaboration architecture problem, and AI is exposing it as a memory crisis.
By memory, I do not mean more storage, longer chat histories, or another searchable archive. I mean the ability to preserve the context of work, from what was agreed and why it mattered, to which customer it affected and what the team learned.
AI is exposing fragmented work
The modern workplace already asks people to coordinate work across too many disconnected systems. Harvard Business Review has reported that workers toggle between apps and websites roughly 1,200 times a day, with the reorientation cost adding up to just under four hours a week.
Collaboration has also become more time-consuming. Another Harvard Business Review analysis found that collaborative work has risen by 50% or more over the past decade and now accounts for 85% or more of most people’s working weeks.
The hidden cost is the articulation work that sits between tools, from finding what was agreed and re-explaining it to waiting for information that already exists elsewhere. That work is necessary, but without preserved context, it becomes a burden.
This is where AI programmes fail. An agent can summarise a document or retrieve a message, but if organisational knowledge lives across chats, tickets, CRM records, code repositories, dashboards, emails, and people’s heads, it is still operating on fragments.
Channels were never designed to become memory
These systems were built around channels and threads. That made sense when the main goal was communication. It works less well when the goal is persistent understanding.
Work does not happen inside a single channel. It happens around entities – accounts, opportunities, tickets, incidents, customer commitments – and the context should sit with those entities, where ownership, history, dependencies, permissions, and accountability reside. A channel treats this as conversation. A knowledge graph treats it as connected work.
When those relationships are left implicit, the stitching falls back to people. New employees learn naming conventions before they understand the business, teams revisit old decisions because the rationale is buried elsewhere, and customers feel the gap every time they have to repeat themselves.
Search is not understanding
“PostgreSQL migration” in one channel, “database move Q3” in another, and “backend infrastructure change” in a third may describe the same decision. Text search can return those messages. What it cannot do on its own is cluster the discussion, identify the underlying choice, connect it to downstream work, or understand which version is important now.
Information can exist and still be functionally dead if the cost of finding and interpreting it exceeds the value of having it. This is why I believe the next phase of enterprise AI is a systems design problem, not a prompting problem. Larger context windows may help individual interactions. They do not create durable organisational memory. More agents do not fix broken context.
The infrastructure layer AI needs now is shared memory, built around a structured, permission-aware representation of work, people, decisions, customers, products, and outcomes. That means moving from channels to knowledge graph-based collaboration, where conversations are embedded in the work objects they affect.
For AI to become genuinely useful at enterprise scale, it must deliver three things simultaneously: precision, efficiency, and safety. Precision means answers are grounded in authoritative business data rather than inferred from disconnected documents. Efficiency means reasoning from an existing organisational memory rather than repeatedly reconstructing context through expensive prompts and retrieval cycles, reducing overall token usage and, therefore, cost. Safety means understanding not only what information an agent can access, but also what actions it is allowed to take and how those actions can be governed, observed, and reversed.
From personal productivity to team intelligence
McKinsey’s 2025 State of AI survey found that 88% of respondents say their organisations regularly use AI in at least one business function, but only about a third have started scaling AI programmes across the enterprise. It also found that 23% are scaling agentic AI, while another 39% are experimenting with agents.
This shows us that AI adoption is no longer rare. But scaled, reliable AI operations remain hard because agents need to reason from the same organisational context as the teams they support. Without shared memory, organisations struggle to achieve the precision required for trustworthy answers, the efficiency required for sustainable economics, or the safety required for production deployment.
Team intelligence is the institutional know-how that persists across people, tools, time zones, and boundaries.
In a memory-aware organisation, a support ticket is not isolated from the account, product issue, engineering discussion, or decision history. A new team member does not need six months of messages to understand why a choice was made. An agent can surface relevant history, identify missing stakeholders, or avoid recommending something the team already tried.
The customer is part of that fabric, too. They should not need to join an internal channel to understand progress, and internal teams should not reconstruct a customer’s history at every boundary. The same is true for AI agents. They are most useful when they work with the team and have the right visibility, rather than as generic assistants outside the organisation.
The UK is right to focus on AI as part of its productivity agenda, but businesses will not become more productive simply by adding agents to every workflow. If every important answer still requires a human to reconstruct scattered context, AI becomes another layer of activity atop the same fragmentation. Britain will not close its productivity gap by adding more agents. It will close it by building organisations that actually remember.
Deploying AI on top of fragmented memory does not create intelligence. It accelerates amnesia.
