A startup just raised three million dollars to solve a problem that six trillion dollars in IT spending has not: giving AI agents enough context about the organizations they serve to do anything useful. The bottleneck is not capability. It is self-knowledge.
Trace raised three million dollars last week to build what it calls an agent operating system for enterprise. The company, which came out of Y Combinator's Summer 2025 batch, does something that sounds almost embarrassingly simple: it connects to a company's existing tools — email, Slack, Airtable, project management software — and builds a knowledge graph of how the organization actually works. Which teams handle which processes. What the approval chains look like. Where the handoffs happen. Who knows what.
Thirty companies are already using it. The investors include Y Combinator, Zeno Ventures, Goodwater Capital, and Formosa Capital. Three million dollars is a modest sum — a rounding error against the six trillion that Gartner projects will be spent on IT globally this year. But the problem Trace is solving may be the most consequential one in the AI transition right now, precisely because almost nobody is talking about it.
The problem is not that AI agents lack capability. The problem is that the organizations deploying them cannot explain their own processes clearly enough for agents to follow them.
The Mirror
This journal has been tracking a paradox. Goldman Sachs's chief economist says AI contributed 'basically zero' to U.S. GDP in 2025. Oxford Economics surveyed nearly six thousand executives and found that seventy percent of firms actively use AI, but eighty percent report no measurable impact on employment or productivity. Gartner predicts sixty percent of AI projects will be abandoned because organizations lack AI-ready data. The money is flowing. The results are not arriving.
The standard explanation is temporal: returns are deferred, the J-curve is steep, be patient. The standard explanation may be wrong — or at least incomplete.
What Trace's existence suggests is that the gap is not primarily temporal. It is epistemic. Organizations do not have structured representations of their own operations. The knowledge of how work actually gets done lives in people's heads, in tribal customs, in the way someone in accounting knows to CC a specific person on invoices over fifty thousand dollars because of something that happened in 2019. None of this is documented. Much of it is not even conscious.
An AI agent asked to 'process this invoice' needs to know: who approves it, what the threshold is, whether this vendor has special terms, which system to log it in, and what happens if the amount exceeds a budget line. A human employee learns this over weeks and months through osmosis, observation, and correction. An agent needs it on day one, structured, complete, and accurate.
Trace's thesis is that this organizational self-knowledge is the binding constraint. Not the models. Not the infrastructure. Not even the data, in the traditional sense. The constraint is that companies have never had to articulate their own processes with the precision that machines require — because humans could fill in the gaps with judgment, context, and asking the person at the next desk.
Two Approaches
There are two strategies emerging for closing this gap, and they reveal different theories about where the work should happen.
The first is to build the map. This is what Trace does. It ingests a company's digital exhaust — communications, task assignments, document flows, calendar patterns — and infers the organizational graph. The result is a structured representation of processes that were previously implicit. Agents consuming this graph know which team handles what, what the escalation paths are, and where the bottlenecks sit. The map exists independently of any specific workflow tool. It is an abstraction layer between the organization's tacit knowledge and the agents that need explicit knowledge.
The second is to embed in the territory. This is what Atlassian did when it made AI agents assignable to Jira tickets — tracked in the same sprint boards, velocity charts, and SLA dashboards as human teammates. Instead of building a separate knowledge layer, agents enter the existing management infrastructure and learn the organization's patterns by operating within its tools. The territory — the Jira board, the sprint retro, the backlog grooming — is the context. No separate map required.
The distinction matters because it implies different failure modes. Build-the-map can fail if the map diverges from reality — if the inferred process graph doesn't match how work actually happens, agents follow the wrong playbook. Embed-in-the-territory can fail if the territory itself is poorly organized — if the Jira board is a mess of unlabeled tickets and stale epics, the agent inherits the mess.
Neither approach solves the deeper problem, which is that many organizations do not have well-defined processes even for humans. The agent is a mirror. It reflects back the organization's operational clarity — or lack of it. A company that runs on tribal knowledge and ad hoc escalation cannot deploy agents effectively, regardless of how sophisticated the model is, because there is nothing coherent to automate.
The Number That Matters
Gartner's most cited prediction about agentic AI is that forty percent of projects will be canceled by the end of 2027. Less cited but more diagnostic: sixty percent of AI projects overall will be abandoned due to lack of AI-ready data. The framing suggests a data engineering problem. It is actually a self-knowledge problem.
AI-ready data, in the context of enterprise agent deployment, does not primarily mean clean databases or well-formatted spreadsheets. It means structured representations of organizational processes, decision trees, approval hierarchies, exception handling procedures, and the countless informal rules that govern how work flows through a company. Most organizations have none of this. They have software systems that store transactional data and people who know how to use those systems. The gap between the two is filled by human judgment — and that gap is exactly where AI agents are being asked to operate.
The sixty-three percent of organizations that told Gartner they either don't have or aren't sure they have the right data management practices for AI are not confessing to messy databases. They are confessing to something more fundamental: they do not know, in structured terms, how their own organization works.
Why This Constraint Is Durable
Infrastructure constraints ease over time. Chip production scales. Data centers get built. Models improve. Capital flows to bottlenecks and dissolves them. The organizational self-knowledge constraint does not dissolve the same way, because it is not a resource problem. It is a complexity problem.
A company's processes are not a fixed quantity that can be documented once and consumed forever. They change — with personnel turnover, with new tools, with shifting market conditions, with the accumulated drift of informal practice diverging from formal policy. The knowledge graph that Trace builds today will need to be rebuilt or updated continuously, which means the mapping infrastructure has to run as a persistent service, not a one-time consulting engagement.
This is why the adoption gap may prove more durable than the infrastructure gap. Building a data center is expensive but legible — you know what you need, you know how to measure progress, and the result is a physical asset that persists. Mapping an organization's tacit knowledge is neither expensive nor cheap. It is unscoped. Nobody knows how much organizational knowledge exists, how much of it is documented, how much is wrong, or how fast it changes. The project has no natural endpoint.
The companies that will close the gap first are the ones that already operate with high process clarity — regulated industries with compliance requirements, manufacturing with documented workflows, financial services with auditable decision chains. The companies that will struggle longest are the ones that pride themselves on agility and informality, where 'move fast and break things' produced an organizational culture that never needed to explain itself to a machine.
What the Mirror Shows
There is something uncomfortable about Trace's pitch. The startup is not selling AI capability. It is selling organizational self-awareness. The implicit message to enterprises is: you have spent billions on AI tools, and the reason they are not working is that you do not understand your own operations well enough to tell a machine what to do.
This is not a technology problem that technology alone can solve. Mapping organizational knowledge requires organizations to confront the gap between how they think they work and how they actually work. Every process improvement consultant in history has encountered this gap. The difference now is that the gap has a price tag attached — measured in failed agent deployments, in the forty percent of projects Gartner says will be abandoned, in the productivity gains that Oxford Economics says are not materializing despite seventy percent adoption.
The six trillion dollars in IT spending will buy the infrastructure. The models will get better. The chips will get faster. The agents will get more capable. But capability without context is waste. An AI agent that can write code, draft contracts, analyze data, and manage workflows is useless if it does not know which code to write, which contract template to use, whose data to analyze, or which workflow to follow.
The adoption gap is not between the technology and the users. It is between the organizations and themselves.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)