A Fortune report from March 2026 put a number on something most operators already felt: roughly 70% of operational decisions inside the average enterprise have never been formally documented. That figure explains a lot. It explains why AI agent pilots keep stalling between demo and deployment. It explains why the same company can spend $200K on a procurement agent and still have three people manually routing exceptions nobody thought to map. And it explains why Interloom — a Munich-based startup building what they call "enterprise memory" for AI agents — just raised $16.5 million in seed funding from DN Capital and Air Street Capital. The money isn't chasing another agent framework. It's chasing the infrastructure that makes agents actually work.
The 70% You Never Documented
Every organization runs on two operating systems. There's the documented one — SOPs, runbooks, process maps. Then there's the real one: the workarounds your operations lead figured out in 2019, the vendor escalation path that only two people know, the approval logic that lives entirely in someone's head.
That second system handles most of the actual work. And it's invisible to every AI agent you deploy.
This isn't abstract. Commerzbank deployed Interloom to analyze millions of customer support emails and found a 50% gap between documented and actual operational knowledge. After building what Interloom calls a "Context Graph" — a continuously evolving model of how decisions actually get made — they reduced that gap to 5% (per Interloom, March 2026). Zurich Insurance, JLL, and logistics firm Fiege are running similar deployments. The common thread: before you can automate a process, you have to understand how it actually works. Not how the SOP says it works. How it works.
Dozens of Agents, Zero Shared Context
Techaisle's 2026 SMB predictions describe a near-term future where the average mid-market company runs dozens of autonomous agents from different vendors. Procurement agents. Support agents. Data pipeline agents. Each vendor ships their own model, their own integrations, their own context window. None of them share memory.
Lio — which raised $30M from Andreessen Horowitz in March 2026 — demonstrates what's possible when the context layer works. Their procurement agents automate 75% of previously outsourced work at one global manufacturer, freeing the equivalent of 10 full-time employees. Walmart's global procurement lead told TechCrunch the "ROI compressed from years to weeks."
But Lio succeeds precisely because procurement is one of the most documented functions in enterprise operations. Purchase orders, supplier evaluations, contract terms — the data exists in structured form. Most operational functions aren't that clean. Customer success, field operations, partner management, executive decision-making — these run on tacit knowledge. And until that knowledge gets captured into infrastructure that agents can actually access, the orchestration problem keeps getting worse.
Building the Institutional Memory Layer
The companies closing this gap are building three things.
Knowledge capture at the process level. Not another wiki. Not another documentation sprint. Infrastructure that ingests real operational data — support tickets, email threads, call transcripts, work orders — and extracts decision patterns from how experts actually resolve problems. Interloom's Context Graph is one approach. Internal teams are building similar systems on RAG architectures tuned for operational data rather than general knowledge.
Shared context across agents. Every agent deployment needs access to the same institutional memory. When your procurement agent approves a vendor, your compliance agent should know. When your support agent escalates a case, your account management agent should have the full history. This is the orchestration layer most agent stacks are missing entirely.
Continuous learning loops. The 70% gap doesn't close once. Operational knowledge changes with every new hire, every process update, every edge case. The memory layer needs to update continuously — not through manual documentation, but by observing how work actually gets done.
The Competitive Window
The AI agent market is projected to grow from $1.5 billion to $41.8 billion by 2030 (Tracxn, March 2026). The models will keep improving. The integration tooling will mature. But the companies that pull ahead won't be the ones with the best agents — they'll be the ones whose agents actually understand how the business operates.
That means building the memory layer now, while the gap between "what's documented" and "what actually happens" is still a competitive advantage to close. The enterprises doing this today — Commerzbank reducing their knowledge gap from 50% to 5%, manufacturers automating 75% of procurement — are building moats that compound. Every process captured, every decision pattern extracted, every edge case logged makes their AI agents smarter and harder to replicate.
For growth-stage companies running $5M–$100M in revenue, this isn't a theoretical concern. It's the difference between AI agents that demo well and AI agents that actually run your operations.
If you're deploying AI agents and hitting the institutional knowledge wall, let's talk about building the memory infrastructure that makes them actually work.
Also published on TouchpointAI and MBC Partners Resource Library.
If your AI agent deployments are stalling between pilot and production, the missing piece might be institutional memory infrastructure. Schedule a conversation with our team to discuss building the knowledge layer that makes agents work.
