Memo No. 001 Category Investigation 03 May 2026
The Company Brain is the most over-articulated, under-shipped primitive of 2026.
A Factory Research category memo · 3,300 words · 34 citations
Tom Blomfield wants someone to build a living map of how every company in the world actually works. The thesis is correct and the prize is real. But the moat lives in places a small studio cannot reach in one quarter, and the field is already crowded with seven-billion-dollar incumbents and freshly funded second movers.
Tom Blomfield's Company Brain RFS[1] argues that the AI agents are now smart enough, but companies cannot deploy them because the operational know-how that makes a business actually run is locked in heads, Slack threads, support tickets, and dead email accounts. The proposed primitive is not a search box and not a chatbot over documents. It is a structured, executable map of how refunds get handled, how pricing exceptions are decided, how engineers respond to incidents. The map gets fed to AI agents as a "skills file" so they can do the work safely and consistently.
The thesis is correct. Garry Tan has already shipped a public reference implementation called gbrain[2][3], which beats vector-only RAG on retrieval precision by 31 points on his own benchmark. Glean has built a $7.2 billion business proving the adjacent search version of this idea works[4]. Interloom just raised $16.5 million on a near-identical pitch[7]. The category is real and venture-fundable.
Our verdict, after thirty-four citations of category research, is not yet. The thesis tier is genuinely 4. The execution tier, for a small studio, is closer to 2. A defensible Company Brain product requires SOC 2, HIPAA, ACL hydration, fifty-plus connector integrations, an enterprise GTM motion, and the patience to do six-month POCs against incumbents with billion-dollar war chests. The studio that wins this is sales-led, not PLG, and it ships in a vertical wedge, not horizontal. We recommend revisiting if Wes finds a wedge with a captive ICP and a co-founder who has sold into security review before.
The enterprise AI knowledge layer is currently three overlapping markets that buyers and analysts confuse for one. Sub-segment one is enterprise search, where Glean became the unambiguous leader. Glean reached $200 million ARR in roughly nine months after first crossing $100 million, raised a $150 million Series F at a $7.2 billion valuation in early 2026, and reports more than 100 million agent actions taken on its platform annually[4][5][6]. The Glean homepage lists Duolingo, Databricks, Webflow, Grammarly, Booking.com, Instacart, Samsung, Zillow, Rivian, and Intuit among customers[14]. This is not a startup. This is the incumbent.
Sub-segment two is the enterprise wiki and document layer, where Confluence, Notion, and SharePoint already host the existing knowledge and have shipped their own AI bolt-ons. Atlassian's Rovo is bundled into all paid Confluence plans at roughly $5.16 per user per month. Notion AI is an add-on at $10 per user per month. Microsoft Copilot for M365 is $30 per user per month. Slack AI was folded into all paid Slack tiers in 2025 with no add-on fee. None of these are great at the Company Brain job, but all of them sit inside the buying motion the buyer already has.
Sub-segment three is the new category that Tom Blomfield is actually pointing at, which the industry is starting to call the agent context layer or context graph. The most direct comparable is Interloom out of Munich, which raised $16.5 million Series A from DN Capital in March 2026 to capture tacit knowledge from operational records[7]. The next is Lovelace, founded by Andrew Moore, the former head of Google Cloud AI, building a system called Elemental[8]. Open-source efforts include Onyx (formerly Danswer) at roughly $16 per user per month for the enterprise edition, and Garry Tan's own gbrain repository, which serves as a sort of reference architecture for the entire category[2][12]. Dust positions itself as the operating system for AI agents and reports 5,000+ organizations using the platform[13]. Guru, Coveo, Algolia, GoSearch, eesel, Capacity, and a dozen others crowd the alternatives lists[12].
The published total addressable market numbers depend on which sub-segment you draw a circle around. Fortune Business Insights has knowledge management software at $26.4 billion in 2026 growing to $74.22 billion by 2034 at 13.8 percent CAGR[15]. Market.us has the AI-driven KM segment growing at 25 percent CAGR[16]. Mordor Intelligence has enterprise search alone at $7.47 billion in 2026, projected to $11.66 billion by 2031[17]. Whichever number you trust, the answer is the same: this is a tier-4 category by anyone's accounting.
Glean is excellent at what it was built for, which is search-first knowledge retrieval. But Blomfield's RFS is explicit: this is not a search tool and not a chatbot over documents[1]. The unmet need is the executable layer above retrieval. Today, when a Glean answer says "refunds over $500 require manager approval," it cites a Confluence page from 2023. What it does not produce is a structured procedure object an agent can follow with confidence in the live ticket queue. The translation from "this is the policy" to "here is the deterministic skill the agent will execute" is left as an exercise.
The second gap is the tacit knowledge gap. Interloom's pitch quantified it: at Commerzbank, the difference between what was documented and what the company actually did to handle support tickets was around fifty percent of the operational know-how. Interloom claims they closed the gap to roughly five percent by mining email archives and ticket histories rather than asking employees to write more wiki pages[7]. Independent industry estimates put the share of undocumented enterprise processes in large companies at around seventy percent. AI agents that try to operate against the documented thirty percent fail in the same predictable ways[21][23].
The third gap is the long tail of mid-market companies who are not large enough for Glean's six-figure minimum but are still drowning in scattered context. Glean's pricing starts at fifty dollars per user per month with a 100-seat minimum, and buyer-reported contracts routinely exceed two hundred thousand dollars annually before infrastructure costs[9]. Paid POCs run up to seventy thousand dollars before any production deployment[9]. There is genuine room below this floor for a focused offering at, say, ten dollars per user per month with no POC fee, that handles a narrower but high-value slice of the company-brain job.
The fourth gap, and the one that may matter most strategically, is verticalization. Glean and Dust are horizontal platforms. A Company Brain for law firms, for medical billing, for property management, for SaaS support operations, would beat the horizontal players on time-to-value because the skills and the schema can ship pre-built. The horizontal players have to be everything to everyone. The vertical entrant only has to be excellent at one shape of business.
The strongest customer-side evidence is not in product reviews. It is in the operational failure data. RAND Corporation, cited in Composio's 2026 agent report, found that the number-one cause of AI project failure is that stakeholders miscommunicate what actually needs solving[33]. This is not a model problem. It is a knowledge-articulation problem, which is exactly the problem Company Brain claims to address.
From Question Base's complete guide to Slack AI features, on what teams discover after they buy Slack AI: "Teams enable Slack AI and discover that the answers to their most critical questions are not in any channel. They are in the heads of three or four senior people who have never written them down anywhere."[28] This is the verbatim shape of the unmet need. It is not a search problem. The data is not there to search.
From Glean's own POC implementation guide, candidly admitting the failure rate of the product category they lead: "Most enterprise AI projects never reach production. The failure rate sits somewhere between sixty and eighty percent depending on the research."[26] Glean lists three specific reasons POCs fail: information retrieval gaps, lack of trust and traceability, and operational integration friction. All three are direct hits on the Company Brain thesis.
From the Stanford 2026 AI Index, cited by CIO in their Lovelace coverage: hallucination rates across the top twenty-six models range from twenty-two to ninety-four percent depending on the task[8]. From the SQ Magazine 2026 hallucination statistics roundup, sixty-two percent of enterprise users name hallucinations as their single biggest barrier to AI deployment, well ahead of the twenty-eight percent worried about job losses[19]. The Suprmind 2026 benchmark reports retrieval grounding reduces hallucination by seventy-five to ninety percent, and tool grounding by sixty-five to eighty percent[18]. Customers want this product. They are not finding what they want.
From Hypersense's research on production deployments, eighty-eight percent of AI agent projects never reach production at all. Of the twelve percent that do, the differentiator was "proportionally more spend on evaluation infrastructure, monitoring tooling, and operational staffing, and proportionally less on model selection and prompt engineering."[21] A Company Brain product is, fundamentally, an evaluation and governance layer in disguise.
The biggest blocker to AI automation of companies is no longer the models, they just got so good so quickly. Now the blocker is the domain knowledge.Tom Blomfield, YC Summer 2026 RFS
The first shift is model capability. Frontier hallucination rates collapsed from the double digits in 2024 to a 3.1 to 19.1 percent range across reasoning configurations in 2026[18]. The blocker is no longer "the model is too dumb." Once the model is smart enough, the architectural bottleneck moves to context. This is the precise shift Blomfield is pointing at, and it dates to roughly the back half of 2025.
The second shift is enterprise procurement. SOC 2 Type II, GDPR, HIPAA, and ISO 27001 are now baseline-not-optional for any AI deployment in financial services or healthcare[27]. ACL hydration, the discipline of carrying source-system permissions all the way through retrieval and into the model context, became a named primitive in 2025 and is now standard table stakes[25]. This raises the bar to entry significantly, which is bad news for hobbyist studios and good news for any incumbent who has already cleared it.
The third shift is distribution. The release of MCP, the Model Context Protocol, in late 2024 and its rapid adoption through 2025 created a standard interface for plugging context providers into agent harnesses. This means a Company Brain product can ship as an MCP server and be consumed by Claude Code, ChatGPT Enterprise, Cursor, and every agent platform that ships in the next eighteen months without bespoke integration work. Garry Tan's gbrain ships as an MCP server already. This is genuinely new and changes the distribution math.
What has not changed in the right direction is buyer maturity. Enterprises are still buying enterprise search and AI assistants from name-brand incumbents because the buyer's career risk is asymmetric. Buying Glean is defensible. Buying a small studio's product to run on top of customer support data is not, until the studio has a logo wall the buyer can hide behind.
A Company Brain product that competes credibly in 2026 is not a side project. The minimum viable feature surface includes a connector layer that ingests at least Slack, Google Drive, Microsoft 365, Notion, Confluence, GitHub, Salesforce or HubSpot, Zendesk or Intercom, and Jira or Linear. That is roughly ten connectors before a customer says yes. Each connector requires permissions-aware retrieval, ACL hydration, change-data-capture for incremental indexing, and at least basic schema reconciliation across systems[25][30].
On top of the connector layer sits the structured-knowledge layer that distinguishes a Company Brain from a search engine. This is the executable skills file or context graph: typed entities (customers, products, accounts, incidents), typed relationships (owns, escalates-to, depends-on), and procedures (refund process, hiring loop, on-call rotation). Garry Tan's gbrain does this with markdown skill files and Postgres-backed graph storage. Lovelace's Elemental does it with knowledge graphs that compress queries from tens of millions of tokens to ten thousand[8]. Either approach works. The differentiator is the extraction quality and the freshness loop.
The third layer is the agent-facing interface. MCP server is the obvious shape. A REST API is the fallback. A web chat UI is required for the human-facing demo even if it is not the primary value prop. A skill-authoring interface, in the gbrain pattern, lets domain experts encode procedures in a form that is half markdown and half typed code.
Compliance posture for a tier-4 sale is non-negotiable. SOC 2 Type II, an executed BAA for HIPAA, GDPR data residency options, and a Trust Center page are all required before the first enterprise contract. Mid-market enterprises (200 to 1,500 employees) routinely budget $250K to $900K in year one for agentic infrastructure builds[27], but they will not write that check to a vendor that cannot pass procurement.
The GTM motion is sales-led, not PLG. Glean does not sell self-serve. Dust sells self-serve only into the long tail and books enterprise deals through a sales motion. Interloom went directly to Commerzbank and Volkswagen as design partners. Time to first paying customer, in this category, is six to nine months from product readiness, not six weeks. Time to first $1M ARR is twelve to eighteen months. Pricing converges on roughly $25 to $50 per seat per month for the mid-market wedge, with custom enterprise contracts above. Usage-based pricing on agent actions, the model Glean is now experimenting with, is the likely 2027 endgame.
Enterprise search platform with 100+ connectors, permissions-aware retrieval, and a knowledge graph layer powering "agent actions". Now positioned as the agent context layer for the Fortune 1000.
Munich-based context graph that mines emails, tickets, and transcripts to recover tacit knowledge. Reduced documented-vs-actual gap at Commerzbank from ~50% to ~5%. Won Zurich Insurance's company-wide AI competition for an underwriting use case.
Public "thin harness, fat skills" architecture. 34 skill files, MCP server, Postgres-backed graph. Beats vector-only RAG by 31.4 P@5 points on its own benchmark. Functions as the canonical reference architecture for the category.
The worst case, which is also the most likely outcome for a small studio entering this category cold, is the failure-by-incumbency case. The studio ships a product, gets some Hacker News attention, lands two or three design partners through warm intros, and then runs into a six-month enterprise security review that the studio cannot afford to fund. By month twelve the studio has $50K to $150K in pilot revenue, no SOC 2 Type II, and no path to a series A. The studio either pivots, gets acqui-hired by a buyer like Glean or Atlassian for talent value, or quietly winds down. Probability: roughly 50 percent.
The expected case is the vertical wedge case. The studio picks a single tight vertical, say mid-market customer support operations for SaaS companies between fifty and five hundred seats. The studio ships pre-built skills for the most common support workflows, integrates with the three help-desk tools the ICP actually uses, and prices at $20 per agent per month with a $1,500 floor. By month twelve the studio has fifteen to thirty paying customers, $300K to $700K ARR, and a credible path to series A on the strength of net dollar retention and verticalized differentiation. Probability: roughly 35 percent.
The best case is the kingmaker case. Wes finds a co-founder with deep enterprise sales credibility, lands two Fortune 500 design partners through that relationship in the first ninety days, ships a credible permissions-aware product in six months, and rides the YC RFS tailwind into a $5M to $8M seed at a $30M to $50M post by month twelve. ARR by month eighteen is $1.5M to $3M with three to seven enterprise logos. Probability: roughly 15 percent, conditional on the co-founder existing.
None of these cases are the Glean trajectory. Glean was founded by Arvind Jain after he sold Rubrik. The studio does not have a Rubrik to sell, and that matters more than any feature decision.
The thesis is right and the prize is real, but the moat is enterprise sales and compliance, not engineering. Without a vertical wedge and a sales partner, this is a six-figure ARR project against billion-dollar incumbents. Better to ship two narrower products that each carry a slice of this idea.