Introducing OM2: Your enterprise just started thinking

Today, AI is doing genuinely complex work inside most companies. Yet, ask any CIO how it's going, and you'll hear the same frustrations on repeat: their token costs are exploding, ROI is unclear, and there's a very long way to go until AI is able to truly understand and automate their business.

The problem: AI models don't learn

Models are mostly trained on generic data. They have no understanding of your organization. So, in the past few years, different ways have emerged to get company context into a model to help AI 'learn' about your business.

Each is broken in its own way. Picture for a second that howling screech you used to hear connecting to dial-up internet. That's about where things are today with enterprise context retrieval.

You'll probably be familiar with the most common approach: connecting your tools, like Slack, Jira, GitHub, Google Docs, and more, to Claude or your agent via MCP. That's useful for recall of specific data, but it means rebuilding your entire company context from scratch every session at runtime. It's slow, often misses critical information, and is extremely expensive because it dumps a lot of useless info into the model: over 50% of enterprise token budgets are spent on context retrieval alone. And any useful context that does accumulate via chat memory remains locked in Claude or ChatGPT.

Things get a little more complex from there: some companies have invested long hours consolidating and maintaining data lakes for vector RAG, while others have bought expensive traditional knowledge graphs like Glean. You could author a lengthy (and dull) book on the problems of vector RAG. It results in agents that reason confidently while never realizing they're missing half the data, hallucinations ($3m looks dangerously like a $30m with cosine similarity), an inability to see the absence of data (often equally useful signal), and permissions are difficult to enforce correctly. And while traditional knowledge graphs bring some much-needed structure (hello identity resolution and enumeration!), they rely on very coarse-grained indexing to figure out which raw documents to throw into an LLM at runtime. That means they still burn a lot of tokens, often on information that isn't quite right.

The fundamental problem here is that an LLM is not a human brain. Human brains are incredible. They've evolved to continuously take in new info, learn by rewiring themselves, and harden certain pathways to improve with practice. Where today enterprise AI greedily burns tokens, human brains are astoundingly efficient at context retrieval.

OM2 makes your existing AI 9x cheaper and 64% faster, all while being preferred on quality 84.5% of the time

To be clear, these improvements do not rely on model routing, nor do they require any change to your existing workflows. We measure them by comparing Claude using the OM2 MCP against Claude using off-the-shelf connectors, with statistical significance across hundreds of tasks. (In fact, when you do pair OM2 with our Optimized Routing, that 9x becomes 51x savings.)

So how are we driving such a significant improvement in cost, speed, and quality? In short, by making AI work much more like the human brain.

You can think of OM2 as a context graph. But because of its density and plasticity, we refer to it as a neural graph to distinguish it from traditional document-level knowledge graphs, which are orders of magnitude sparser. Connect OM2 to your apps and data, and it'll start learning. Continuously taking in new information, making connections, rewiring itself, hardening paths that are used frequently, and invalidating outdated information. Then, when you or an agent need to read or write across those apps, it traverses ever-evolving, pre-computed pathways to retrieve and execute on exactly the right info and actions, and nothing more. Because it's able to formulate structured queries against this dense data fabric, you get far greater accuracy, determinism, and predictability, all while being able to do things like enumeration, negation, counting, structural ranking, multi-hop traversal and more. All this, while enforcing source permissions on traversal and maintaining clear provenance chains to the underlying sources.

Whether you're working in Claude, ChatGPT, Perplexity, or Coworker's native apps, the end result is far more efficient, higher quality, and reliable AI.

The benchmarks

We benchmarked Claude paired with the OM2 MCP against Claude using its own off-the-shelf native connectors, holding the agent harness constant so the only variable was the context layer. The question set, 114 tasks across seven categories (business operations, engineering, general business, marketing, people, product, sales), was derived from real customer workloads, run against a representative connector set (Google Drive, Gmail, Google Calendar, HubSpot, GitHub, Jira, Slack, Notion, Stripe), and repeated to statistical significance. Quality was judged by blind human pairwise preference: reviewers compared anonymized, order-randomized answers against a correct, human-authored reference.

Note that none of the gains below include savings or efficiency that come from model routing; with Optimized Routing to frontier open models, the 9x savings becomes closer to 50x.

Where the learning gains come from: when Coworker solves a given class of nuanced retrieval tasks, constructing the right graph query, Jira JQL, GitHub search filters, or CRM query, it learns the path for future. Learning helps most where agents would otherwise iteratively rebuild queries each time, which is why the gains are largest on context-heavy workloads like Jira, GitHub, Slack and CRM. In production today, roughly 40% of conversational prompts already invoke learning successfully, and for automation-style agents that figure is above 90%.

Two details worth noting for the skeptics (we'd ask the same): on roughly 5% of questions, Claude's native-connector baseline couldn't answer at all. Those were excluded from cost and speed comparisons so the baseline wasn't unfairly deflated, but retained in the quality comparison, since not answering is itself a quality signal. And even with OM2 turned off, Coworker's MCP tool implementations alone still meaningfully outperformed the native connectors: the memory layer compounds on top of engineered-for-reliability tooling.

OM2 is built for the questions CIOs ask

• Fully permissioned. OM2 is a dense web of interconnected semantic units and data, and access policies strictly inherited from underlying sources travel with every fact and connection. That means everyone, human or agent, sees only what they're allowed to see and governance is built in from day one.
• Security in every atom. Your data stays in its own isolated environment, is never used to train models, and every answer traces back to its source. SOC 2 Type 2, GDPR, and CASA Tier 2 compliant with VPC available. OM2 runs on an isolated single-tenant infrastructure in our cloud or via VPC, with SOC 2, GDPR, and CASA Tier 2.
• Works with the AI you already use. OM2 connects via MCP or API. That means it works in Claude, ChatGPT, Gemini, Perplexity, your custom agents or anywhere else you're working.
• Look forward and back. OM2 learns backwards over historical data at setup, while continuously evolving as it sees more each day.
• Portable context. Your organizational memory, as well as the skills you create, now travel with you whatever open or closed model you choose, with no lock-in to a single provider.
• Always current, automatically. OM2 continuously reads and updates itself throughout the day, with no uploads, no tagging, and no re-syncing projects to maintain.

Already tested with customers

OM2 already powers every Coworker enterprise product, trusted by 300+ companies, and runs in production with customers including Harness, RapidSOS, Harri, and Column. These are organizations that need AI to deeply understand how their business works, while maintaining strict permissions and security.

We are huge proponents of Coworker. Before Coworker and OM2, keeping up with how fast the business moves meant our RevOps team spent hours stitching together Salesforce, Slack, and meeting notes to trust an answer. Now the answer is already there, current and correct, the moment we ask.

Anna Waring, Director, Revenue Operations and Systems, RapidSOS

What happens next

Connect the tools you already use and OM2 starts reading, with no schema to design, no tagging, and no migration. It builds itself in the background, and everything you point at OM2 draws on the same shared neural graph, which gets stronger with every day and interaction.

If you're curious to see what OM2 can do, come chat with us.

Benchmark methodology. 114 questions repeated until statistical significance and then replicated across 7 categories (business operations, engineering, general business, marketing, people, product, sales), run on a Claude harness against ~8 connected sources (Google Drive, Gmail, Google Calendar, HubSpot, GitHub, Jira, Slack, Notion, Stripe). Cost = total input + output tokens per session; speed = total end-to-end time to completion; quality = blind human pairwise preference against a human-authored reference answer, order randomized. Questions the baseline could not attempt (~5%) were excluded from cost/speed comparisons and retained in quality comparisons.