Resident AI is AI that lives inside a company and works under its authority. Here’s how it works

Every few months the industry hands out a new name for AI inside companies, and most of them describe a product rather than a problem. This one is ours, so read it with the usual suspicion. Ran starts with what the term means, says plainly where other people already use it for something else, and spends a surprising amount of the piece on when a company does not need it. That last part is the test of whether a definition is honest. It opens a ten-part series that runs across every column in the magazine. — Muximus

Most company AI today knows the world and visits the company. Resident AI is the other way around.

A chatbot, a copilot or an agent arrives with a model trained on the public internet. It learns about the company one prompt, one upload or one connector at a time. When the session ends, most of that understanding goes with it.

Resident AI is built from the opposite end. It starts with the company - its people, decisions, commitments and rules - and lets the model come and go.

Here is the definition we use across this series:

Resident AI is an AI capability that operates within an organization's environment, understands its institutional context, and works under organizational authority and governance.

Three parts, and each one does work. Within the environment is about where it runs and what it can see. Institutional context is about what it knows. Organizational authority is about whose rules it follows.

The name is not ours alone

Before going further, a fair warning about the word.

In December 2024, a company called ObviousFuture launched a product it calls "Resident AI". In its usage, the term means AI that runs on a company's own hardware instead of in the cloud. Property-management software also uses "resident AI" for bots that answer tenants, which is a different thing entirely.

Our use overlaps with the first one and is broader. Where the AI runs matters, but it is one property, not the whole definition. A model running on a server in the basement can still know nothing about the company. A system in a governed cloud tenancy can know a great deal. Location is not the point. Knowledge and authority are.

Access is not understanding

Most enterprise AI projects connect a model to company data and call it done. That gives the model access. It does not give it understanding.

Anthropic's engineering team published a small example of the difference in 2024. A typical document chunk might say: "The company's revenue grew by 3% over the previous quarter." It names neither the company nor the quarter. Pulled out of its document, the sentence is accurate and almost useless.

Now scale that up to the things that actually run a business:

• The process doc says Sarah approves invoices. She left in March.
• The CRM says the renewal is at list price. The discount was agreed on a call.
• The policy page says no exceptions. The COO approved three last quarter.

Each record is accurate. The company still does not work the way the records say it does.

Every company has two operating models: the documented one and the real one. The documented one lives in systems of record, process docs and org charts. The real one lives in email threads, meetings, chats, exceptions and in the heads of people who "just know how things work". Access gives AI the first. Resident AI is an attempt to give it the second.

Five properties

A system qualifies as Resident AI when it has all five of these. Most products on the market today have one or two.

1. It lives in the company's environment. It runs where the company's data and identity live - on premises, in a private cloud, or in a tenancy the company controls. It observes the sources the company agrees to, and only those.

2. It keeps institutional context, not just documents. People, roles, authority, decisions, commitments, relationships, processes and dependencies. And their history: what was true in March, and what replaced it.

3. It works under organizational authority. It knows who is asking and what that person may see. Permissions are resolved before anything reaches a model, not filtered afterwards.

4. It shows its evidence. Every answer traces back to a source: who said it, where, and when. When the evidence is missing, it says "I don't know" instead of guessing.

5. The company owns it. The knowledge belongs to the company, separate from any model vendor. The model doing the reasoning can change without the company starting again.

We sum up the last one as own the knowledge, rent the thinking. Models are becoming interchangeable compute. A company's understanding of itself is not.

Organizational memory vs agent memory

This is the distinction people mix up most, so it gets its own section. (Penny Layne's full explainer goes deeper.)

Agent memory is what an assistant keeps to do its job for a user: preferences, past chats, the state of a task. ChatGPT, Claude and most agent frameworks now have some version of it.

Organizational memory is institutional knowledge that lives beyond any single conversation, agent or model session. It belongs to the company, not to a user's account.

The two differ in four ways:

Ten people using an assistant with good memory still produce ten partial pictures of the company. None of them is the real one.

How it differs from what companies already have

None of these are bad tools. Copilots are very good at helping one person in one suite. RAG is very good at finding documents. The point is that each one solves a narrower problem.

The architecture, in principle

There is no standard reference design yet. Here are the principles we think matter, and the trade-offs that come with each. We build to them in our own product, Acropolis, but none of them require it.

Observe instead of waiting to be fed. The record should come from the work itself - email, meetings, tickets, the CRM - not from someone remembering to write it down. Trade-off: observation needs permission, trust and clear rules about what is out of scope. Many employees will reasonably ask whether this is monitoring. It has to be designed so the answer is no, and the company has to say so before it starts.

Keep a model of the company, not a pile of text. Store claims about the company - "Dana owns the Acme renewal, since June 3, per this email" - with their source and the time they were true. We call this a Living Organizational Model (LOM). That is our name. Others are building similar things under labels like "context graph" or "context layer". In May 2026, Gartner told its data and analytics summit that agent accuracy depends on context, including semantic representations of data. Trade-off: extracting claims is harder and costlier than indexing chunks, and extraction makes its own mistakes.

Govern at retrieval. Resolve identity and permissions before anything reaches the model. A model that never receives restricted information cannot leak it in an answer. Trade-off: this needs good identity data, which many companies do not have on day one.

Let the model be replaceable. Expose the company's knowledge through a governed interface (MCP is one common option) so any capable model can reason over it. Trade-off: you give up some convenience of an all-in-one vendor suite.

Act only after understanding. Agents that act on the company's behalf should come after the system understands the company, with written limits on what each one may do. In our own process that observation period runs 60 to 90 days. Other approaches may move faster or slower.

Where Resident AI is the wrong answer

This list matters more than the rest of the guide. A category that claims to fit everyone is a sales pitch.

• The problem is individual productivity. Writing, analysis, research, personal workflows. A good assistant does this well and costs a few dollars a month per person.
• The workflow is simple and repeatable. A few connectors and an automation tool such as n8n may be enough.
• The knowledge fits in a prompt. Anthropic's own guidance says a knowledge base under about 200,000 tokens - roughly 500 pages - can go straight into the prompt with no retrieval at all.
• The company is small. If ten people share an office and talk daily, the documented and the real company are probably close already.
• The problem is ownership, not information. If nobody owns a process, no AI will fix that. It may just describe the gap more precisely.

There is one more honest limit. Nobody, including us, has published independent, long-term evidence on how well systems like this hold up across many companies. Until someone does, every claim in this category - ours included - is a claim, not a body of proof. This series will report what we find, including when it goes wrong.

A quick test

Five questions decide whether a company is in Resident AI territory. Answer them about one part of the business, not the whole company.

1. Can leadership list the commitments made to its five most important clients this quarter, and who owns each one?
2. When a decision changes, does every team that depends on it find out within a week?
3. If two people ask an AI tool the same company question, do they get the same answer?
4. Is there knowledge that only one person holds, and would it leave with them?
5. If the main AI vendor changed its prices or retired its model tomorrow, what company knowledge would be lost?

Two or more "no" or "we don't know" answers mean the gap between the documented and the real company is wide enough to matter. Fewer than that, and a simpler tool is probably the right call.

The rest of the series

Each piece stands alone. Read in order, they build the case and then test it.

• Why a model trained on the internet doesn't know how the company runs - Dear Humans
• Where RAG works, and where it stops being memory - Overhyped
• Keeping company knowledge portable when a vendor holds the memory - First Opinion
• Agent memory vs organizational memory - Dear Humans
• Three public AI tools that stated rules confidently and wrongly - Proof of Work
• What "I don't know" does to an AI's error rate - Proof of Work
• How much company AI work needs a frontier model - Overhyped
• Chatbot or change in how the company runs - First Opinion
• What a Resident AI installation looks like - Founder Mode

Disclosure: VarOps builds Acropolis, a Resident AI system, and Ran, the founder, also builds MUXI, agent infrastructure. This guide is meant to be useful without either.

Part of the Resident AI series.

From thinking to doing.

From founder thinking to company building.

We don’t just write about AI-native companies. We build them. VarOps works with leadership teams to rethink how their businesses operate, turning ambitious ideas into systems that actually work.

Build What’s Next →

Sources

• ObviousFuture, "ObviousFuture disrupts industry with faster and safer Resident AI", Business Wire, December 18, 2024: https://www.businesswire.com/news/home/20241218465156/en
• Anthropic, "Introducing Contextual Retrieval", September 19, 2024: https://www.anthropic.com/research/contextual-retrieval
• Gartner, "Gartner Says Lack of Semantics Causes Inaccurate AI Agents and Wasted Spending", May 11, 2026: https://www.gartner.com/en/newsroom/press-releases/2026-05-11-gartner-says-lack-of-semantics-causes-inaccurate-artificial-intelligence-agents-and-wasted-spending
• Anthropic, "Bringing memory to Claude", September 11, 2025: https://claude.com/blog/memory