What AGI Would Actually Change for an Organisation That Keeps Its Data at Home

Artificial general intelligence is the most overloaded phrase in technology. Depending on who is speaking, it is a product launch next quarter, an existential threat, a fundraising line, or a distant philosophical curiosity. None of those framings helps a managing partner, an NHS trust's information governance lead, or a council's head of digital decide what to do on Monday morning.

So here is our attempt at a calmer view. We research the road to AGI ourselves, and we build private AI for organisations today. Both jobs push us towards the same conclusion: the more capable these systems become, the more it matters where they run and what they are allowed to see.

What we mean by AGI, and what we do not

We use a deliberately practical definition. A system is approaching general intelligence when it can take on most of the cognitive work a capable professional does, across many domains, and pick up new kinds of task without being rebuilt for each one. That is a description of usefulness, not of consciousness, and it does not arrive on a single date. It is a gradient, and organisations will feel it long before anyone agrees it has been crossed.

What we do not mean is a machine that is infallible, or one that should be trusted because it sounds confident. Fluency and reliability are different properties. Today's best systems are often fluent well beyond their reliability, which is exactly why the questions below matter.

What is genuinely here now

It is worth being precise about the present, because both the hype and the dismissal get it wrong. Agents can now complete real, bounded pieces of work: drafting, summarising, searching, writing and checking code, working through a defined process. METR's long-running measurements show the length of task an agent can finish reliably has been doubling roughly every seven months for years. That trend is real, and it compounds.

What is still missing is just as important. Long tasks remain fragile. Agents rarely know when they do not know. They struggle to weigh competing obligations the way an experienced professional does, and they can be wrong in ways that look exactly like being right. Capability is climbing quickly; judgement and dependability are climbing more slowly. Any sensible plan has to live in the space between those two curves.

What changes for organisations

If you take the trend seriously without swallowing the hype, five things change for an organisation that holds sensitive information.

The data boundary becomes the main question. A more capable system is more useful in direct proportion to what it can see. The temptation will be to give it everything: every share, every mailbox, every case file. Where that information goes, and who else can reach it, stops being a procurement detail and becomes the central decision. A capable assistant that sends your documents to someone else's servers is a capable assistant that has quietly moved your data boundary.

Permissions stop being an IT chore. When a system can read across the whole organisation, the only thing standing between a junior employee and a confidential board paper is whether the system respects the access controls you already have. Mirroring existing permissions exactly, rather than building a parallel set that drifts, becomes a safety property, not a convenience.

Audit trails become governance. If decisions start to lean on machine-generated answers, you need to show what was asked, what was retrieved, and what was answered. Regulators, auditors and courts will ask. An answer you cannot trace back to its sources is an answer you should not act on.

The cost model matters more. Pricing that charges per query or per token makes sense while usage is light. As capability rises, the work you want to hand over grows, and a metered bill grows with it. Running models on hardware you control turns a variable, rising cost into a predictable one.

People need a plan, not a press release. Capable systems change jobs. We do not think that should be left to chance. Our commitment is simple: automate the work, fund the worker. Every deployment we make supports reskilling and stipends for the people whose roles change, because the benefits of automation should not land on one side of the table only.

Why local matters more as capability rises

There is a common assumption that private, local AI is a stopgap: something you put up with until you can afford the frontier. We think the opposite is true.

The more an organisation relies on AI, the more it depends on whoever operates it. That dependence covers price, availability, terms of service, data handling and the choice of model itself. For many organisations, and almost all regulated ones, accepting that dependence for their most sensitive work is not a risk they can sign off. Local inference keeps the data, the model and the decision about when to change either of them inside the organisation.

We should be honest about the trade-off. Models that run on your own hardware will often trail the very largest models offered as paid cloud services. For a great deal of everyday knowledge work, that gap matters far less than people assume, because the hard part is usually finding the right document and staying faithful to it, not producing dazzling prose. A grounded answer with a source you can check beats an eloquent guess. And the gap itself keeps narrowing as open models improve.

There is also a quieter benefit. A system that runs inside your walls can be inspected, measured and constrained by you. You can test it on your own questions, decide what it may read, and switch it off without asking permission. As systems become more general, that kind of local control is worth more, not less.

Where Atheneum AI stands

Athena is the product we sell and support today. It searches your files, email and code in place, answers with sources, respects the permissions you already have, and keeps every query on your own infrastructure. It is not AGI and we do not pretend it is. It is the practical first step: a grounded, auditable assistant that makes your own knowledge usable without moving it anywhere. We publish how we measure Athena and the conditions behind every figure we quote.

Talos is our research programme on the road to AGI. We share our thinking openly, in posts like this one, and we keep the implementation to ourselves. That is a deliberate choice. Research into more general systems should be discussed in public; the details of how a particular system is built do not need to be.

What to do this year

You do not need a view on timelines to prepare well. The same steps help whether general intelligence arrives in three years or thirty.

1. Know where your knowledge lives. Map the shares, sites and mailboxes that hold what your people actually need. Most organisations are surprised by the answer.
2. Fix permissions before you automate. An AI system will faithfully expose any access control that is already wrong. Clean it up first.
3. Decide what must never leave. Write down which categories of information may not be processed outside your network, and hold every tool to that line.
4. Measure on your own questions. Vendor benchmarks tell you little about your documents. Build a small set of real questions with known answers and test against it, repeatedly.
5. Start with bounded, checkable work. Pick tasks where an answer can be verified against a source. Expand only as reliability earns it.
6. Plan for the people. Decide now how time saved will be reinvested and how affected roles will be supported.

The calm version of the future

The loudest voices on AGI tend to be selling either fear or certainty. We would rather sell neither. Capability is rising quickly and unevenly. Reliability lags behind it. The organisations that do well will be the ones that keep their data at home, insist on answers they can trace, control the systems they depend on, and bring their people with them.

That is not a prediction about a dramatic moment. It is a description of good practice that gets more valuable every month the trend continues. We intend to keep building for it, in the open where it helps and quietly where it should be.