Artificial General Intelligence
Artificial general intelligence (AGI) is a proposed form of AI able to learn, reason and perform across the full range of cognitive tasks at a broadly human level — rather than one task at a time, as every deployed system does — and, in the labs' stated missions, the goal they are building toward. The term has no agreed definition; for business the practical meanings are the labs' timelines, their contracts and the capability trajectory of rented models.
- Also known as
- AGI, General AI, Human-level AI, Strong AI
- Used at
- the labs' missions and safety policies, the AI-safety and alignment fields, policy debates and the EU AI Act's systemic-risk tier, investment theses, and the contractual definitions that decide, for example, when OpenAI's obligations to Microsoft change
- Binding
- nothing legal attaches to the term; the regulatory proxies are training-compute thresholds (10²⁵ FLOPs in the EU AI Act) and the labs' voluntary frontier-safety frameworks, which gate deployment on capability evaluations; the Microsoft–OpenAI agreement's AGI clause is the best-known contractual definition, tied to a board determination
- Typical definitions in circulation
- economic (outperforms humans at most valuable work), cognitive (matches humans across tasks including learning new ones), test-based (passes a battery of benchmarks or a "Turing test" variant), and levels (Google DeepMind's 2023 paper proposes emerging, competent, expert, virtuoso and superhuman levels, judged by breadth and depth); current frontier models are placed at "emerging" general and "expert" or higher on narrow tasks by that scheme
- Typical timelines offered
- lab leaders have stated horizons from a few years to a decade; surveys of researchers give medians of decades; the forecasts have shortened since 2022 and remain disputed
- Parties
- the labs (whose missions, safety policies and marketing all use the term), the safety researchers, the policymakers, investors pricing the labs, and, at a remove, every company whose vendor is pursuing it
How the idea works in practice
Today's deployed systems are narrow: a fraud model scores transactions, a language model writes and reasons about text, a vision model reads documents. A frontier language model is general in a limited sense — it handles a very wide range of text tasks it was not specifically trained for — and narrow in others: it does not learn from experience after training, act in the world without tools, or reliably do what a competent person does across a job. The AGI question is whether scaling and new methods close that gap, and when. The labs say yes and soon; their frontier-safety frameworks are built on the possibility; their fundraising is priced on it. The sceptics point to the gaps that remain and the history of forecasts. The 2024–2025 reasoning models and agents narrowed some gaps — long tasks, tool use, code — and left others.
For a business the term is best translated into three questions. First, what can the rented models do this year and next, on the company's own tasks — an evaluation, not a forecast. Second, what does the vendor's pursuit of AGI mean for the relationship: the capability and price can change quarterly, the vendor's safety policy constrains what the product may do, and the vendor's own products may compete with the company's. Third, what would a step change do to the business model — a lender whose underwriting is a model, a support operation that is an assistant, a software company whose product is a workflow that an agent could perform — and whether the plan needs a scenario for it. For AI M&A, AGI is the thesis behind the labs' valuations and the licence-and-hire deals for the teams that might reach it; for the application layer, it is a reason the moat must be data, distribution and workflow rather than the model.
For a fintech, AGI is a vendor's mission statement and a scenario for the strategy offsite; the models that exist are evaluated, the models that might exist are planned for, and neither is bought on the word.
Example
A fintech's board asks for a view on AGI. The CTO's paper defines the term's uses, translates it into the three questions, and reports: on the company's evaluation sets, the current frontier models automate 60% of document work, 40% of support conversations and 25% of engineering tasks, up from 30%, 15% and 5% two years earlier; the vendors' capability roadmaps and pricing have changed every quarter; and a step change to reliable agents across the company's workflows would cut operating cost by a third and expose the product to competitors who could build it faster. Actions: an evaluation programme that re-measures capability quarterly; vendor diversification; a data strategy to make the company's proprietary data the asset a step change would not erase; and a scenario in the three-year plan for agent-driven operations. No forecast of AGI is made, and the board approves the paper on that basis.
Common mistakes
- Planning on a date for AGI rather than on measured capability. - Reading a lab's mission as its product's capability today. - Treating the term as defined; it is not, and contracts that use it define it themselves. - Building a business whose moat is the model, which the pursuer of AGI supplies to everyone. - Ignoring the vendor-relationship consequences of a supplier pursuing a step change.