How to Prepare for AGI

Most writing on AGI readiness has a structural problem: it asks you to prepare for an event with no agreed definition, no arrival test, and expert forecasts spanning from this year to 2050. That is not a plan, it is a posture.

Kapoor and Narayanan put the objection precisely: "If a company declares that it has built AGI, based on whatever definition, it is not an actionable event. It will have no implications for businesses, developers, policymakers, or safety." Helen Toner's version is that the term is now "almost useless." If they are right, and the definitional record strongly suggests they are, then any preparation that depends on AGI arriving is unfalsifiable.

So invert it. Prepare for measurable capability increments, which are arriving continuously and can be verified. Everything below pays off whether or not anything deserving the label ever shows up.

1. Redesign a workflow, do not just buy tools

This is the single best-evidenced finding in enterprise AI. McKinsey's 2026 global survey found 73% of high performers fundamentally redesigning workflows, against 25% of everyone else. Meanwhile 80% of individuals report productivity gains while the share of organisations reporting any EBIT impact sat flat at 37%.

Individual speed does not aggregate into company results on its own. Something has to change about how the work moves.

2. Solve the mapping problem first

A field experiment on 515 high-growth startups identified the real barrier and named it: firms cannot work out where in their own production process AI creates value. The intervention was not technology. Treated firms were given information about how other firms had reorganised around AI. They discovered 44% more use cases, completed 12% more tasks, were 18% more likely to acquire paying customers, and achieved 1.9x higher revenue.

An information-only intervention nearly doubled revenue. The constraint is knowing where to apply this, not affording it.

3. Go deep on few things rather than shallow on many

US Census Bureau data shows 57% of adopting firms use AI in three or fewer business functions, and 65% of task-level adopters restrict it to three or fewer tasks. Adoption is a mile wide and an inch deep, and shallow deployment is exactly what produces individual gains with no organisational result.

Pick the work you do repeatedly, where the output is verifiable and the cost of an error is survivable. Go properly deep there before adding a second front.

4. Measure cost per completed task, not per token

Inference cost per token fell more than 280-fold between late 2022 and late 2024. Cost per finished job has not followed, because reasoning and agentic workflows consume vastly more tokens per task. Epoch AI estimates token consumption growing around 10x per year against capacity growing 3.4x, and expects frontier prices to rise. McKinsey found 20% of organisations saying AI operating costs have actively limited adoption.

Two measurements will tell you more than any vendor benchmark: what it costs to get one verified, usable output, and what proportion of attempts need rework. Princeton's agent evaluation found the most expensive model on the cost-performance frontier in only one of nine benchmarks, and found that increasing reasoning effort produced equal or worse accuracy in 21 of 36 combinations. Spending more is not a strategy.

5. Test reliability the way your customers experience it

Anthropic's engineering team made this distinction public and it is the most practically useful idea in agent evaluation. pass@k asks whether any of k attempts succeeds. pass^k asks whether all k succeed. "At k=1, they're identical... By k=10, pass@k approaches 100% while pass^k falls to 0%."

Demos and benchmarks report the first. Customers live the second. Before anything touches a client, run the same task ten times and count how many times all ten were acceptable. That number is usually a shock, and it is the one that predicts whether you will be apologising later.

6. Get governance in place before you need it

The gap here is wide and measurable. Deloitte's survey of 3,235 leaders across 24 countries found only 21% with mature governance models for agentic AI. A separate survey of 418 IT and security professionals found 82% of enterprises had unknown AI agents in their environments, while 68% believed they had strong visibility. Only 21% had a formal process for decommissioning an agent.

For anything acting autonomously, two rules are worth adopting immediately. Simon Willison's lethal trifecta: access to private data, exposure to untrusted content, and the ability to communicate externally are fine in any two combinations and dangerous in all three. Meta's agents rule of two follows: an autonomous agent should have at most two, and all three requires a human to approve the action.

The dates that are actually real

Forget AGI timelines. These are on the calendar and enforceable.

• 2 August 2026, already in force: EU AI Act Article 50 transparency obligations. Disclosure that people are interacting with AI, machine-readable marking of AI-generated content, disclosure at the moment of contact rather than in fine print, deepfake labelling even without intent to deceive. Penalties up to €15 million or 3% of worldwide turnover.
• 2 December 2027: high-risk obligations for standalone Annex III systems, postponed from August 2026.
• 2 August 2028: AI embedded in regulated products.

Worth noting for anyone building agents: the 2026 omnibus that moved those deadlines does not treat AI agents as a distinct regulatory category at all. The gap is real and it will be filled eventually.

What this adds up to

Nothing on this list requires believing any particular AGI timeline. Redesigning a workflow, knowing where value sits in your own process, going deep on a few things, measuring cost per finished job, testing reliability the way customers experience it, and knowing what your autonomous systems can reach are all worth doing if capability plateaus tomorrow.

That is the test for any AGI-readiness advice worth taking: would you still do it if the curve flattened? If the answer is no, it was not preparation. It was a bet on a date nobody can name.