Last spring I watched a COO ask her leadership team one question: what changes if artificial general intelligence arrives in 2029 instead of 2045? The room went quiet. Someone made a Terminator joke. Then everyone returned to the budget spreadsheet and nothing changed.
That reaction is normal, and it’s expensive. You don’t need to forecast AGI to prepare for it. You need a process that turns a contested, foggy idea into decisions you can actually make this quarter — hiring plans, data investments, product roadmaps. Here’s that process, in five steps, with enough detail to run it yourself on a Thursday afternoon.
Step 1: Break job titles into task lists
AGI forecasts are useless at the level of “will this replace lawyers.” They get useful when you shrink the unit of analysis down to a single task.
Pick one role and write down what that person genuinely does in a week. Not the job description — the real work. A paralegal might spend six hours redlining contracts, four hours chasing signatures, three hours writing client updates, and two hours deciding which clause is odd enough to flag for a partner. Those four items have completely different automation timelines.
For each task, ask three questions:
- Can success be verified cheaply? Grading a translation is easy. Judging whether a strategy memo is good is not.
- How much context does it need? Answering a billing question requires one account record. Advising a founder on whether to sell requires years of relationship history.
- What happens when the system is wrong? A bad subject line costs nothing. A bad dosage calculation costs everything.
Tasks with cheap verification, narrow context, and low error costs fall first. That ordering alone will tell you more than any timeline chart. If the terminology itself is tripping people up in your meetings, this breakdown of the two meanings of general AI is worth ten minutes before you start.
Step 2: Build a capability delta table
Now put numbers on it. Four columns: task, hours per week across the team, current AI performance, and the gap between them.
Be blunt about performance. A 40-person support team handling 12,000 tickets a month might find a current model drafts an acceptable first reply about 70% of the time on billing questions and roughly 35% of the time on anything involving a policy exception. Those two numbers imply opposite staffing plans, and most teams never separate them.
If you want a realistic picture of what today’s commercial models can and can’t do before you fill in that column, a rundown of what OpenAI’s 2025 models actually ship with is a reasonable starting reference.
The table’s value isn’t precision. It’s that it forces you to name the 20% of work where a jump from 70% to 95% would wreck your headcount math — and the 80% where it wouldn’t move a thing.
Step 3: Run a 90-minute scenario exercise
Set a specific, inconvenient date and describe it plainly: It’s March 2029. A system has just cleared a broad reasoning benchmark and can pick up unfamiliar white-collar tasks from a two-page brief with roughly human reliability, at a tenth of the cost.
Split into groups of four and give each group twenty minutes on one question, then rotate.
What breaks first?
Usually it’s not the product. It’s the pricing model, the onboarding process, or the fact that your entire sales motion depends on being the only person in the room who understands the client’s data. A legal software company I spoke with realised their moat was a two-week implementation service that a general system would make pointless overnight.
What gets dramatically cheaper?
Name the specific line items. Drafting, translation, first-pass review, market research, code scaffolding. Then ask who else in your market gets the same cost collapse at the same moment. If the answer is “everyone,” cheaper isn’t an advantage — it’s the new floor.
What gets more valuable?
Almost always: accountability, physical presence, licensing, taste, and relationships people can’t easily verify with a model. A surgeon’s signature and a notary’s stamp are worth more in this scenario, not less. So is a ten-year client relationship nobody can fabricate.
If your group keeps drifting into speculation about whether AGI is even possible, the arguments in this look at the race to build machines that truly think give you a shared vocabulary to argue with instead of past each other.
Step 4: Make three bets, one per time horizon
The exercise is theatre unless it produces commitments. Cap it at three.
- Next 90 days: build an evaluation set for your riskiest workflow. Fifty real examples with known good answers. Without it, you’ll never know whether a new model is better or just more confident.
- Next two years: own something that can’t be scraped. Proprietary outcome data, a licensed position, a distribution channel, a certification. This is the only defensible answer to a world where raw capability is a commodity.
- Next five years: decide which human layer you’re doubling down on. Trust, judgement under ambiguity, physical execution, or regulation. Pick one and put real money behind it.
Write each bet with an owner and a date. A bet without a name attached is a wish.
Step 5: Pick signal sources and ignore everything else
Most teams fail here, not at the workshop. They drown in AGI commentary and end up reacting to whichever headline was loudest that week.
Choose five to eight sources and check them on a schedule. A useful filter for sorting genuine reporting from recycled speculation is this guide to telling real AI coverage from the noise. For the commercial side — who’s actually shipping and who’s mostly raising money — this overview of which AI companies are leading the market is a sensible quarterly read.
Then set one recurring calendar block, 45 minutes, once a quarter. Same table, updated numbers, new bets if something genuinely shifted.
What it looks like for a 12-person accounting firm
They ran this in a single afternoon. Task audit found that 30% of staff hours went to reconciling and categorising transactions, work that current tools handle at maybe 80% reliability with a human check. Another 25% went to tax planning conversations with long-standing clients, where no one could see a path to automation within a decade.
Their three bets: a 90-day project to fix the reconciliation review step so one person could supervise three times the volume; a two-year push to collect structured outcome data from every planning engagement; and a five-year commitment to advisory work and local reputation. Total cost of the exercise: one afternoon and a stack of sticky notes. Cost of not running it: hiring two junior accountants in 2026 for work that may not exist in 2029.
Start with the role that scares you most
The instinct is to begin with the easy, low-stakes work. Resist it. Start with the role whose disappearance would genuinely threaten the business, and make the first task list yourself rather than delegating it. The discomfort is the point — it’s where you’ll find the one assumption that’s been quietly holding your strategy together, and the sooner you find it, the more options you have when it finally breaks.