Track 5 · Go Further
From "I've tried it" to it being part of how you work.
Three roads lead out of here: use it well in daily life, use it well at work, or build something with it. They share a first mile. This page is that mile, and then the fork.
Pick your road
Three roads, with real price tags
Confident everyday user
Uses it naturally for life admin, learning, writing and thinking. Knows what it's bad at. Checks without being told.
Cost: ~20 min/day for a month.
Reachable by: anyone. No technical background at all.
The person at work who gets it
Has rebuilt a real part of their own job around it, can show the before-and-after, and is trusted to advise colleagues.
Cost: Road A, then 3–6 months of applying it to real work.
Reachable by: anyone with a job that involves words, numbers or process.
Builder or founder
Builds tools or a product on top of AI. Understands cost per use, reliability, and what stops a competitor copying it.
Cost: 6–24 months, real money, real risk.
Reachable by: fewer people than the internet implies. Read the honest section below first.
Every road starts with fluency and the checking habit. A founder who can't tell a good answer from a plausible one will build a product that confidently ships wrong answers to customers — which is the most common way these things fail.
Road A
Four weeks, twenty minutes a day
One job per week, five things in each. Tick as you go — saved on this device.
Week 1 · Fluency — get comfortable
Week 2 · Your own life — make it useful
Week 3 · Checking — the week that earns the trust
Week 4 · Make something — proof it's real
That last item is the deliverable. Anyone who can write those five lines from their own experience is no longer guessing in the dark — and is ahead of most people currently being told to embrace AI.
Road B
Finding the AI-shaped work in your actual job
Your employer said "use AI" and stopped there. Here is the rest.
The four task types worth handing over first
| Type | Looks like | Why it's first |
|---|---|---|
| Reformatting | Notes → minutes. Data → report. Long → short. One template → another. | The answer is already in what you supply, so there's little to invent and it's easy to check. |
| First drafts | Anything you currently stare at a blank page for. | You were always going to rewrite it. The AI's mediocrity costs nothing. |
| Reading piles | Long documents, many emails, transcripts, feedback, applications. | Volume is exactly where humans get expensive and inattentive. |
| Rehearsal and review | Practising the pitch, pre-mortems (Gary Klein's technique — imagine it has already failed, then explain how), "what will they object to?" | Zero risk, and nobody else's time is spent. |
If you can't do the task yourself, you can't tell when the AI does it wrong — so you've automated a step you can no longer supervise. Start where you're competent. Expand outward as your checking improves, not as your confidence does.
How to propose it without it going badly
Check the rules first — before anything else
Which tools are approved, what data is allowed in them, and whether clients or patients must be told. Do this first. Quiet use of an unapproved tool on real work can already be a policy breach or a data leak — and in law, medicine, accounting and finance, confidentiality duties are obligations, not preferences. Getting this wrong is also the fastest way to get AI banned for everyone.
If your company has no rules yet — an "embrace AI" announcement and nothing else — that is not permission, it is a gap. Ask the person who owns data or IT, in writing, one narrow question: "Is there an approved tool for this? If not, may I use [tool] on non-confidential work while we decide?" You get an answer or you get on record asking — both protect you. In a company with no policy, the person who asks the question in writing often ends up shaping the policy.
Then prove it on your own work
One task, one month, timed before and after. An unarguable before-and-after beats any amount of enthusiasm.
Lead with the checking, not the speed
"Here's how I verify the output" is what makes managers comfortable. "It's so fast" is what makes them nervous.
Give away the method
Hoarding it buys you a few months. Teaching it makes you the person the organisation asks — which is worth considerably more, for longer.
The uncomfortable part
What this honestly does to your job
Nobody knows how this lands in ten years, and anyone who says they do is selling something. The direction of travel is visible enough to plan around.
Getting less valuable
- Producing competent first drafts
- Formatting, tidying, transcribing
- Knowing facts others could look up
- Routine boilerplate of any kind
- Being the only person who can operate the tool
Getting more valuable
- Knowing which question is worth asking
- Telling a good answer from a plausible one
- Deep domain knowledge — the thing that lets you check
- Taste, judgement, and taking responsibility
- Everything physical, relational, or requiring accountability
Every skill in the right-hand column is built by doing the work in the left-hand column for years. If juniors never write the bad first draft, the pipeline that produces the judgement stops. If you're early in a career, use AI to go further, not to skip the part where you learn to tell good from bad — that part is the career.
The reasonable expectation isn't "AI takes your job" or "nothing changes." It's that the mix inside your job shifts, faster in some fields than others, and that the people who stay useful are the ones who moved deliberately rather than waiting to be told. Which is what the previous four pages were for.
Road C
Starting an AI company, with the awkward parts included
It has never been cheaper to build a working product. That is exactly why building one is no longer the hard part — and why most of the advice you'll see is optimistic in the wrong places.
01A thin wrapper around a model is not a business
If your product is a text box that passes the request to someone else's model with a clever instruction attached, then a competitor rebuilds it in a weekend, and the model provider may ship it as a feature for free. That's not a reason not to start — it's a reason not to stop there.
What actually holds: access to data nobody else has; a workflow so deep in a specific job that switching is painful; distribution and trust in a market that's hard to enter; being the one who carries the regulatory or liability burden customers won't carry themselves.
02Every use costs you money — this isn't normal software
Traditional software has near-zero cost per extra user. AI products don't: every answer burns computation you pay for. A free tier can lose money on your most enthusiastic users, and the heaviest users are often the ones you most wanted.
Work out early: what one typical customer costs you per month to serve, and what happens to that number if they use it ten times more than you assumed. Do this before you price anything, not after.
03The demo is the easy part
An impressive demo can take a weekend. Almost all of the remaining work is what happens when it's wrong in front of a paying customer: how you detect it, how the product behaves, who's accountable, what you promise, and how you improve it without breaking what already worked.
Teams routinely underestimate that second part, because the demo felt like the product.
04Do it manually for five customers first
The cheapest validation there is. Find five people with the problem and solve it by hand — you, an AI chat window, and an afternoon each. No product, no code.
- If nobody will let you do it free, they won't pay for the automated version.
- Doing it by hand teaches you the actual workflow, including the messy parts that would have broken your design.
- You find out what "good enough to be trusted" means in that job — which is the number your whole product has to hit.
05Build on shifting ground, deliberately
Models, prices and capabilities change on a timescale of months. Two consequences worth designing for from day one:
- Don't marry one provider. Keep the model swappable, and keep your own set of test cases so you can tell whether a swap actually improved anything.
- Assume the base models get better. If your entire value is a gap in the current model's ability, the next release may close it. Build where improvement helps you rather than erases you.
06You are accountable for what it outputs
"The model said it" is not a defence to a customer, a regulator or a court. If your product gives advice, handles money, touches health, or makes decisions about people, get real advice about your obligations before you launch, not after.
It is also, incidentally, hard to copy. Being the company willing and able to stand behind the output is a real advantage, not just a cost.
07The unglamorous version usually wins
The durable opportunities are rarely the exciting general-purpose ones. They're narrow, specific and slightly boring: one industry, one workflow, one painful hour of someone's week that you know intimately and your competitors don't.
Which means your existing career is an asset, not a handicap. The AI expertise is learnable in months. The twenty years of knowing exactly where a particular job hurts is not.
Staying current
How to keep learning without drowning in it
No reading list here, on purpose: links rot and a stale list looks authoritative. What doesn't rot is the routine.
- Use it weekly on real work. Everything else is optional. Nothing substitutes for this.
- Re-run your five prompts each quarter. That's your entire "keeping up with model news" obligation.
- Read the model makers' own documentation when you want the truth about a product, rather than commentary about it.
- Follow two or three people who show failures, not highlight reels. Anyone whose examples always work is showing you a demo.
- Repeat your calibration test on your own area of expertise twice a year. Your working assumption about how often it's wrong should move as the tools move.
- Teach someone. The fastest way to discover the parts you only think you understand.
The goal was never to become an AI expert. It was to stop guessing in the dark — to know roughly what's possible, to get real help from it, and to be able to tell whether what you got back is any good. That's a normal, achievable competence, and it's most of the value on offer.