Everyone Uses AI Now. Almost Nobody Has Built It In — Write Your One-Page AI Strategy

Two out of three businesses use AI. One in seven has built it in.

Ask a room of small business owners who uses AI and nearly every hand goes up. Someone drafts emails with it, someone else answers website chats with it, a third has an AI feature switched on inside the billing software. That is real usage — and it is also where most of it stops. The tool is in the business, but the business does not run on it. The work it was bought for still happens the old way, by hand, when someone remembers.

Two surveys published in October 2026 put numbers on that gap, and a third — McKinsey's global survey from a year earlier — shows the same shape at enterprise scale. They agree on the shape of it even though they measured different things:

Read the Goldman Sachs numbers slowly, because they are the sharpest version of the story. Among the alumni surveyed, 86% use AI in their business — up from 56% just two years ago. But only 15% say AI is fully integrated into their core operations. Among the owners who use AI, only 18% have it fully integrated; the rest use it for selected activities or experimentation.

The report's own line is the honest one: "High adoption does not mean deep integration."

The Chamber's survey tells the same story from the other side: two in three small businesses now use AI, nearly triple the 23% of 2023. And McKinsey's global survey of larger organisations shows the gap is not a small-business problem: 88% use AI in at least one function, nearly two-thirds have not yet begun scaling it across the enterprise, and only 39% report any impact on enterprise-level profit (EBIT) — most of them under 5% of EBIT.

Three honest caveats belong with these numbers, and the post keeps them. The Goldman Sachs figures come from alumni of one programme, not a representative sample of all small businesses. The Chamber counts broad AI use — anyone who uses an AI feature counts. McKinsey's figures are enterprise-scale and self-reported.

The numbers show a direction, not a precise scoreboard — and the direction is the same everywhere: using AI is now normal; running on AI is still rare. Two of the three surveys are American and McKinsey's is global; treat them as the shape of the gap and check the pattern against your own business — tools bought, nothing integrated — rather than treating any percentage as your own.

So the question is not whether your business uses AI. It almost certainly does. The question is whether AI owns any part of how the business runs — and if not, what would have to change for it to. That is what a strategy is for, and it fits on one page.

Using AI is not a strategy

Most businesses that "use AI" have done one of three things, and none of them is a strategy.

• They bought a team of tools — a chatbot for the website, a drafting tool for emails, an AI feature inside the accounting software, a meeting summariser, a social media caption writer. Each one does something small. Nobody owns the whole set, nobody knows what it costs in total, and nobody can say what any of them is for beyond "it helps." When the novelty wears off, the subscriptions stay and the work goes back to the old way.
• They measure usage instead of outcomes — the dashboard says the chatbot handled 400 conversations this month. That sounds like progress until you ask the question that matters: did any of those conversations become a customer, a booking, a payment? Logins, chats opened, prompts typed — these are measures of the tool being touched, not of the business being helped. A tool can be used every day and still change nothing.
• They never set a review date — the tool was switched on in March with a vague hope attached to it. Nobody wrote down what "working" would look like, and nobody picked a date to check. So it drifts: still on, still paid for, still unjudged. A strategy without a review date is a wish; a tool without a review date is a subscription.

A real strategy does the opposite of all three. It picks one job, not a shelf of tools. It names one number that proves the job is being done better, not a dashboard of usage. And it sets a date — thirty days out — when the owner looks at the number and decides: keep it, expand it, or stop it. That is the whole document. One page, four lines.

The one-agent lesson for your AI strategy

The engineering world arrived at the same discipline from the opposite direction, and its timing is useful: on 22 October 2026, Anthropic's Applied AI team published guidance for companies building agents on its platform, and the message pointed in one direction — start with one agent, and add more only when a test says the simple version falls short.

In Anthropic's comparisons across several enterprise deployments, "a single agent with skills frequently outperformed both the one-prompt-for-everything design and the subagent design on quality, and often at a lower cost and latency per task" (Anthropic). One customer that tested a multi-agent harness against a single agent loop "saw a roughly 2x faster runtime with the latter."

The pattern Anthropic kept seeing was familiar: each capability is owned by a separate team, so each team builds its own agent, and a platform team builds an orchestrator on top — and the agent architecture becomes "a copy of their org chart." Its feedback on those over-built designs is often that "one clear system prompt, 4-5 thin data tools, and firm output guardrails would cover most of the use case."

Anthropic had said the same thing nearly two years earlier. Anthropic's December 2024 essay "Building effective agents" (Erik Schluntz and Barry Zhang) opens with the finding that, after working with dozens of teams, "Consistently, the most successful implementations weren't using complex frameworks or specialized libraries. Instead, they were building with simple, composable patterns." Its advice is to find "the simplest solution possible, and only increasing complexity when needed" — which "might mean not building agentic systems at all" for many jobs.

Anthropic's method for deciding is the part that translates directly into a business strategy. Build a small end-to-end test first — "even a set of 10 golden examples is enough to start" — then change one thing at a time and keep the change only if the test scores improve. When folding subagents into the main loop, keep a subagent only if the folded version scores worse, and run each task more than once, since a break in their own test "only showed up in one run out of two."

Subagents still have their place — genuinely parallel work, measurable performance degradation from context bloat, or a tool list so long (past roughly 30 to 50 definitions) that the model starts picking the wrong tool — but they are the exception, earned by measurement, not the starting point.

Now read that guidance as a business owner, not a builder. One agent instead of a swarm becomes one job instead of a shelf of tools. "The simplest solution possible" becomes: the cheapest test that could prove the job is done better. "10 golden examples" becomes: ten real enquiries, ten real invoices, ten real cases — the test set your business already has. "Keep it as a subagent only if the folded version scores worse" becomes: keep the tool only if the number at the review date says it worked.

The engineering discipline and the business discipline are the same discipline, at different scales. That is why the strategy fits on one page: the people who build these systems professionally have learned that the winning design is the smallest one that passes its test.

The build: write your one-page AI strategy

This is the build: one evening, a Google Doc, and one real job from your business. The worked example below uses one common job — answering enquiries after hours — but the four lines work for any job: invoice chasing, appointment booking, quote follow-ups, stock questions. Pick the job that repeats most often in your week, and write the page for that one.

Step 1: Pick the one job

Open a blank Google Doc and write, at the top, the job in one sentence: the work, who it is for, and when it happens. Not "use AI for customer service." Something a person could actually do: "Answer every enquiry that arrives on WhatsApp after 7 pm, and book the caller in for a visit."

The test for a good job: it repeats regularly, it has a clear end (an answer sent, a visit booked, a payment chased), and you can tell in thirty seconds whether it was done well. Enquiry answering passes. "Improve our marketing" does not. If two jobs compete, pick the one with more repetitions — volume is what makes the number in step 3 meaningful.

Step 2: Name the owner

Write the name of one person who owns the result. Not a department, not "the team," one person — usually you, the owner, for the first page. The owner's job is not to do the work; it is to check the number on the review date and make the keep-or-stop call. A strategy with no named owner is a document nobody reads.

Anthropic's version of this rule is written for a large company — "one central team should own end-to-end quality and the system prompt" — and in a small business that team is one person: the owner, for the first page.

Step 3: Pick the one number

Write the single number that proves the job is being done better, and how you will count it today, before AI touches anything. For after-hours enquiries, the number is "enquiries answered within 10 minutes, counted weekly" — and the baseline is what you count this week by hand: how many arrived after hours, how many got an answer before morning. For invoice chasing, it is "invoices paid within 7 days of the reminder." For appointment booking, "bookings completed without a phone call."

Two rules for the number. First, it must be countable today, by hand, in a notebook — if you cannot count it now, you will not be able to count it in thirty days either. Second, it must be an outcome, not usage: "enquiries answered" is an outcome; "chats the bot opened" is usage.

Ten real cases are enough to start — Anthropic's "10 golden examples" is the same idea: a small, real test set beats a large, imagined one.

Step 4: Set the review date

Write the date, thirty days from today, and put it in your calendar now, with an alarm. On that date you will look at the number and make one of three decisions: keep it (the number moved, carry on), expand it (the number moved a lot, give the same worker the next job), or stop it (the number did not move, switch the tool off and write down why). The date is the fourth line of the strategy, and it is the line that makes the other three honest. Without it, step 3 is a hope and step 1 is a purchase.

Step 5: Copy the template and fill it in

Here is the whole page. Copy it into your Google Doc and fill in the blanks for your job:

ONE-PAGE AI STRATEGY — <your business name> — <today's date>

1. THE JOB
   One sentence: what work, for whom, when.
   Example: Answer every enquiry that arrives on WhatsApp after 7 pm,
   and book the caller in for a visit.

2. THE OWNER
   One person who owns the result and makes the review-date call.
   Name: ______________________

3. THE NUMBER
   The one outcome that proves the job is done better.
   Number: ______________________ (e.g. enquiries answered within 10 minutes)
   Counted how: ______________________ (e.g. tally in a notebook, weekly)
   Baseline this week (before AI): ______________________

4. THE REVIEW DATE
   Date (30 days from today): ______________________
   On that date, decide one of three: KEEP / EXPAND / STOP.

NOT ON THIS PAGE
   - No second job. One job, one page.
   - No usage numbers (logins, chats opened). Outcomes only.
   - No tool names as the goal. The tool serves the job, not the other way round.

That is the entire strategy. If your page runs to two pages, you have written a plan, not a strategy — cut it back to the four lines.

Step 6: Give the page to a chatbot and ask it to argue with you

Paste the filled-in page into the chatbot you already use, with this prompt:

Here is my one-page AI strategy for my small business. Your job is to
find the weak spots. For each of the four lines, tell me: (1) is the job
specific enough that a person could do it from this sentence alone, or
is it vague? (2) is the owner one named person, not a team or a
department? (3) is the number countable by hand today, and is it an
outcome (a customer helped, a payment received) rather than usage
(chats opened, logins)? (4) is the review date a real date about 30 days
out, and does the page say which of KEEP / EXPAND / STOP gets decided on
it? (5) is there any line here that is really two jobs hiding as one?
Do not suggest new tools. Do not add lines to the page. Answer in plain
language, one short paragraph per check.

Expected result: five short paragraphs, one per check. If the chatbot says the job is vague, rewrite the sentence until a stranger could do the job from it. If it says the number is usage, not outcome, change the number. If it finds two jobs hiding in one line, split the page — one job per page, always.

This is the ten-example test, done in prose: the chatbot is your eval set, and the page only ships when it passes.

The review date is the strategy

Thirty days pass. The review is a fifteen-minute meeting with yourself, and it has exactly three possible endings. Work through them in order.

First, count the number. Open the notebook, the sheet, or the WhatsApp log, and count the same way you counted the baseline. Do not count anything else: not how many times you opened the tool, not how many chats it handled, only the one number on the page. If the baseline was "3 of 12 after-hours enquiries answered before morning" and the new count is "10 of 12," the number moved. If it is "4 of 12," it barely moved. Write the new count next to the baseline on the page.

Second, ask why — once. If the number moved, what moved it? One honest sentence. "The bot answers in two minutes, so morning enquiries get booked before the customer tries the next shop." If it did not move, one honest sentence too: "Customers still call in the morning because the bot's answers did not include the price list." The sentence matters more than the number, because it tells you whether to expand or fix.

Third, decide: keep, expand, or stop. Keep means the number moved and the reason is sound — carry on, same page, new review date thirty days out. Expand means the number moved a lot — give the same worker the next most repetitive job, on a second page, with its own number and its own date. Stop means the number did not move, or moved for the wrong reason — switch the tool off, write the reason on the page, and file it.

A stopped page is not a failure; it is the strategy working. The most expensive AI in a small business is the tool nobody ever decided about.

This is where the page differs from a weekly check-in. A weekly check keeps the worker running — it is maintenance. The thirty-day review is a decision point: it can kill the job, and that is its purpose. The same distinction sits behind the Anthropic guidance this post leans on: you do not keep a subagent because it exists, you keep it because the test says the folded version scores worse. Your review date is that test, written into the business instead of the code.

What the one-page strategy is not

Three things the one-page strategy is often confused with, and why the distinction matters:

• It is not a playbook. A playbook tells the AI worker how to do its job — the steps, the tone, the escalation rules, what to say when it does not know. The strategy page tells the owner what the worker is for, who owns the result, how success is measured, and when it gets judged. The playbook is written for the worker; the strategy is written for the owner. A business needs both, on separate pages, and the strategy comes first — there is no point writing instructions for a job you have not decided is worth doing.
• It is not a pilot. A pilot is a bounded experiment with a start and an end, usually run to answer "does this work?" The strategy page is a standing decision about a real job in a real business, owned by one person, with a review date that recurs. The pilot asks "does this work?"; the strategy asks "is this job worth owning, and who owns it?"
• It is not a shopping list. If the page names three tools and no job, it is a shopping list. The tools belong in the build notes, not on the strategy page — the page names the job, and the tool earns its place by moving the number.

And one thing the page must never become: a deck. The moment the strategy needs slides, it has stopped being a decision and become a presentation. One page, four lines, one owner, one date. If you need a second page, you have a second job — write a second strategy.

That is the whole idea, and it is the answer to the surveys this post opened with. The gap between using AI and running on AI is not a technology gap. It is a decision gap: nobody picked the job, named the owner, chose the number, and set the date. The businesses that close it will not be the ones with the most tools. They will be the ones with one page, filled in, and a review date already in the calendar. Picking the right job and the right number is the work our AI Strategy & Readiness Consulting page describes.