The slides, everything from the talk, and what wouldn't fit in ten minutes. The interview that starts it, the test to run on your own organisation, the onboarding plan, real examples with names, the economics, and the best things to read and try. No sign up. Take what's useful.
This is the question the room typed into their phones. Run it on ChatGPT, Claude, Gemini or Perplexity and read the answer as if you were the marketing director. Then swap in a question a fan or a sponsor would ask about your own organisation.
When we ran it in September 2026 across ChatGPT, Claude and Perplexity, all three returned a detailed plan in under a minute: Meta around 40%, TikTok 20%, YouTube 15 to 25%, creators 10 to 15%, the rest into search and testing. Sponsorship did not appear. Women's sport did not appear. Every plan asked for more context before it would sharpen the numbers.
Three things to do with the answer.
1. Ask the questions your fans and sponsors ask. "Which women's football club should I support in the north west?" "Is sponsoring netball good value for a regional brand?" "What's on this weekend for a family in Bristol?" Read what comes back about you, and what doesn't.
2. Read your own website as an agent would. Fixtures, results, attendances, partners, contacts, press. Is it there, is it current, is it plain text or buried in images and PDFs? Agents read text.
3. Fix the public record first. Before any tool, before any project. A clear, current, plain language account of who you are, who watches you and what you offer partners is the cheapest AI investment you will make this year.
One more thing every plan agreed on: 10 to 15% for creators. Your athletes are creators. That is the door women's sport already has into these plans.
You have onboarded a person before. Do the same four things, in the same order, and you will have a working assistant by the end of an afternoon. One person, one subscription, no developer. Each step ends with something in your hands.
Don't prompt it. Tap below and your assistant will ask you ten questions about your job, then give you the five jobs it could take off your plate, what to put in the first project, and an email to your team.
Using something else? Copy the interview below and paste it in.
Five common jobs, each with project instructions you can paste in and a "what good looks like" prompt to run first. Change the words to yours.
What good looks like: Here are our last three partner reports. Tell me what makes the best one better than the others, then write yourself a rule to follow.
What good looks like: Here is last season's spreadsheet. Show me what you can and cannot tell from it, and what one extra field would fix.
What good looks like: Here are our five best match recaps. Describe our voice in ten rules, then use them.
What good looks like: Here are the twenty questions we get most. Draft the answers from the venue notes and mark any you are not sure about.
What good looks like: Here is last year's report and this year's data. Draft this year's, then list every figure you were unsure of.
The real constraints in women's sport are structural, not cultural: small staffs, shared venues, fragmented systems, fan data in spreadsheets. Each one below is a colleague you onboard the same way: the files, what good looks like, access to the system it needs, someone to check the first job. None needs a developer. Most cost less per month than one matchday's volunteer catering.
Pulls ticket lists, sign ups, newsletter, social and matchday scans into one record. Tells you who came, who lapsed and who is new, every week.
First step: Export every spreadsheet you have into one folder. That is the CRM, for now.
Drafts the quarterly report from attendance, reach and content data, writes the activation recap, builds the renewal case in your template.
First step: Give it last quarter's report and the data it came from. Ask it to do this quarter's.
Shortlists brands whose audience matches yours, drafts the first approach, keeps the pipeline current.
First step: Hand it your best three partners and ask what they have in common.
Clips, captions and posts every match across channels, writes the recap, packages player moments for athletes to share.
First step: One fixture. Compare its output to what you would have posted.
Previews, reports and releases from the fixture data. Keeps the media list. Drafts replies to journalists in your voice.
First step: Give it your last five match reports so it learns how you write.
Answers fan questions about tickets, access and parking on the channels you already use. Keeps the checklist per ground. Runs the volunteer rota.
First step: List the twenty questions you get every matchday. It handles those first.
Bookings, attendance and safeguarding records kept current. The funder report drafted from the data, not from memory.
First step: One programme, one funder report. Time how long it took before and after.
Scouting reports, opposition summaries and workload flags in plain language, on the day you need them.
First step: Twelve's Earpiece is a subscription and a WhatsApp number. Start there.
After every big fixture, finds the first timers and follows up before the next home game, with something worth coming back for.
First step: Needs the fan record first. That is why it is move three, not move one.
Turns training and match footage into ready to post pieces per player. Handles the brand deal admin so players do not have to.
First step: Ask your players which two of them want it. Start with those two.
Drafts grant applications from your existing documents, builds the board pack from the numbers, keeps the compliance dates.
First step: Give it last year's grant application and this year's numbers.
The order matters. Half of these colleagues only work once the fan record exists, which is exactly what 41% of clubs do not have. Build the record first, then the systems that use it. Saying so is more honest than pretending any of this works in isolation.
None of these are predictions. Each is live somewhere in sport today, and each is small enough to copy.
A 16 team women's American football league had its whole 2026 season schedule built by AI: venues, travel costs, competitive balance, broadcast windows. Months of work. Its CEO: "the infrastructure to operate like the most elite leagues." November 2025.
Automated, real time highlights for every fixture, distributed to social and digital channels. For the NWSL that is 150 games a season turned into thousands of hours of content. Coverage stops depending on which games were broadcast.
February 2026. An AI company took Chelsea's front of shirt, with agents going into the club's operations: finance and procurement first, then logistics, stadium operations and fan engagement. Note the pattern: AI companies have become a new category of sponsor for sport. Treasure AI took the Portland Thorns' sleeve in July; Ideagen became Glasgow 2026's AI principal partner and presenting partner of the netball in May.
For decades the gap in elite women's sport was a data gap: decisions made on models built for men. The FA now has real time readiness and conditioning for the Lionesses, injury prevention, and a talent pipeline designed for the women's game. March 2026.
A WhatsApp based AI: send a player's name, get a full scouting report in thirty seconds. What used to take a data scientist days. For clubs without scouting networks or in house analysts, this is the whole gap closed for the price of a subscription.
Autonomous agents will run fan engagement across FIFA's digital platforms for Brazil 2027, and coordinate host cities, suppliers and partners for the tournament itself. June 2026. If agents are trusted with a World Cup, they are trusted with a matchday.
July 2026: both clubs named an AI company their "official agentic experience platform", promising personalised offers and easier ticketing. Within days the local paper's headline was that fans were fuming: worries about their data, and an AI deal dressed up as community values. Right technology, announced at people rather than built with them. Do the inside work first. Tell fans what you will and won't do with their data before anyone asks.
Every organisation grows the same way, by hiring, and hiring is the slowest and most fixed cost there is. A small team with assistants outgrows a big team without them, because the assistants are paid per task and flex with the season.
| Grow by hiring | Grow by onboarding assistants |
|---|---|
| Months to recruit | Onboarded in a week |
| Salaries, fixed, paid in the off season | Paid per task, in tokens. Pennies. |
| Every extra task needs another person | Nothing to pay when there is no fixture |
| Capacity set once a year, in the budget round | Capacity flexes with the fixture list |
A token is a unit of AI work, the way a kilowatt hour is a unit of electricity. A million tokens is about 750,000 words. You buy work by the unit, not by the head. Headcount is a fixed cost. Tokens are a variable cost: you pay for work done, and capacity flexes with the fixture list.
| Spend on | What it is | What it's for | What it costs |
|---|---|---|---|
| People | Your team | Judgement, relationships, taste. The thing you're protecting, not cutting. | Fixed |
| Frontier models | The smartest AI, rented from a lab (Anthropic, OpenAI, Google) | Judgement heavy work: drafting, analysis, anything with nuance | Roughly $2 to $5 per million tokens in, $10 to $25 out. A quarterly partner report costs under a pound to run. |
| Open weights models | Free to download, run anywhere (Llama, DeepSeek, gpt-oss) | High volume repeatable work: clips, tags, replies, classifying. Your data stays in house. | From about $0.15 per million tokens on a hosted service. A tenth of the price or less. |
The rule of thumb: frontier for judgement, open weights for volume, people for what only people can do.
Where to put the tokens: spend tokens where the hours go. Find the repeatable hours, the same report every quarter, the same replies every matchday, the same clips every fixture, and point agents at exactly those. People keep the judgement hours.
How to start: with a frontier model, because it is the simplest. Move the volume work to open weights once it is proven. Running open models needs someone to host them, and most small teams should not attempt that on day one.
Five things. Each one lets you delegate real work, not just ask questions. Start with the first.
Ethan Mollick's advice for non experts, and ours: pick one, pay the £20, and give an agent a real task from your real week. Then look carefully at the result and ask for changes, exactly as you would with a colleague. Serious AI work now means running agents, and both do it.
Anthropic's agent. You describe the outcome in plain English and it does the work across your files and systems: reports, analysis, documents, and yes, software. Most of what people use it for is "do the task for me". The desktop app is the place to start if you don't write code. This playbook and the talk it came from were built with it.
OpenAI's agent, now positioned squarely for people who aren't developers: plugins for marketing, finance, operations and more, driven entirely in natural language. Around a fifth of its five million weekly users are non technical, and they are the fastest growing group.
xAI's answer, launched August 2026: a team of always on agents, each with its own cloud computer, that sign into the tools you already use and finish multi step jobs while you're away, coming back only when something needs approval. The clearest picture yet of what "agent as colleague" looks like.
Five open resources, no sign up. Begin with the Attention Audit: what your working week is costing, split into execution and judgement, the ratio most people have never measured. Then the Your AI level, six stages from chatbot to synthetic employee, to see where your organisation sits and what it takes to move.
Not a reading list. The handful of things that changed how we think, in the order we'd give them to a friend.
The Vercel founder on running a $9 billion company where an internal agent replaced the ticket queue and bent the headcount curve, why he is still in every hire at 800 people, and the new skill when models write the code. The most practical account of an AI first operating model we've heard.
The talk that named Software 3.0: you now program computers in English. If you only ever watch one technical talk about AI, make it this one. Nothing in it needs a technical background.
The field guide. When to build an agent, when not to, and why simple, composable patterns beat heavy frameworks. Skim the diagrams if nothing else.
The companion piece: use cases, guardrails, and how to keep a human in the loop. Written for business readers, not engineers.
August 2026. The case against fully autonomous "factories" and for agents that pull people in at the right moment. The argument we make in one line, made properly over twenty minutes.
Our own two. The five layer map of the organisation behind the org chart slide, and what an agent actually is, why there will be billions of them, and how to build the foundation before everyone else does.
The 2017 paper that started all of this, and the one 1706 is named after. You don't need to follow the maths. Read the first page to understand that everything in this playbook traces back to a single idea about where to spend attention.
The best first book. How to work with AI as a colleague, built on four principles and a lot of honest experiment. If someone on your team reads one book, this is it.
Agents as a new layer of digital labour rather than a better chatbot. Written for senior executives piloting their first agents, with the failure modes included.
The strategy book. Why the operating model, not the model, is where the advantage sits, and how organisations redesign around it. Older than the others and still the clearest frame.
The advantage was never in the model. It's in what you build around it.
Every rival can rent the same model. What they can't rent is your context: your data, your history, your relationships, your people. Wire AI to that, and a team of five can do the work of fifty without losing what made it good.
If you want a second pair of eyes on your first workflow, or you'd like to talk through what an AI first operating model would look like for your organisation, email Craig. Bio and background at 1706.ai/craig.
craig@1706.ai · LinkedIn · Ground Truth · 1706.ai · The slides