Updated after the day with the learning from the room: what we covered, what to keep doing, and the resources to come back to any time.
Thank you for a brilliant day. This section captures the things that landed in the room, written down so they keep working on a Monday morning.
Ask the right question, and if you don't know it, ask AI to write it for you. Drown it in context, because the more you give it, the better the pattern it builds. And never take the first answer: ask it to review and improve its work, in a role if you like. "Act as an expert copywriter, review and improve this."
Keep going with the tone-of-voice work you started: two or three of your own sent emails as the examples, the document saved to SharePoint so Copilot can be pointed at it, and updated whenever it gets you wrong ("I never sign off that way, please update the document"). Teams sharing an inbox can build a shared voice for it alongside their own.
Treat AI like a bright new starter: check its early work until you know where you can rely on it. Ask it to flag its own assumptions and inferences so it has an off ramp instead of guessing. And be careful not to plant what you want to hear: "find the mistakes" will find mistakes that aren't there.
Talk, don't type, because you speak faster and self-edit less. Work in parallel: while one chat is thinking, open another and keep moving. Rename the threads you'll come back to. And keep one thread open as your trainer, where every "why?" and "how do I?" question goes. Pestering it is the skill.
When you've gone round in circles and it still isn't giving you what you want, paste this in.
Still to come from us: the video walkthrough of the Kanban to-do board exercise we ran out of time for, the responsible and sustainable AI resources asked about in the room, a simple method for sharing the Claude seats so everyone gets their own workspace, and the daily wrap-up method Mary mentioned. They will appear here as they land.
Two that came up in the room, both close to real Guild work, both a strong place to start.
Point AI at your member data to sort and shortlist: the two-star winners, say, or members turning over one to five million, so the sales team knows who to reach first. Keep that data inside your approved, signed-in tools and share only the minimum it needs. Let AI do the filtering and the first-pass list, and a person makes the calls.
Plenty of what lands in a shared inbox is formulaic: judges asking key dates, the same handful of questions. AI can sort those and draft the standard replies in the Guild's voice, ready for a person to check and send. The team's time then goes on the ones that genuinely need a human. A natural next step for the tone-of-voice work you started.
The whole team in the room, working hands-on, on your own work. Here is what we covered together.
The mindset shifts, how these tools actually work, hallucinations and how to catch them, data safety, and the Guild's own rules of engagement. AI is a skill, not a tool, and psychological safety is where the skill starts.
Everyone playing in the room, on their own work. The Time Thieves exercise finds where your week is leaking. True North turns the tools on something everyone is an expert in: themselves. Trust gets built by doing.
Open questions, what each person does on Monday morning, and how the team keeps the momentum going.
Everything from the session, to open and come back to any time. The two decks we worked through together, and the worksheet of prompts to run whenever you like.
How to think about AI, the mindset shifts, staying safe, and where this goes for the Guild.
The hands-on session, getting more from the tools on your own real work.
The worksheet of prompts to run any time, including the Time Thieves prompt that maps where your week is leaking.
The Guild's instinct is the right one. Cautious, staged, policy-led, with the team brought along rather than pushed. A small team holds a lot of trust here, across a publishing rhythm, an awards calendar that runs a year ahead, and a network of producers, retailers, judges and sponsors who notice tone shifts quickly. AI can quietly take work off the team in all of those places, but the order in which it's introduced matters more than the speed.
The point of this page is to make each next step shorter than the last. A clear water cooler policy you can put up tomorrow. A fuller AI policy bundle you can adapt at your own pace. A starting set of moves for marketing and editorial. A nod to the operational and awards work some of you are already automating. And, on Copilot specifically, the one habit that protects the work the team puts in.
One side of A4. Visual. Memorable. Designed to live where the team will actually see it, not in a folder no one opens. Print it, pin it up, and the bulk of the day-to-day questions answer themselves. The version below is keyed to Microsoft Copilot, since that's the tool the team is using.
One thing before you print it: the Guild's own policy is the authority. Where anything on this page and the Guild's AI policy and data security page differ, the policy wins. The lists below have been updated after the session to match it, with thanks to the sharp eyes in the room.
Our shared rules for getting the most out of Copilot, without putting the Guild, our subscribers, or our partners at risk.
The water cooler policy sets the culture. It doesn't replace the policy your auditors, partners and trustees will eventually want to see. The right shape is three short documents, in three different voices, each doing one job.
Two to four pages. For the team. Permitted and prohibited actions, how sensitive material gets handled, the accepted tools list, who can install what, the escalation path when someone's not sure. Written so anyone at the Guild can read it once and act on it without coming back to you.
One to two pages. For members, sponsors, judges and anyone who asks how the Guild manages AI. Transparency commitments. What we will and won't put through AI tools. The version you can hand to a partner who asks. Short, principled, signed by leadership.
The one above. One side of A4. In plain English. The version your most reluctant team member could read on a Monday morning and walk away knowing what's expected. Direct, human, no legalese.
Paste this into Copilot or Claude. It will ask you a small set of questions about the Guild, the team's appetite, and the data you handle, then write all three documents in order. Edit as needed.
Help me draft an AI policy for my organisation. I need three versions in three different voices. The three versions: 1. INTERNAL POLICY. Two to four pages. For my team. Covers permitted and prohibited actions, how sensitive material gets handled, the accepted tools list, who can install what, the escalation path when someone is not sure. Written so anyone on the team can read it once and act on it without coming back to me. 2. EXTERNAL POLICY. One to two pages. For members, partners, sponsors or anyone who asks how we manage AI. Transparency commitments. What we will and will not put through AI tools. The version I can hand to a partner who asks. Short, principled, signed by leadership. 3. WATER COOLER POLICY. One side of A4. In plain English. Sets the culture in public, not in a binder. The version my most reluctant team member could read on a Monday morning and walk away knowing what is expected. Direct, human, no legalese. To draft these well, ask me questions one at a time, maximum of twelve in total. Cover: - What we do, who we serve, how sensitive the data is - Our current AI tools and which ones we want on the accepted list - Our risk appetite and any regulatory context - Our cultural starting point: are people excited, hesitant, or somewhere in between - Any specific incidents or near-misses that have shaped our thinking When you have enough, draft all three documents. Show them to me in order. Use British English. No em dashes. No jargon. No "in today's fast-paced world".
If only one thing on this page gets done before next week, make it this. On the free Copilot the team is using, the protection that keeps prompts out of the wider training pool only kicks in when each person is signed in with their Guild work account. Personal account, no protection. Work account, commercial data protection applies. Same product, two very different stories.
So the discipline is small, but it is the whole game. Open Copilot. Look at the top corner. If it shows your Guild email, you're safe to draft, brainstorm and edit. If it shows a personal address or none at all, sign out, sign back in with the Guild account, and only then prompt. The two-second check in the water cooler policy above is exactly this habit.
When the Guild moves to a paid Microsoft 365 Copilot plan, an extra setting appears that controls whether prompts are used to improve the model. The simple advice when that day comes: turn it off. Until then, the sign-in habit is your protection. The policy tells people what to put in. The sign-in makes sure what they put in stays inside.
One practical note from the room: there are two Copilot apps, and the one you want is Microsoft 365 Copilot, signed in with your Guild credentials and showing the green shield. It has a voice mode too, which is perfect for talking a task through hands-free, on the move or in the car.
One of the most underused things AI can do for a media business is sit in for an audience before you publish. A synthetic focus group is exactly that. You describe the audience in detail, paste in the piece, and AI plays back how three or four different reader profiles would actually respond. Not as a replacement for real readers. As a way to catch the things you'd otherwise only learn after publication.
The same trick works on a subject line, a campaign idea, a sponsor proposal, an awards announcement. It is fastest at exactly the moment when you are most likely to skip the test, which is the moment that benefits most from one.
Paste this into Copilot or Claude. Replace the bracketed bits. The richer the audience descriptions, the more useful the response.
I want to run a synthetic focus group on a piece of work before I publish it. The audience: [describe in detail. Who they are, what they care about, what they read, what they buy, what they distrust. Three or four distinct profiles is plenty.] The piece: [paste the article, subject line, campaign concept, headline, proposal or whatever you want tested] For each profile, tell me: 1. The first reaction in their own voice (one or two sentences, sounding like them, not a summary). 2. What they would believe more strongly after reading this. 3. What would feel off, generic, or like it was not written for them. 4. The one change that would make this land harder for them specifically. Then, across all profiles, give me three things this piece is doing well, three things it is not doing well, and the single sharpest edit you would make. Use British English. No em dashes. No marketing speak.
A common AI failure mode is taking a piece of writing with personality and ironing it flat. This prompt does the opposite. Paste in two or three short examples of Guild writing you like, paste in the new piece, and Copilot will sharpen it without losing the voice.
Here's a piece of writing I want to sharpen, in the voice of the Guild of Fine Food. To learn the voice, read these two or three short examples first. Notice the rhythm, the level of warmth, the way the Guild talks about producers, food and trade, and the things the Guild would never write. Examples of Guild writing: """ [paste two or three short pieces, 200 to 400 words each, that sound exactly like the Guild] """ The piece I want sharpened: """ [paste the new piece here] """ Show me three things: 1. Two or three lines that already sound like the Guild, that I should keep. 2. Two or three lines that sound generic, formal, or like marketing copy that could have come from anyone, with a specific rewrite for each in the Guild's voice. 3. The opening sentence redone three different ways, so I can pick the one that lands. Do not rewrite the whole piece. Just sharpen what is there. Use British English. No em dashes.
There is already a thread of useful AI work running through the operational side of the Guild, particularly around the awards. Anything that involves reading the same document type at scale, lifting fields out of forms, sorting entries into categories, or generating standard pieces of correspondence is genuinely good ground for AI to take on. Done well, it gives the team back the parts of the cycle that used to disappear into manual handling.
A few principles to keep that work safe as it grows. Anything that touches entry data, judging scores or producer details runs on the work account, never personal. Judging, scoring and the final call stay human, always: AI can carry the paperwork around the awards, never the decisions. Outputs get a human eye before they go anywhere external, even where the AI is doing 90 per cent of the lift. And the rule of thumb is the same one the policy sets: if it's safe in the Guild's own systems, it's safe in Copilot. If it isn't, it isn't.
If it would help to have a fresh pair of eyes on what's already in motion, or to sketch what the next automation could look like, just say. The faster these get pressure-tested against the policy, the more confidently they can scale.
Take what's useful, leave the rest. The water cooler policy is the easiest first move. The sign-in habit on Copilot is the most valuable second. Everything else can be picked up at a pace that suits the Guild.
If anything on this page raises a question, or if a member of the team would like a hand using one of the prompts on a real piece of work, just reply to this and we'll find a way in.
The AI Potential team