T1 · Running the discovery session: teaching notes
Notes for t1-praxis-running-the-discovery-session.html (12 slides) and the printable scoping worksheet. The deck is what you put on screen for the client. These notes are how you run it. The discovery session comes before anything else in Praxis: before the charter, before the scorecard, before anyone buys a subscription. It draws out the one task each person will carry through the programme. The worksheet and the flow come from Roxanne Northover's Claude Academy, where she runs it with client teams in groups. We made it provider-neutral so it works whichever ecosystem the client ends up in.
How to use these notes
- Read T0, how we teach Praxis first. It covers the moves you use in every session, this one included.
- 90 minutes, in person or on a call. Up to about 12 people. Past 6, split them into groups of 3 or 4 so everyone gets airtime on their own task.
- Send the worksheet the day before and ask each person to fill in section 01 only: the tasks that eat their week. People who arrive with a list save you 15 minutes.
- Every person works on their own task. Don't let a group settle on one shared use case. The point is that each person leaves with something they do themselves.
- Stay provider-neutral. If someone asks "so is it Claude or ChatGPT?", park it: "We choose that in T4, once we know what your tasks need."
- Collect the Use Case Cards at the end (a photo is fine). They're the input for the charter (T2), the scorecard (T3) and the ecosystem choice (T4).
- Your assignment: run this session with a real or practice client and submit three completed Use Case Cards plus a short note (half a page) on which card you'd start with and why. Your tutor marks it within 48 hours against the rubric in T9.
Start with your week.
Set the frame: we're here to find one task per person, and the tools come later. People usually arrive expecting a demo. This session is about their work.
"Today we're going to talk about your week. Which AI tool to buy comes later. By the end of the 90 minutes each of you has one task, written on one card, that we'll work on together for the rest of the programme."
Ask everyone to have the worksheet in front of them, printed or on screen. If they did section 01 in advance, thank them for it out loud. It sets the tone that their input matters.
Seven steps, one worksheet
Orient, don't teach. The timings add up to 90 minutes including a 10-minute welcome. Steps 01 and 02 take the longest because that's where people do the real thinking.
"Here's how the time goes. The first two steps take the longest, because choosing the right task is most of the work. After that it's quick: we test it, split it, check the risks, say what success looks like and put it on a card."
Twenty seconds on this slide. If you're running late later on, cut time from 04 and 06 and protect 01, 02 and 07.
The task decides the tool
The whole Praxis method in one line: task first, tool last. Clients who buy a subscription before they know what it's for end up with licences nobody opens.
"A lot of companies start by buying licences and then wonder why nobody uses them. We do it the other way round. Once we know your tasks, we know which files AI has to read, which systems it touches and what could go wrong. That tells us what to set up. Your cards from today are the brief for everything we do next."
- "Has anyone here already got an AI subscription at work? How much do you use it?"
(wait: their answers usually make the point for you)
A senior person who has already chosen a tool ("we're a Microsoft shop, it'll be Copilot"). Don't argue. Say that's useful to know, and that today's cards will show whether it covers what the team needs.
What eats your week?
Get the raw list out before anyone judges it. Quantity first. People filter too early ("AI couldn't do that") and throw away their best candidates.
"Write down everything that eats your time, however boring. Especially if it's boring. How often, how long each time and what makes it slow. Don't ask yourself whether AI could do it. We'll get to that."
- "What's the task you'd most like to never do again?"
- "What do you do every month that feels like you're copying and pasting your own work?"
(give them 8 minutes to write on their own before anyone shares)
- Finance
- month-end packs, reconciliations, supplier queries, expense coding
- Operations
- status reports, rota changes, supplier chasing, incident logs
- Sales
- proposals, CRM notes after calls, follow-up emails, account reviews
- HR
- job ads, onboarding packs, policy questions, interview notes
- Leadership
- board packs, weekly updates, reading and replying to long threads
The example on screen is fictional. Use it to unstick people, then take it down. If someone just copies Ana's rows, ask them what their own month-end looks like.
Pick one task and make it concrete
Narrow down to one task, then describe it well enough that a stranger could picture it. The inputs row is the most important one for later: it tells us which systems and files AI has to reach, which drives the provider choice in T4.
A good first task has three things: a clear start (something triggers it), a clear output (you could hold it up) and a named reviewer (someone can judge it). If one is missing, the task is too vague to learn from yet.
"Pick one, the one you'd most like help with that you actually do yourself. Now describe it: what starts it, what goes in, what comes out, who checks it. Be specific about the inputs: which file, which system, which format."
- "Where does the data for that task actually live?"
- "Who would notice first if the output was wrong?"
(walk the room, read over shoulders, push on vague inputs)
People choosing a task for their team or their boss. Bring them back: "Which part of that do you do?" Also watch for "everything I do in Excel", which is a category. Ask for the one spreadsheet they opened this morning.
Is this a good first use case?
Six quick checks. Any "no" or "unsure" means narrow the task or pick another, and it doesn't mean the idea is bad. The high-stakes check matters most: first use cases help with the work, and the decision stays with a person.
"Six checks, yes or no. If you get a no, it doesn't kill the idea. It usually means the task is too big, so take a slice of it. Ana can't hand over deciding what to tell an unhappy landlord. She can hand over drafting the note that she then edits."
- "Who got a 'no'? Which check was it?"
(take one or two and narrow them live, in front of everyone)
The person who ticks all six in 30 seconds. Ask them to say out loud what "good" looks like for their task. If they can't, the last check is really a no.
AI helps. You own it.
Split the task in two columns. This is where people stop fearing replacement and start seeing help. It's also the first draft of the hand-over decision you'll make properly in T3.
"Take your task and split it. On the left, the parts AI could help with: reading, comparing, drafting. On the right, the parts that stay yours: checking, deciding, signing off. You're still the one accountable. AI does the first draft of the boring parts so you spend your time on the judgement."
- "What's on your right-hand column that you'd never want to hand over?"
(wait: this is where the room relaxes)
- Finance
- AI drafts the variance commentary, you sign off the numbers
- Sales
- AI drafts the proposal from the call notes, you set the price
- HR
- AI drafts the job ad from the role profile, you decide who to interview
- Leadership
- AI summarises the thread, you decide and reply
Anyone who puts everything in the left column. Ask: "If it goes wrong, who explains it to the client?" That's the right-hand column.
Quick risk check
The risk flags are the raw material for the company AI policy in T2. Every "yes" and its safeguard becomes a line in the charter. Say that out loud so people take it seriously.
Personal data on a consumer plan is the most common real risk you'll find. The safeguard is nearly always the same: a business or team plan where the provider doesn't train on your data, and only the data the task needs. That's why the subscription choice in T4 comes after this session.
"Six questions. A yes is fine, most real tasks have a couple. What matters is the safeguard next to it. And keep these: when we write your company's AI policy together, it's built from exactly these answers."
- "Which of your yeses would worry your manager most?"
- "Is there anything AI should never do on its own in your task?"
(wait)
People answering "no" to everything to get it over with. Read the first flag out and ask who has a customer name or a salary anywhere in their inputs.
What would success look like?
Success in terms you can measure and compare. This row becomes the scorecard in T3: the target is the KPI, the quality bar is the counter-metric that stops speed wrecking accuracy.
"Write today's number and the number you'd like. Then the quality bar: what has to be true for you to trust it. Speed without the quality bar is how people end up sending a wrong figure faster."
- "How would you know, a month from now, that this worked?"
(wait)
Targets with no baseline. If they don't know how long it takes today, the homework is to time it next time they do it.
Your Use Case Card
The card is the output of the session. The sentence at the top forces the task into a shape AI can work with: a trigger, an action, the inputs and an output.
"Fill in the sentence first: when this happens, I need to do this, using these things, to produce this. Then the rest of the card. The last box matters most: what's the smallest version you could try this week? For Ana, it's three landlords from last month, so she can compare the drafts with what she actually sent."
- "Read me your sentence."
(go round the room. Each person reads theirs. Fix vague ones live.)
A first experiment that's the whole task. It should be small enough to try in an hour, on real data from the past, where they already know the right answer.
Your card is the brief for what comes next
Close the loop. Show people that the card has a life after today. This is also where the client sees the shape of the full programme.
"Here's what happens to your card. The risks you flagged become your company's AI policy. The success measures become your scorecard. And the inputs, the files and systems, tell us which provider and which plan you need. Before next time, run your first experiment and bring the result, whatever it is."
Collect the cards (a photo each is enough). Log attendance and who has homework. Roxanne keeps a simple sheet per group: name, card done, experiment done. It's the easiest way to see who needs a nudge before the next session.
One task. One card. Your first experiment.
End on the homework. Keep it short.
"You came in with a busy week and you're leaving with one task on one card and a small experiment to try. That's all we need to start. See you next time with the result."