AI in
Practice
A working guide to use across all three sessions, from getting a process ready, to using generative AI well, to designing an agent that can carry the work forward.
Emerging Leaders Network · AI Learning Series · October 2026
The example is recruitment, but the point is broader: learn how to define the work, give AI the right context, decide where it can act, and connect the result to business value.
Three sessions. One hiring journey.
Are You Ready?
Start before the tool. Clarify the need, map the process, find the data and name the owner. We also set up the difference between AI that responds and AI that acts.
Solve a Pressing Operational Problem
Go one level deeper. Every step of the process runs on knowledge. Build a knowledge library for each step, give each folder a README.md as its memory, and see how much better AI works when it can read it.
Agents That Move Money, Not Just Minutes
Hand the end-to-end chain to an agent that reads the library you built in Session 2, keep people in the decisions that matter, and work out where the commercial value sits.
Before AI: can the process run?
Take one real hire. If any of these four are unclear, the technology is not the first problem to solve.
Intent
Do we know what we are actually trying to achieve: role, property, timing, banding and constraints?
Process
Can we describe what really happens from “we need to hire” to interview, including the variations?
Data
Is the information needed to run the process available, usable and in a place we can access?
Owner
Is there a named person accountable for the process and the decisions inside it?
When AI responds. When AI acts.
Generative AI
Useful when you want an output and a person will decide and execute what happens next.
Agentic AI
Useful when a defined workflow can continue across several steps inside agreed permissions.
From process to knowledge
Every step of the hire runs on knowledge that already exists somewhere: a spreadsheet in finance, an old file on a shared drive, or a Head Chef's head. Gather it into one folder per step. That library is what the agent reads in Session 3.
Approve
- Headcount approval form
- Budget and cost centre
- Pay bands and salary scales
- Org chart and reporting line
Update
- Current job description
- Job description template
- Brand tone of voice
- Job spec: band, approvals, criteria
Advertise
- Salary survey for the market
- Approved job boards and agencies
- Advert template
- Pay transparency rules by market
Screen
- Shortlist criteria
- Scoring sheet
- Interview question bank
- Right to work checklist
Trial and offer
- Trial shift checklist
- Offer letter template
- Contract and sign-off route
- Onboarding plan
What a good knowledge library looks like
One current version
Old versions go to an Archive folder. The AI reads only what is current.
Names that explain themselves
Role, document, version and date. "JD final FINAL" does not count.
A README in every folder
What is here, who owns it and when it is next reviewed.
Rules written down
If it lives in someone's head, the AI cannot use it.
Decide what stays out
CVs and personal data never sit in the library.
AI forgets. Your folder remembers.
Andrej Karpathy, a founding member of OpenAI and former Director of AI at Tesla, describes today's AI as having "anterograde amnesia": it keeps its skills but cannot form new memories. Whatever you taught it on Monday is gone by Tuesday, unless it is written down somewhere the AI can read.
His answer: an LLM Wiki
A small library of plain text files that the AI reads and keeps up to date, so knowledge is built once and kept current.
- Raw sources · documents as they are, never changed. For us: old job descriptions, pay bands, the rota.
- The wiki · short pages with summaries, an index and a log of changes. For us: our folders and READMEs.
- The rules · one file that explains how the library works. For us: the house rules.
Three habits
The same habits work for a hotel folder.
- Add · a new document arrives and the pages it affects are updated.
- Ask · read the index first, then only the pages you need.
- Check · now and then, look for gaps, clashes and anything out of date.
Every time AI gets something wrong, add one dated line to the folder's README. Next time, it reads it.
Sources: Andrej Karpathy, Y Combinator AI Startup School (June 2025); Karpathy, "LLM Wiki" (April 2026).
How to make an .md file
An .md (markdown) file is plain text with a few symbols. People can read it, AI reads it reliably, and there is no formatting to break. Four symbols are all you need.
#a heading##a smaller heading-a list item>a note or quoteOn Windows
- Open Notepad.
- Paste the template below and fill it in.
- File, Save as. Set Save as type to All files.
- Name it README.md, encoding UTF-8, and save it in the folder.
On a Mac
- Open TextEdit, then Format, Make Plain Text.
- Paste the template below and fill it in.
- File, Save. Name it README.md.
- If asked, choose Use .md, not .txt.
Let Copilot draft it
- Paste your rough notes about the folder into Copilot.
- Ask: "Turn this into a README.md in markdown using these headings: purpose, what is in this folder, rules, owner and review, memory."
- Check it, then save it as above.
Some systems will not open or attach an .md file. Save the same text as README.txt. It works the same way.
# [Folder number and name] · [what it is for] > Purpose: [one line on what this folder is for] ## What is in this folder - [file name] · CURRENT - [file name] - [file name] · INTERNAL, never shared outside ## Rules - [a rule that lives in someone's head today] - [another rule] ## Owner and review - Owner: [name or role] - Reviewed: [date] · next review: [date] ## Memory: what we learned - [date] · [what the AI got wrong] · [the rule we added]
# 02 Update · Sous Chef job description > For refreshing the job description on any kitchen backfill. ## What is in this folder - job-description-sous-chef-v3-2026-10.docx · CURRENT - jd-template-brand-standard.docx - tone-of-voice.md - job-spec-sous-chef.docx · INTERNAL, never in an advert ## Rules - Reports to the Head Chef. There is no Executive Chef. - Always covers the brasserie and banqueting. - Level 2 food hygiene certificate is required. - Never use "young", "dynamic" or "native speaker". ## Owner and review - Owner: Head Chef, with HR - Reviewed 23 October 2026 · next review April 2027 ## Memory: what we learned - 2026-10-23 · AI kept the old Executive Chef line. Rule added. - 2026-10-23 · AI invented a gym perk. Perks now listed in job spec.
The job description sells the role. The job spec governs it.
A polished answer can still be wrong. The quality of the rewrite depends on what the leader gives the AI, and what the leader knows to challenge.
Job description · goes in the advert
External. Written for a candidate. It should explain the role clearly and represent the property and brand well.
- Purpose of the role
- Key responsibilities
- Experience and capability required
- Property and market context
- Brand tone of voice
Job spec · stays in the library
Internal. It carries information needed to hire and manage the role that does not belong in the external advert.
- Banding / grading
- Compensation parameters
- Reporting line
- Approval route
- Internal constraints and decision criteria
Which Copilot do you have?
Copilot Chat
With your work account. Chat, summarise and draft. Attach a file and ask about it. Build a simple agent.
- 6 November: follow along and build a simple version
Microsoft 365 Copilot
Paid licence. Works inside Word, Excel, Outlook and Teams, and with your company's files.
- 6 November: build the agent on your own library
Copilot Studio
Usually set up by IT. Agents that follow rules and take actions across your systems.
- 6 November: see where this goes next
Hand the workflow to an agent
An agent is not “AI with a better prompt”. It needs a goal, a process, authorised tools, boundaries and a way to know when to stop or ask for help.
Goal
Define the result, timing, constraints and approvals.
Plan
Turn the process into steps, owners, data and decisions.
Execute
Use authorised systems and data to carry out the work.
Evaluate
Check quality, gaps, delays and progress against the goal.
Adjust
Change the next action when the result is weak or conditions change.
Complete
Return the outcome, evidence, exceptions and items needing a person.
Fill this approved role in 21 days without rejecting a candidate or making an offer without human review.
Move money, not just minutes
What changes financially or operationally if this process becomes faster, more consistent or less dependent on external support?
Keep the calculation defensible. Avoid converting every minute into a fictional headcount saving.
Keep people in the decisions that matter
Set before you automate
- A clear business outcome
- The real process, including exceptions
- What data the agent may use
- What systems it may access
- Where human approval is mandatory
- A named person accountable for the result
Do not delegate blindly
- Do not give unrestricted access to candidate or employee data
- Do not assume a shortlist or recommendation is correct
- Do not let a machine reject a candidate without the agreed review path
- Do not automate an unclear process and expect a clear outcome
- Do not deploy without an escalation route
- Do not measure success only in minutes saved
Build one folder before 6 November
Take one step of a process you own. Design its folder, then write its README.md using the template in this guide. Save your answers in your Field Guide notes and bring it to Session 3.
Use the same thinking on your own process
Choose one repeated process from your own area. You will use this in Session 3, but you can start filling it in at any point in the series.
Terms we will use
www.deep-hospitality.com
Caroline Hardman · Caroline.hardman@deep-hospitality.com
Nitin Thariyan · Nitin.thariyan@deep-hospitality.com