Capture the context.
What percentage of the intelligence your organization created yesterday survived yesterday? Not the documents. The thinking. The reason a decision went the way it did.
AI cannot reason from context that disappeared.
Intelligence you are already losing
- Meetings nobody wrote down
- Stay interviews and exit interviews
- Manager coaching conversations
- The hallway version of the real problem
- Why a policy exception was granted
- Handwritten notes from a session like this one
Start with one surface
Do not try to capture everything. Pick the single recurring conversation where losing the thinking costs you the most, and capture only that one for thirty days. For most HR teams it is either manager coaching or investigation intake.
Consent first, every time. Tell people what is being recorded, why, where it lives, and who can see it. In Pennsylvania, recording an oral conversation requires the consent of all parties, so get it explicitly and confirm your own organization's rules before you record anyone.
Your mouth has higher bandwidth than your thumbs. You will never type five paragraphs of context. You will happily talk for five minutes about the thing driving you crazy.
If you do start capturing, the next question is what to do with what you collect. I wrote a book about that. Transcript Alchemy is 101 prompts for turning the meetings, calls, and voice notes you already have into something useful. It was built from my own transcripts, including the one where my own AI graded me a six out of ten.
Connect the intelligence.
Most people are still using AI at the chat level. Open a window, type something, get an answer, close it. Tomorrow you start over. These are five different problems.
| Layer | What it solves | The HR version of it |
|---|---|---|
| Chat | One task, right now | Draft this note. Summarize this document. Answer this question. |
| Project | Persistent context | A workspace holding your handbook, your policies, your templates, and your prior work, so the machine stops meeting you for the first time every morning. |
| Skill | A repeatable process | A standard operating procedure for your AI. How your team reviews manager documentation. How an investigation summary gets structured. Every time, the same way. |
| Connector | Access to real systems | Your calendar, your approved drive, your document library, a public website. The place where the truth already lives. |
| Agent | Work that gets done | It goes and gathers, compiles, drafts, and organizes while you run your day. You review and you approve. |
When a prompt works once, save it. When a process has to work every time, skill it.
A skill is not a longer prompt. It is the written process, plus the examples and templates that go with it, handed to the machine once instead of re-explained forever. This is how the expertise inside your best people stops being trapped inside your best people.
Connected does not mean permitted.
Turning on a connector is a technical act. Deciding what may flow through it is a governance act, and it is yours. See section eight before you connect anything to a system that holds employee information.
Direct the work.
Three lessons. If you remember nothing else from this morning, remember these.
- Build a team of experts Ask yourself who you would want in the room if money and availability were irrelevant. An HR business partner. An employment attorney. An organizational psychologist. A process engineer. A communications advisor. Then name them in the prompt. And when you do not know who you need, ask the machine first: "before we solve this, what expertise should be in the room to think it through well?"
- Make it interview you before it solves anything Most people drop half a thought into a model and then blame the model for a mediocre answer. You do not have to know what the perfect prompt is. You only have to say what great looks like, and then make the machine pull the rest out of you. This single sentence is the highest leverage thing in this guide: "Before you solve this, ask me one question at a time until you have enough context to give me an excellent answer."
- Talk. Do not type. Typing produces short prompts, missing context, and transactional behavior. Talking produces nuance, examples, exceptions, history, and the part you would never have bothered to write down. The richer the input, the more usable the output. Use voice.
The first answer is not the finish. It is the first draft of the thinking. Push back, fork it, and ask it what it is missing.
The prompts, ready to run.
The structure is always the same five parts. Experts, context, issue, goal, command. Replace the bracketed pieces and say it out loud rather than typing it.
Make it interview you
The one to use when you own a process and cannot articulate it yet. This is how tacit knowledge gets out of a person's head.
You are an HR operations architect, an employee experience designer, a process engineer, and a change management expert working together. I am [your role] at [organization, size, industry]. The process I want to improve is [process]. Do not propose a solution yet. Interview me one question at a time until you understand the current state, the outcome I want, the constraints, the systems involved, the stakeholders, the risks, the exceptions, and how we would know it worked. Ask your first question now.
Find the experts you did not know to ask for
Use this before any complex or sensitive piece of work, so you are not limited by your own vocabulary.
I am working on [issue] at [organization type and size]. Before we do anything else: what expertise should be in the room to think this through well, and what would each of those people want to know first? List the experts, then list their opening questions. Then wait for me.
Grade it, then fix it
Never accept the first draft. Make it score its own work against the outcome you named, then close the gap it found.
Score what you just produced out of ten against this standard: [what great looks like for this audience]. Be a hard grader. Tell me the score, then name the three specific things costing you points. Then rewrite it as a ten.
Turn one captured page into many usable assets
Photograph handwritten notes, a whiteboard, or a printed page. Then ask for the transformation rather than the transcription.
These are my handwritten notes from [session or meeting]. First, extract exactly what they say. Do not embellish and do not invent anything that is not on the page. If something is illegible, say so rather than guessing. Then give me twelve useful things this single page could become for [audience], ranked by how much time each one would save me.
Delegate a real task to an agent
Notice the guardrails. Read only, public sources only, nothing irreversible, and a named place for it to flag uncertainty. Never hand an agent an action you would not let a new hire take unsupervised.
Research task, public information only. Find [number] viable [venues, vendors, providers] for [purpose] in [location]. For each one, give me name, location, capacity, accessibility and parking notes, contact information, and the source URL. Constraints: do not contact anyone, do not submit any form, do not make a purchase, and do not take any irreversible action. If information is missing or you are unsure, flag it for human review rather than filling it in. Put the result in a table I can paste into a spreadsheet.
Map a workflow end to end.
This is the sequence we ran live on stage. It works on any process you own. Do it once with a colleague in the room and you will never build one the slow way again.
- Name the problem, not the toolWhat eats your time, what does it cost, and who feels it. If you cannot say what great looks like, you are not ready to prompt yet.
- Let it interview youUse prompt 01. Answer out loud. You are transferring what is in your head, not writing a specification.
- Map the current stateTriggers, steps, owners, systems, handoffs, exceptions. Include the ugly parts. The exceptions are where the real process lives.
- Name the deployable outputNot "a better process." A checklist, a message map, a template, a tracker, a script. One artifact a human can pick up Monday.
- Build it and set the human gateDecide exactly where a person reviews and signs off, and make that step impossible to skip. The higher the consequence, the earlier the gate.
- Pick the measureTime saved, cycle time, error rate, manager satisfaction, completion rate. One number. Check it in thirty days.
Two workflows to start with.
Pick one. Not both. The second one is easier once the first is real.
Employee onboarding
Onboarding is the right first build, and not because onboarding is exciting. It is because onboarding is where the knowledge is most obviously trapped in one person's head.
- Preboarding, day one, week one, thirty days
- Triggers and owners for every step
- What the manager does versus what HR does
- The communication map and who sends what
- A thirty day pulse check that actually gets run
Deployable output: a manager onboarding checklist and a communication map.
Investigation documentation
High consequence, high volume, and heavily structured, which is exactly the profile where a documented process pays off. Also the one where the human gate matters most.
- Intake structure and clarifying questions
- Which policy areas are potentially implicated
- Interview question sets by role
- A consistent summary and memo structure
- A named reviewer before anything is finalized
Deployable output: an intake template and a standard summary structure. The finding is always a human's.
Use AI to structure the work and surface what you might have missed. Never to reach the conclusion. That name on the decision is yours.
One source, many experiences.
The cost of transforming knowledge has collapsed. That is the part that changes HR, and most people have not noticed yet.
Take one trusted source. Your own public website, a policy document, a completed plan. Constrain the tool to that source only, so every answer is grounded in something you can point at. Then stop asking it questions and start asking it for formats.
| The need | The format | Why it lands |
|---|---|---|
| New hire briefing | A written report | Scannable, forwardable, easy to keep current. |
| The first week | An infographic | People remember the shape of a thing before they remember the words. |
| The commute | An audio overview | Learning time you were never going to get any other way. |
| Manager orientation | A slide deck | Gives a manager something to actually stand behind. |
| Did it land? | A quiz | Retrieval beats exposure. Testing is the learning, not the check on it. |
| Culture and connection | Music | An onboarding anthem, an enrollment deadline people finally remember, a recognition moment that feels personal. |
You are not creating six pieces of content. You are creating one source of truth and six experiences.
Two cautions. Ground it in a source you control and can verify, and check the citations rather than trusting the summary. And make it purposeful. A song that exists to prove AI can make a song is a party trick. A song that carries your benefits enrollment deadline is a communication strategy.
Clear the rights first. This one catches people.
- Commercial use is tied to your plan tier. On most generation tools the free tier does not grant commercial use, and internal employee communication often counts as commercial. Read the terms for the tier you are actually on, not the one you read about.
- Ownership is not automatic. In the United States, material generated from a prompt without sufficient human authorship is generally not copyrightable. You may be permitted to use something while owning nothing, and be unable to stop anyone else from using it.
- Do not use a real voice or likeness without written consent. That includes an employee, an executive, and any recognizable artist.
- Keep a record of the tool, the plan, the date, and the prompt. If ownership is ever questioned, that record is the only thing you will have.
General guidance for HR practice, not legal advice. Run anything you intend to publish past whoever owns that call at your organization.
Governance and the hard conversation.
Your employees are not waiting for your AI policy. Shadow AI is not a future risk. It is the current state of most organizations, including the ones that believe they have blocked it.
As the consequence of being wrong goes up, so should the amount of human judgment in the loop.
Never put this into an unmanaged tool
- Anything from an open investigation, including the allegation itself
- Medical, disability, or accommodation information of any kind
- Discipline and performance documentation naming a real person
- Compensation data tied to identifiable employees
- Anything you would not want read aloud in a deposition
Unmanaged means a personal account, a free tier, or any product your organization has not approved for that category of information.
Before anything leaves your hands
- Every factual claim traced to a source document you opened yourself
- Every citation checked, not skimmed. A confident citation can still be wrong
- Names, dates, dollar figures, and policy numbers confirmed against the record
- Nothing asserted that you could not defend if asked where it came from
- The human reviewer named, and the review itself documented
Treat every output as an unverified draft. That is not pessimism. It is the standard your function is already held to.
Do not repeat what you were told
Be careful with absolutes like "our enterprise plan means everything stays inside our walls." What may be processed is determined by your organization's specific plan, its configuration, the vendor's terms, and your own policies. Read those four, or ask the person who has. The confident summary is how organizations get surprised.
Have you had yours?
You spend your days telling managers to have the difficult conversation. The one in front of you is with your CEO, your general counsel, or your IT lead, and it starts with a single question: what are we actually doing about this? Put it on a calendar before you leave the building today.
Settings worth checking today.
Menus move and vendors change their terms. Treat this as where to look, not as what you will find, and confirm the current behavior in your own account before you tell anyone it is handled.
Check three things
- Whether your conversations are used to improve the model, usually under a data or privacy control in settings
- Whether chat history and memory are on, and what the tool has already remembered about you
- Which connectors you have authorized, and to which account
Ask three questions
- Which AI tools are approved here, and who decided
- What our agreement actually says about how our inputs are handled and retained
- Who is accountable when an AI assisted output turns out to be wrong
If nobody can answer the third one, that is your Monday.
Your seven day commitment.
Conferences create inspiration and then Monday happens. Fill this in before you leave the room. A workflow with a name and a date is the only version of this that survives the drive home.
It is not hard. It is foreign. Those are different problems, and only one of them goes away by starting.