Karney AI University: Modules 2, 3 and 4
Everything from the session, in one place.
Summary
Christi Brown led AI training on tool choice, costs, prompting, and building Copilot and Claude agents. The group identified information fragmentation as the main opportunity and set follow-ups.
Training Hub and session setup
- The session opened with a recap of the prior AI introduction and framed today as practical, task-focused training.
- Christi shared the Training Hub with Q&A, downloads, recaps, one-pagers, and Karney-specific resources.
- Questions submitted in the hub are public, visible to Christi, and meant to guide examples.
- The Karney Properties version will stay online for over 60 days unless extended.
Choosing AI tools and managing costs
- Claude is best for long documents, careful reasoning, repeatable skills, and client-facing writing.
- ChatGPT is strong for web research, images, brainstorming, and detailed strategic outputs.
- Copilot is best when answers live in Outlook, Teams, SharePoint, OneDrive, or sensitive Microsoft data.
- Tokens drive cost, and higher models plus deeper thinking cost more.
Homework examples and early use cases
- One attendee tried using AI to improve a to-do list but found the result too complicated.
- Another compared bids, generated contractor follow-up emails, and analyzed a Cambridge injury claim.
- ChatGPT was described as more detailed and strategic, while Claude was briefer and more conversational.
- AI reduced work that would have taken hours into a few minutes of drafting and analysis.
Discovery themes
- Participants wanted help organizing email, tracking ownership, and surfacing only what matters.
- Key organizational questions included who owns what and how trusts, partnerships, and owners connect.
- Facilities data was another need: paint, roofs, electrical capacity, fire life safety, and capital planning.
- Lease and partnership agreement variance review was raised as a major quality control opportunity.
Prompting, skills, and agents
- Christi reinforced prompting with role, context, task, format, and examples.
- Copilot can prioritize Outlook email using AI summaries plus high and low priority rules.
- Claude skills can create repeatable routines, such as a daily inbox briefing.
- Copilot agents were demonstrated for Daily Briefing and Chief of Staff workflows.
Next steps
- Karney: send the training documents around after the session.
- Karney: forward the Training Hub links to attendees.
- Karney: share the processed meeting recording link with Christi.
Decisions made
- Karney will not use the open source Chinese AI tools discussed.
Downloads
Prompts
Copy, paste into the tool named on the card, and adjust the details to your work.
Daily Briefing (inbox sweep)
The one Christi runs every morning, made generic. Save it as a Claude skill or a Copilot agent (both how-tos are on this page) and run it first thing. Replace the bracketed parts with your own names and rules.
You are my Daily Briefing assistant. Your job each morning is to turn my inbox, calendar, and task list into one short, trustworthy summary so I can see in two minutes what needs me, what is already handled, and what is waiting on someone else. Two things matter more than anything else: nothing important slips past, and nothing goes out in my name without me seeing it first. Sources: read my [Outlook inbox / other mailboxes], my calendar, and [task list, e.g. Microsoft To Do or Planner]. If a source cannot be read, keep going with the others and say so at the top of the summary. A silently skipped inbox is worse than a late one. Window: the past 24 hours. On Monday, go back to Friday morning so the weekend is not lost. Also read my Sent Items and Drafts from the past 7 days so you know what I have already answered or promised. Classify every message in this order, because the order protects the important things from the noise filters: 1. Surface immediately: anything from or about [named executives, key clients, or VIPs]. These go at the very top, above urgent, with no filtering. 2. Skip entirely: automated system notifications, alerts, receipts, and newsletters. The test is whether a person wrote it. A human escalation always surfaces. 3. Move to Deleted Items: obvious spam and social network notifications. Never permanently delete anything. 4. Bills and invoices: check Sent and the thread to see whether it was already forwarded to [accounting contact]. If not, draft the forward. Report either way. 5. Calendar items: out of office notices, new hire start dates, events that need preparation. Check for an existing entry before creating one, because this runs daily and duplicates pile up. 6. Vendors: tag each as [Security / IT / Facilities / other categories that fit us] and say which ones need me. 7. Support cases: escalate when a vendor is asking a question that needs my input, or a case has gone 24 business hours with no vendor response. 8. Everything else is a candidate for the summary sections below. Give extra attention to [priority clients or projects]. Drafting rules: never send anything. Every reply or forward goes to my Drafts folder for me to review. Before drafting, check Sent for the past 7 days to confirm I have not already replied, and check Drafts for an existing draft on the same thread. If one exists, do not create another. Never route to, forward to, or loop in [people who must never be included] on any topic. Write drafts in my voice: conversational, direct, short. Build the summary in this structure, leaving out any empty section except the sources line: Daily Briefing [date] Sources: [each source, read or failed with reason] VIP: sender, subject, what they need Urgent: sender, subject, what is needed, why it is urgent Needs a response: sender, subject, what is needed, and "draft ready" if you made one Decisions pending: what is being decided, who is waiting, by when Follow ups I owe: things I said I would do in Sent mail that have not happened Quick wins under five minutes: one line each Vendors: vendor, tag, what is happening Support case escalations: vendor, case, reason Bills: bill, entity, already forwarded or forward drafted Calendar added: person or event, date Tasks: task, due date, days overdue Housekeeping: counts only of what was moved, labeled, or drafted Keep every line tight: who, what, what I need to do. I read this on my phone, so the first screen should tell me whether my morning is on fire. Never: send anything; permanently delete mail; include [excluded people] anywhere; let system alerts into the summary; create a duplicate draft, event, or record; put a VIP item anywhere but the top; skip a source without saying so.
Organizational Coordination Assistant
Built live in the session as a Copilot agent. Who is doing what, what is overdue, what changed, what is at risk. Start in Microsoft 365 Copilot; add Planner or To Do only for explicit tasks.
You are my Organizational Coordination Assistant and Executive Chief of Staff. Goal: Help me understand who is doing what, what is overdue, what changed, what is at risk, and what requires my attention. Before answering, search the Microsoft 365 information I am permitted to access, including Outlook, Teams, meeting transcripts or notes, SharePoint, OneDrive, and any available task sources. Do not wait for me to paste the information if you can retrieve it from those sources. Scope: [today / this week / named project or team]. Find evidence of: - assigned work and accountable owners - commitments and follow-ups - deadlines and overdue items - blockers, risks, escalations, and unanswered questions - decisions or approvals waiting on leadership - recent changes in priority, status, or ownership Return: 1. Immediate attention 2. This week 3. Monitor 4. Active workstreams 5. Meetings requiring follow-up 6. Ownership gaps or conflicting assignments 7. Top three recommended leadership actions For every item include: topic, owner if explicitly identified, status, due date if explicitly stated, why it matters, recommended next action, and source. Rules: - Cite the source for every item. - Do not invent owners, deadlines, status, or project details. - Distinguish verified facts from your interpretation. - If evidence conflicts, show the conflict. - If information is incomplete, say what is missing. - Do not send messages, create tasks, change assignments, or approve anything unless I explicitly ask and review the action first. - Filter aggressively. I want to know as little as possible while still knowing everything I need to know.
Chief of Staff: know less, know enough
The one built for Jeff in the session. A daily or weekly executive brief that surfaces only what has a material effect. Works as a Copilot agent or a Claude skill.
Act as my AI Chief of Staff. I want to know as little as possible while still knowing everything I need to know. Search the Microsoft 365 information I am permitted to access for [today / this week]. Bring forward only items with a material effect on revenue, risk, operations, strategic objectives, tenant or customer experience, or leadership commitments. Answer: 1. What are the three most important things I need to do? 2. What decision or approval is waiting on me? 3. What work appears stalled or at risk? 4. What should I delegate, and to whom only if the accountable role is explicitly documented? 5. What can safely wait? 6. What might I miss if I do not look deeper? For each item provide the source, date, and confidence. Do not invent urgency, ownership, impact, or deadlines. Do not take action without my approval.
Ownership Intelligence Assistant
Who owns what, across trusts, partnerships, and entities. Inventory and validate the records first, put approved documents in SharePoint, then run this against them. Research support, not a legal conclusion.
You are an Ownership Intelligence Assistant supporting leadership, finance, legal, and asset management. Question: [Who owns PROPERTY / ENTITY? Show the ownership chain as of DATE if the records explicitly support that date.] Before answering, search only the approved ownership sources I am permitted to access, such as [SharePoint library], [validated ownership table], partnership agreements, trust documents, entity records, and approved finance or legal files. Create: 1. Direct owner(s) 2. Intermediate entities or trusts 3. Ultimate ownership chain, only where supported 4. Ownership percentages, only when explicitly stated 5. Effective dates, only when explicitly stated 6. Source documents and relevant sections 7. Conflicts, missing links, expired records, or uncertainty 8. Questions requiring legal or finance validation Rules: - Do not infer ownership from file names, email discussions, or incomplete references. - Do not combine records from different effective dates without labeling them. - If exact ownership cannot be verified, say so. - Treat the output as research support, not a legal conclusion. - A designated legal or finance reviewer must validate the result before business use.
Property Knowledge Assistant
Building facts across 80 to 90 properties. Paint, roofs, electrical capacity, fire life safety, capital items. Keep cleaning up in Excel if it helps, but point this at a governed SharePoint library or the property system.
You are a Property Knowledge Assistant. Property: [PROPERTY NAME / ID] Question or decision: [QUESTION] Before answering, search the approved property sources I am permitted to access, including [property system], [SharePoint library], [capital plan], inspection reports, invoices, project files, and the validated property inventory. Return: 1. Direct answer 2. Building facts relevant to the question 3. Last recorded maintenance, replacement, or inspection dates 4. Current projects and future capital items explicitly listed 5. Remaining useful life only if an approved source provides it 6. Data gaps, stale records, and conflicting entries 7. Sources, document dates, and responsible data owner if stated 8. Recommended follow-up questions for the property or facilities team Rules: - Never interpret missing data as proof that work was not performed. - Do not calculate remaining useful life unless the approved methodology and required dates are provided. - Label estimates and assumptions clearly. - Do not update the source record or approve capital work without human review.
Contract Variance Review Assistant
Compare a renewal, lease, or partnership agreement against the standard and flag every clause that changed. Claude is the strongest tool for this. Start as a human-reviewed pilot; a qualified reviewer makes every consequential decision.
You are a Contract Variance Review Assistant supporting a qualified human reviewer. Compare: - Standard or baseline: [APPROVED TEMPLATE / ORIGINAL AGREEMENT] - Document under review: [RENEWAL / LEASE / PARTNERSHIP AGREEMENT] - Review scope: [ALL CLAUSES / NAMED SECTIONS] Before answering, read both complete documents and use only their text and any approved review standard provided. Produce: 1. Executive summary of material differences 2. Clause-by-clause variance table 3. Added, removed, or changed language 4. Business or operational implications stated as review questions, not legal conclusions 5. Missing exhibits, definitions, dates, signatures, or cross-references 6. Items that match the standard 7. Items requiring counsel review 8. Exact source section or page reference for every finding Rules: - Quote only the minimum text needed to identify a variance. - Do not invent a standard that was not supplied. - Do not provide a final legal opinion, approve terms, or send communications. - Flag uncertainty and OCR or legibility issues. - A qualified reviewer makes all consequential decisions.
Market Intelligence Assistant
Submarket research, offering memorandums, and transactions for strategy sessions. ChatGPT Deep Research or the Copilot Researcher agent. Keep external research separate from internal portfolio context, and require source links and dates in every briefing.
You are a Market Intelligence Assistant supporting strategy discussions. Research question: [MARKET / SUBMARKET / PROPERTY TYPE / DECISION] Time scope: [DATE RANGE] Approved external sources: [LIST LICENSED OR PUBLIC SOURCES] Internal context: [APPROVED PORTFOLIO FILES OR SHAREPOINT LOCATION] First, gather current external evidence from the approved sources. Then review the internal context I am permitted to access. Keep external facts and internal facts clearly separated. Create: 1. Executive summary 2. Recent market activity explicitly supported by sources 3. Offering memorandums or transactions reviewed 4. Comparable evidence and important differences 5. Relevance to the current portfolio 6. Opportunities, risks, and open questions 7. Source list with publication or record dates 8. Recommended questions for the investment team Rules: - Cite every factual claim. - Do not present missing data as zero activity. - Do not use unapproved or inaccessible sources. - Label conflicting sources and stale information. - Do not make an investment recommendation or decision. Frame tradeoffs for human review.
How-tos
Step by step, for the things shown in the session.
Let Copilot prioritize your Outlook inboxCopilot
Shown live in the session. Works in classic Outlook; the new Outlook has the same feature under a slightly different layout.
- In Outlook, click the Copilot button in the ribbon, then Prioritize.
- Turn on Let Copilot prioritize my email on arrival. It will show AI summaries and mark items high or low priority. Out of the box it is already good at spotting the high priority ones.
- Add your own rules on the Higher priority tab. Example: anything from a specific person at Karney.
- Add rules on the Lower priority tab. This is where most people get the real value. Examples: newsletters, invitations to conferences or meetings from vendors you have no relationship with, automated notifications you are only copied on.
- Look for the up and down arrows in your inbox list. Read the up arrows first.
Tip: if Copilot is marking something wrong, add a rule for it instead of ignoring it. The rules are what teach it your inbox.
Turn a repeatable process into a Claude skillClaude
A skill is a saved recipe: a process you are tired of explaining, written once so Claude runs it the same way every time. The Daily Briefing was built this way in the session.
- In Claude, start a new chat and type / to see your skills. Choose Skill Creator. Anthropic provides it.
- Tell it what you want, for example: "Build a daily briefing skill. Here are my instructions." Then paste your instructions. The Daily Briefing prompt on this page is a good starting point.
- Claude drafts the skill in an artifact on the right side of the screen. Read it. Adjust anything that is not how you work.
- Click Save skill at the top of the artifact.
- From now on, type / in any new chat and pick the skill by name. You can rename it later.
Tip: a skill only sees what Claude is connected to. If it should read your inbox, connect Outlook under Settings, Connectors first. Karney data stays in Karney's Claude account.
Create an agent in Copilot with Agent BuilderCopilot
What Claude calls a skill, Copilot calls an agent. Both the Daily Briefing and the Chief of Staff were built this way in the session.
- Open the right Copilot: the Microsoft 365 Copilot app (the one with the M365 badge on the icon), or go to office.com and open Copilot from there. Make sure you are signed in with your Karney account.
- In the left sidebar, choose Create agent.
- Either describe what you want and let Agent Builder draft it, or click Skip to fill it in yourself.
- Give it a name (for example, Chief of Staff) and paste your instructions. Add a short description.
- Under Knowledge, tick what it should be able to read: Outlook, Teams, cloud files. Add specific files as attachments if it should always have them, such as a RACI chart or a priorities list.
- Agent Builder requires at least one suggested prompt. Ask it in the left pane: "Create suggested prompts for me."
- Click Create. Start a chat with it, for example: "Give me today's executive brief."
Sharing: you can share an agent with certain people or the whole organization. Permissions still apply. If the agent reads a file you can open and a colleague cannot, they still cannot see it through the agent.
Shortcut: Microsoft ships a Plan My Day agent. Open More agents, find it, then create a new agent that references it and edit the instructions to fit you.
AI opportunities for Karney
Five places where AI can do real work at Karney, and the tool that fits each one best.
| Opportunity | Best tool | Why it fits best |
|---|---|---|
| Organizational Coordination Assistant | Microsoft Copilot + Planner + Teams + Power Automate | Most of the information already lives in Outlook, Teams, meetings, tasks, emails, and documents. The value comes from coordinating work, surfacing actions, reminders, follow-ups, and status updates. |
| Ownership Intelligence Assistant | Enterprise Knowledge Agent (Copilot Studio + SharePoint/SQL ownership repository) | The challenge is finding and explaining ownership structures across partnerships, trusts, and entities. This is fundamentally a retrieval and reasoning problem over structured and unstructured data. |
| Property Knowledge Assistant | Knowledge Agent + Document Intelligence + SharePoint Repository | Property data exists across inspections, reports, spreadsheets, leases, plans, and contracts. A knowledge assistant that can answer questions across all sources creates the most value. |
| Contract Variance Review Assistant | Claude (analysis) + Contract Repository + Copilot Agent | Claude's strength is document comparison, legal language analysis, and exception identification. This use case is heavily document-centric and benefits from deeper reasoning capabilities. |
| Market Intelligence Assistant | ChatGPT Deep Research or Copilot Researcher Agent | This opportunity relies heavily on external market data, offering memorandums, transactions, and market research. Research-focused agents are best suited for continuously gathering and synthesizing external information. |
Where to start with each one
- Organizational Coordination. Start in Microsoft 365 Copilot. Add Planner or To Do only for explicit tasks. Add Copilot Studio once the brief and the ownership rules are validated. Add Power Automate only for defined notifications or approvals.
- Ownership Intelligence. Start by inventorying and validating the records. Put approved documents in SharePoint and the relationships in one authoritative table or system. Add a Copilot Studio retrieval experience after validation.
- Property Knowledge. Keep cleaning up in Excel if it is useful, but define a governed destination. Use SharePoint or the approved property system for authoritative records, then add a knowledge agent.
- Contract Variance. Start with a controlled, human-reviewed document comparison pilot. Move reusable standards and agreements into a governed library. Automate routing only after the review method is stable.
- Market Intelligence. Separate external research from internal portfolio context. Keep approved research outputs in SharePoint, and require source links and dates in every briefing.
The matching prompt for each is in the Prompts section above. Every one tells the assistant to find information before answering, cite evidence, separate facts from interpretation, and leave the decision with a person. Replace the bracketed text with approved scopes, standards, properties, or dates.
Usage, costs, and the golden rule
Tokens are the meter. Think of every word in and out as a cost. Higher models cost more per word, and deeper thinking returns more words, so the cost climbs with both. Within each family the handout lists models from most to least expensive: Opus at the top, Haiku at the bottom for Claude; Astra at the top, Luna at the bottom for ChatGPT.
Prices are moving. In the 48 hours before the session, ChatGPT cut prices on all its models by about half, and Claude was expected to follow. Treat any number you hear as a snapshot.
Your allowance. Most Karney users are set to about $25 of usage per month, which covers everyday work comfortably. Where Crimson has seen a single heavy user run past $500 in a month at other clients, ceilings are set, typically at $50. If someone runs out and needs more, that is a quick adjustment, not a problem.
Per person, not pooled. Claude Team plan limits are per member: each seat has its own allowance and one person cannot borrow another's unused usage. Pooled consumption is an Enterprise plan feature. What the owner controls is the extra usage budget, with a spend limit set per person, so the fix for a heavy user is to raise that person's limit or move them to a Premium seat.
Leave the thinking dial on default. Turn it up for comparing documents or contracts. Most people will never need Max or Ultra.
The golden rule. Karney data only goes into Karney's business accounts: Karney's Claude, Karney's ChatGPT, Karney's Copilot. Never a personal login. Some personal accounts cannot turn off model training on your data. Grok's personal accounts are one example, and the open source Chinese models are another.
Decision made in the room: Karney will not use the open source Chinese AI tools.
Discovery: what the group said
Christi asked five questions to shape which examples get built for Karney. The answers, lightly cleaned up.
If you could hire one more person tomorrow, but they could only do one thing all day, what would it be?
- Organize my email.
- Make sure everyone in the organization knows what is going on, keep it organized, figure out who should do what, follow up on whether it got done, and let everyone know.
- Keep the trains running on time: what the priorities are, what needs to be communicated to whom and when.
What is important but nobody has time to do consistently?
- The growth strategy. The day to day crowds out the highest level objective.
- Market intelligence: download, organize, analyze, and track offering memorandums and sales data in Karney's submarkets, and produce an underwriting and investment summary for strategy sessions. Today that is an analyst or two's full time job.
What question do you wish you could answer today, but the information is scattered across too many places?
- Who owns what. Trusts, partnerships, layers, and owners, with no single map of the spiderweb.
- What do we have in a portfolio of 80 to 90 buildings: when was it last painted, last roofed, electrical capacity, fire life safety, remaining useful life of each component. This feeds the five year capital plan and lets Karney fact check the plan the construction management vendor puts together. It is being built in Excel today, with AI used to find the information.
- For every lease and every partnership agreement, what is off standard. A variance and quality control question against 75 years of documents, and a way to onboard the new general counsel without reading all of it by hand.
What work do you secretly hope gets done while you are on vacation?
- All of it.
The pattern. Operational visibility and follow through, analysis work handed off, strategic work that keeps getting cut, and above all information fragmentation. That is what the five opportunities above are built around.
Roadmap, guardrails, and how to judge a pilot
Four phases
Phase 1: Validate quick wins. Pilot the Organizational Coordination Assistant with a small leadership group and a defined scope, one week or one project. Measure whether it finds real commitments, reduces manual follow-up, and avoids false urgency. Inventory the ownership, property, and contract sources without building a production agent yet.
Phase 2: Build governed knowledge foundations. Define approved sources, metadata, owners, retention, permission boundaries, and validation steps. Pick the authoritative home for ownership relationships and property facts. Define the standard language and the materiality rules for contract review.
Phase 3: Standardize repeatable experiences. Create Copilot Studio agents only after the users, sources, prompts, and review process are stable. Use Power Automate for clearly defined triggers, routing, reminders, and approvals, not for judgment calls. Publish approved outputs back to governed locations so future retrieval improves.
Phase 4: Expand decision support. Pilot market intelligence with specific submarkets, approved sources, and a leadership decision cadence. Connect internal portfolio facts only after data quality and permissions are validated.
Questions to answer before building
- Which problem consumes the most staff hours in a typical month?
- Which problem creates the greatest business risk when the answer is wrong or late?
- Who is the business owner for each process and data set?
- Which source is authoritative for tasks, ownership, property facts, contract standards, and market research?
- What information is structured, and what exists only in documents, email, or people's memory?
- What permission boundaries must the assistant preserve?
- What counts as a material contract variance?
- What output would leadership trust enough to use, and what evidence must be shown?
- What may an assistant recommend, and what may it do only after approval?
- What would success look like after a limited pilot?
Guardrails
- Use business-approved accounts and sources for company information.
- Preserve existing permissions. An agent must not give a user access to information the user cannot otherwise open.
- Cite evidence and show source dates for every consequential finding.
- Label missing, stale, conflicting, or low-confidence information.
- People stay accountable for legal, financial, investment, capital, and personnel decisions.
- Explicit approval before sending messages, creating or reassigning tasks, updating records, or starting workflows.
- Do not assume a tool can reach a source until the connector, license, permission, and tenant configuration are verified.
Pilot scorecard
| Measure | Question | Evidence to collect |
|---|---|---|
| Usefulness | Did the output help the user decide or act? | User rating and examples of avoided work |
| Accuracy | Were owners, deadlines, and statuses supported? | Sample verification against source records |
| Completeness | Did it miss material items? | Reviewer-identified omissions |
| Noise reduction | Did it filter effectively? | Items removed or downgraded by users |
| Trust | Could users trace claims to sources? | Citation and source-date coverage |
| Time saved | Did the pilot reduce searching or follow-up? | Before and after task observation, or user estimate |
| Safety | Did it avoid unauthorized actions and unsupported conclusions? | Exception and incident log |
What this is based on
The session transcript gives strong qualitative signals, but it does not prove time savings, error rates, data completeness, or what the Karney tenant can reach today. Those get measured during discovery and a limited pilot. Tool placements above are hypotheses to validate, not commitments.
Follow-ups and what's next
From Christi
- Answer every open question on this page's board. Questions that need a demonstration get a short video, posted here.
- Prompts from the session are in the Prompts section: Daily Briefing, Chief of Staff, Organizational Coordination Assistant, and one for each of the other opportunities.
- The recording is linked at the top of this page (Teams recap, Karney sign-in required).
- Confirmed: Karney's Claude allowance is per user, not pooled. See Usage and costs.
- Crimson is adding two to three hours to the engagement, nine in total, for deeper build sessions on the ideas raised today.
- An in-person Friday in October to work through the discovery questions and build efficiencies together.
- Help build the property knowledge database, starting from the Excel work already under way.
Where to find things
- The Teams meeting chat from the session, which holds the links and prompts shared during the training. In Teams, look under Chat for "AI University (Day 2)".
- This page, for everything else.
From Karney
- Forward the Training Hub link and the training documents to attendees.
- Share the processed meeting recording link with Christi. Done.
Open office hours
Fridays for the next couple of months, open to all Crimson clients active in AI. Bring an idea or an example and it gets built live. These sessions are recorded and shared across clients, so bring ideas, not sensitive data.
What was covered, and what is next
This session covered modules 2, 3, and part of 4: choosing tools, cost, prompt patterns, and building the Daily Briefing and Chief of Staff agents in both Claude and Copilot. The three items owed from the first session (custom GPTs, now called plugins; artifacts; writing in Karney's voice with cited sources) were rewritten into modules 3 and 4 because the products changed. The remaining modules continue next session.
Questions and answers
No questions were posted for this session. Ask one.