← All sessions
Session 08 · Recap
Materials live

AI without the anxiety

Security, safety and terminology for CRE decision makers. Anthony Peters, BranchPattern's IT manager for seven years, skipped the tool demos and started with the questions that should come first: what the words mean, who can see your data, and how much to trust what comes back.

Friday, October 16, 202669 minutesMicrosoft TeamsAnthony Peters · BranchPattern

Anthony opened by naming a tension he lives with. He called it his split screen. At work he uses AI every day and gets excited about what it can do. Then he goes home, opens social media, or talks to friends who are authors and artists, and hears a completely different story. He did not pick a side. He said both are real, the conversation is polarizing, and how we implement AI and talk about it matters as much as what it can do.

"I think it's more important now than ever that people are in the driver's seat and in the decision making chair."

His other framing came from experience: he remembers dialing into the Wichita State University bulletin board on a 14.4 modem, when email, URL and WWW were brand new words and people felt about them exactly the way they feel about tokens and agents now. The vocabulary settles. The judgment is what you have to build.

Three paths, and the one he argued for

Enthusiastic, anxious, uncertain: organizations feel the same three things individuals do, and leaders make adoption and policy decisions from whichever one they are in.

Technology first

Full speed ahead

Push it down everybody's throat, set token quotas, talk about it constantly, and skip the groundwork of building people's skills and their capacity to tolerate the change.

Liability first

Say no first

Write a pile of reactive policies and refuse things before anyone has discovered what they are. His tell: the headlines about firms announcing a policy and rolling it back three months later.

People first

Move at a human pace

Invest in education, let ideas come from the people doing the client work, and have IT and BranchPattern's Digital Enablement team build on those ideas rather than push from the top down.

"People first to me means moving at a rate that people can handle."

He asked the room which path their own firm was on, and whether there is a fourth he had not thought of. That question is still open. Tell us yours.

— / Recording

Watch it back.

All 69 minutes, including the discussion on Claude versus Copilot and connectors into Microsoft 365.

Watch the recording Session 8 · Anthony Peters · Oct 16, 2026

The vocabulary, the frameworks, the discussion and Anthony's one-page reference are all below. Trouble opening the video? Email Ivy and she will sort your access.

01 / Speak the language

New words for new technology.

Twelve terms that come up in every conversation with IT, vendors and leadership. You do not need to be technical to use them. You need them so you can follow the conversation, and push back in it.

Token

The unit of measure for AI: the pieces of text it reads and writes. Limits and costs are counted in tokens, not words. When a tool says you have hit a limit, this is what it is counting.

Agent

AI given a goal, instructions, tools, and permission to take specified steps on its own. The website chatbot and the voice that answers an 800 number are agents.

Anthony's warning for anyone building one: people remember to say what it should do and forget to say what it should not. A restaurant's ordering bot will happily answer your Python question if nobody told it not to.

Model

Software trained on an enormous amount of text to predict likely next words. It is not a search engine, and it does not look anything up unless you, or the tool around it, tell it to.

Copilot is not a model. It is an application layer that sits on top of models, and so are most of the products you use.

Tenant

Your organization's contractual and technical "fenced yard" in the cloud, governed by its contracts, controls and policies. A boundary, not a guarantee that all data stays inside. More on that below.

Prompt

What you ask, plus the context you give it. Poor output very often traces back to an ineffective prompt rather than a bad tool.

Permissions

AI inherits the access you already have. It can surface anything you could open, including the files you forgot you could open.

Grounding

Pointing the tool at real sources, such as your files or a specific document, so the answer comes from evidence rather than from the model's best guess.

Training and inference

Training builds the model. Inference is you using it. On business plans, prompts and responses are often excluded from training foundation models, but that depends on the exact service and contract. Verify yours rather than assuming.

Context window

How many tokens the AI can work with at one time. Past that, it starts losing track of what you gave it earlier.

Prompt injection

Hidden or malicious instructions buried in a file, email or web page that try to hijack what the AI does next. This matters more the more you let it act on its own.

Hallucination

A confident answer that is simply wrong. It is not a glitch that a future update removes; it comes from how models work. Plan for it.

Instructions

A longer, standing prompt that applies to every interaction with an assistant or agent. Write them once, carefully, instead of repeating yourself in every chat.

02 / Match the cost lane to the task

Three ways to pay for AI.

Each one suits a different kind of work. The mistake is picking the lane first and the task second.

Free · public tools

Brainstorms and first drafts

Best for: explaining public concepts, rewording non-confidential text, getting unstuck.

Watch for: no client or firm data, ever. Inputs may be used for training.

Examples shown: ChatGPT free tier, Microsoft Copilot (consumer), Google Gemini free.

Subscription · per seat

Daily email, meetings and documents

Best for: drafting inside the firm's data boundary, routine research and analysis.

Watch for: idle seats. Value grows with regular use, so an unused license is pure cost.

Examples shown: Microsoft 365 Copilot, ChatGPT Enterprise, Claude Team and Enterprise.

Metered · pay per use

Volume and automation

Best for: high-volume document processing, automated workflows and API integrations, multi-step tasks that run unattended.

Watch for: costs scale with volume. Set budgets and monitor usage from day one.

Examples shown: direct model access through an API, Copilot Studio pay-as-you-go.

Rule of thumb

Data sensitivity decides which lanes are allowed. Task volume decides which lane pays off.

Product names are the examples from Anthony's slide. Plans, names and pricing change often; check current terms before you buy.

How BranchPattern got here

Anthony described the cost lanes as their own journey. Two or three years ago everyone was on the free tools. Then the subscriptions arrived, with Microsoft 365 Copilot at about $30 per user per month when they looked at it, and then the newer agent-style tools that, in his words, give you a taste for free and hit you with the costs later. Metered pricing is not new, he pointed out. Anyone who waited until after 7 PM to make cell phone calls has lived it.

He does not advocate free tools for work, and was candid about why. You already use free AI every time you search Google; it is built in. But:

"If it's free, then you're probably not the client, you're probably the product."

The subscription lane is where business use belongs: sensitive information, questions you would not put into a search engine, data you would not throw onto the open web. And not every role needs AI at the same level, which is the point of thinking in lanes rather than licensing everyone for everything.

The $30 figure is what Anthony cited from BranchPattern's experience. Check Microsoft's current pricing for your plan.

03 / The fenced yard

A boundary, not a free pass.

Your tenant is a contractual and technical boundary around your firm's data. It is not a guarantee that everything stays inside, and an approved work account is not permission to upload everything.

1

What is inside

Business systems and inboxes, secure networks, and the AI your firm has authorized.

2

Where the fence is

Account: an approved work identity. Contract: retention and training terms. Access: file and system permissions. Connections: where data can travel.

3

What moves through the gates

Email, sharing links and attachments. Web search and browsing. Downloads and screenshots.

4

What is outside

Personal or unapproved AI accounts, and external services with their own separate terms.

Every user is a gate

Anthony's image for the tenant: a canvas with a fence around it, and a gate for every person with an account. An email you receive, a sharing link you create, a screenshot, a downloaded attachment: each moves data through your gate. So does pasting firm data into an AI tool that still learns from your prompts. Copilot lives inside a Microsoft 365 tenant by default. A business tool like Claude Team connects to it through a connector someone authorizes, and Anthony's advice is to confirm its training setting is off before you do. Either way, someone decides what crosses the fence.

Why BranchPattern chose Copilot, and what that costs them

He was open that this is a choice with an opportunity cost, made for a Microsoft shop. His reasoning, as he put it:

  • Longevity over features. He is confident LLMs will be around and that Microsoft will be. He is less certain about every model provider's long-term future, and expects models to become commodities that are roughly equal in what they can do. So he would rather build strategy on the more stable layer.
  • Who the vendor serves. Microsoft builds for IT departments and executive teams deploying tools across a firm, with the compliance relationships that come with that. BranchPattern passed on Google's AI products because, in his view, Google's business model is selling data.
  • The control panel. Everything on the fenced-yard slide maps to switches IT can actually turn on and off.

The tradeoff is a feature lag. His observation over the last 12 to 18 months: a model provider announces something, the early adopters ask when they can have it in Copilot, and three to six months later it usually shows up. Not every feature, but a lot of them.

These are Anthony's views on vendor strategy for BranchPattern, shared as one practitioner's reasoning. Your firm's answer depends on where your work already lives.

Access follows sensitivity

Inside the yard, data sits at different levels. The AI does not know which is which. It only knows what you can open.

Level 1 · Public

Everyone in the firm

Policies, templates, market research.

Level 2 · Internal

Department or team

Budgets, pipeline, operations.

Level 3 · Confidential

Client or deal team

Leases, LOIs, financials.

Level 4 · Restricted

Named individuals

HR, legal, compensation.

Before a team pilot

AI can surface anything you already have permission to see.

Anthony's example: someone prompts Copilot and gets back an executive slide deck nobody realized they could open. His advice to any decision maker is to ask for an audit report of who has access to what before deploying AI across the board. AI did not create the access. It just lowered the barrier to finding it. Three checks before you turn anything on for a team:

1. Check broad sharing links. "Anyone in the organization" links on deal folders are the usual culprit. 2. Separate client and deal files from general team space. 3. Give connectors only the access they need, not the access that was easiest to grant.

And keep client or firm data out of unapproved public tools entirely. Consumer and business terms differ, and web search and connected agents may handle data differently from the core tool. Check the exact service and connection.

04 / Two different questions

Security and safety are not the same thing.

They get blurred together constantly. Both matter, for different reasons, and passing one tells you nothing about the other.

Security: who can access the data, and where can it go?
  • Account and contract
  • Permissions and connectors
  • Retention and sharing

Lease exampleA confidential lease uploaded to an unapproved account. The summary may be perfect. The data has still left the yard.

Safety: is this output reliable enough for the intended use?
  • Evidence and completeness
  • Professional review
  • Authority to act

Lease exampleThe AI summarizes the lease, in the approved tool, and leaves out a renewal condition. Perfectly secure. Still wrong.

A secure tool can still be wrong. A correct answer can still disclose data improperly.

Security, he said, is an IT discipline that has been around a long time and applies to AI the way it always has. Safety is newer: how do you catch a hallucination on a topic you are not an expert in, what review has to happen, and what authority does it have to act? If you give it access to your email, is it going to delete your whole inbox because it hallucinated a command? Do not confuse the two.

05 / From tool to delegate

Delegate the task. Keep the accountability.

Using a tool and delegating a task are different jobs. The work moves faster when you delegate. The accountability does not move at all.

Using a tool
  • You do the work; the tool amplifies your effort.
  • Results depend on your skill and time.
  • Direct, step-by-step control.
  • Success means using the tool correctly.
Delegating a task
  • Define the outcome and the standard.
  • Provide context, inputs and constraints.
  • Review and verify what comes back.
  • Accountability stays with you.

The old model of software was a hammer: it does exactly what you tell it, and if you hit reply all, that is on you. AI puts an abstraction layer between what you ask and what you get back, with uncertainty baked in. Treat it as delegation and you start thinking like a manager: give it resources, context and accountability, then circle back and check.

"It is a tool. I mean, it's not sentient. It's not its own thing. But it kind of operates like a person sometimes."

Which is why people first is not a slogan

Announce a program to turn everyone into "managers of AI agents" and the first thing a line-level employee hears is: here comes AI for my job. And not everyone has managerial skills yet. Teaching people to delegate well, and how you communicate the rollout, ranks an order of magnitude above which tool you pick or which prompts you use. They are the ones prompting.

Sarah's takeaway from the same point: people say "I had AI do this" almost as a disclaimer.

"We don't disclaim ourselves away from accountability and work products with any other tools we're using."

Before you use it, or let it act

Four checks from Anthony's follow-up reference, in order.

01

Approved tool, work account

Is this the environment your firm approved for this kind of work, and are you signed in to the work account, not a personal one?

02

Permitted data, permitted access

Is the data you are about to use allowed in this tool, and does the tool have only the access it needs?

03

Ground it, then check the evidence

Point it at the real source. Then check that the answer is supported by that source, and that nothing material is missing.

04

Professional review before consequential action

A human with the right expertise reviews it, and a human approves anything that sends, signs, commits or changes something.

06 / Discussion

What the room asked.

The last twenty minutes, lightly condensed. The chat spent much of the hour on Claude versus Copilot, so that is where the discussion went.

"Who makes the first permission calls: IT, or the business user?" Michelle Hamilton

It depends on your culture and how you already make decisions. AI should not force a pivot into a new one. BranchPattern's values include being progressive, so they start open and add constraints that make sense. A military base or a hospital would start the other way, and could still be people first.

He expects a messy phase: experiments, stray Teams, groups nobody can explain. Then comes a reining in, and that is a group decision with executives, IT and the people actually using the tools in the room. Without them, the decisions get wasted. Firms with heavy silos, he warned, will hit communication problems in any AI rollout.

Michelle added the other side of the ledger: at some point all those artifacts cost real tokens, to the company and to the planet. The skill is knowing how much is too much without killing the creative part.

Claude or Copilot: why do they feel so different?

Anthony's read: the experience gap is real, and BranchPattern users notice it too. The newer model companies are competing hard for customers and put them on the leading edge of what their AI can do. Microsoft's game is different: it builds for the IT and executive buyers worried about everything in this session, which means a lag in features.

BranchPattern pays for Copilot firm-wide and a small number of Claude seats, mostly for its Digital Enablement team's coding work. Staff are not stopped from buying a personal subscription, but they are taught to treat it like a personal email account: you do not upload contract documents to it.

"We have to make a decision that leads the whole organization, and not just one person, one department, one prompt."
"If we use Claude, can it get into our Microsoft 365 data?" Andrea

Not by default. Like any application, it only reaches your data when someone builds a connector, which authenticates one application to another under the name of whoever approves it.

BranchPattern has turned off user-level approval for outside apps, so a connection request escalates to Anthony to approve or deny. Once approved, the trust relationship is set, and the tool can reach everything that person can. He compared the approval screen to installing a phone app that asks for your camera and microphone: you are making a trust decision, and AI cannot make it for you.

Microsoft Entra: the admin consent workflow ↗

"Do I need to keep track of which model does what?" Melody Montgomery

No. Nobody can; tens of thousands of people are working on these every day. That is what groups like this one are for. Anthony learns from early adopters on his own team rather than chasing every release himself.

What leaders should watch are the mile markers: the moments it can suddenly do something new that changes the game. Lean on your team and your community to flag those.

Melody's version: forums like this show you things you would never have thought to ask it.

"Can a rule make it always research a question instead of just answering?" Sarah Gudeman

Yes, that is what custom instructions are for, or an agent built for one kind of task. It interprets your phrasing heavily, and a question you type may not mean "go search," so spell it out.

"Can it grade my prompts?" Sarah Gudeman

Anthony's answer went to cost: the less specific you want to be, the more you will likely pay. In his experience Copilot needs more specificity, while Claude handles vaguer prompts better. Knowing that can open up different tools at different price points.

Worth trying regardless: ask your assistant to critique a prompt before you run it. Prompt 05 below is a starting point.

07 / The prompts

Put it to work.

Anthony's session was frameworks, not prompts. These are AI × CRE working prompts that turn his checks into something you can run today. Use them in your firm's approved tool, and edit them before you send them.

08 / What to take away

Six ideas for your next AI decision.

01

Choose the people-first path

Lead with people, not hype or fear, for sustainable growth and sensible policies.

02

Match the cost lane to the task

Free, subscription and metered tools suit different work. Data sensitivity sets the limits.

03

Speak the language

Tokens, models, prompts, grounding and hallucinations shape what AI can and cannot do.

04

Stay inside the fenced yard

Keep firm data in the approved environment, under your firm's contracts and controls, and know what your contract actually says about training and retention.

05

Access follows sensitivity

AI can surface anything you can already open. Tighten sharing before a pilot.

06

Delegate, do not just use a tool

Define the outcome, give context, check the result. Security and safety both stay with you.

Put people first, keep data in the right place, and keep human judgment in charge.

Next month: From one or two design options to five to ten. Sandeep Ahuja, CEO of cove.inc, on what AI changes in the design phase and what to ask for. Friday, November 20, 10:00 AM ET. See the session and register.

— / Materials

The leave-behind.

Anthony's one-page reference, the vocabulary and discussion on this page, and the prompt library. Free, no gate.

Presenter
Anthony Peters
IT Systems Manager, BranchPattern

Anthony has been BranchPattern's IT manager for seven years and has worked in technology far longer, starting on dial-up bulletin boards. BranchPattern recently ran its own internal AI education sessions, and much of this material came from them. His conviction, repeated through the hour: invest in people's skills and knowledge, because people need to stay in the decision-making chair.

With thanks to