From Buzz to Business
Everything from our August 12 fireside chat, expanded with the answers we ran out of time for. Search it, skim it, or open only the parts you need.
Thank You for a Great Conversation
Twenty-nine of us spent ninety minutes on this, and the room did most of the heavy lifting. What follows is the session itself plus everything that came out of your questions. Nothing here requires you to buy anything or hire anyone.
A note on the numbers: Every statistic on this page names its source and the year it was collected. Where a figure is older than it looks, or where the sample skews in a particular direction, that is said out loud. You should be able to check anything here yourself.
Know Your Business Before You Touch AI
The businesses that get the most out of AI are not the ones with the fanciest tools. They are the ones who can describe their own workflow in plain language first.
You can't hand off what you can't describe.
Speed Applied to a Broken Process Is Just Faster Chaos
If you bolt AI onto a process that is already inefficient, or one that needs constant correction along the way, you do not get efficiency. You get the same mess arriving faster and in higher volume. AI adds velocity. Velocity in the wrong direction is not a gain.
This is the single most common reason people try AI, get frustrated, and quietly stop. It usually is not the tool. It is that the tool was pointed at a process nobody had ever written down.
Before you automate anything, describe it. If you cannot explain the task to a sharp new hire in five sentences, you are not ready to explain it to AI either.
Three Questions That Find the Work Worth Handing Off
These are the same three questions used in a paid audit. They cost nothing and they work just as well on a legal pad.
- What is not you? What sits on your plate that does not actually require your specific expertise, judgment, or relationships?
- What repeats? What do you do the exact same way, week after week, with only the details changing?
- Where did you get stuck? Walk through yesterday start to finish. Find the moment you stalled, and ask why.
Answer all three in writing this week. Anything that shows up in two of the three columns is your first AI project.
What Would You Eliminate Forever?
This was the first question asked in the chat, and it is the fastest way to find the real friction in a business. If you could wave a wand and delete two or three tasks from your week permanently, what would they be?
People answer this question honestly in a way they do not answer "where are your inefficiencies." The answers from the room clustered hard around data entry, meeting notes, and inbox triage. Every one of those is solvable today.
What Actually Gets Handed Off, and What Never Does
Reasonable to Hand Off
- First drafts and outlines
- Reformatting information between systems
- Summarizing long documents and calls
- Research starting points you will verify
- Repetitive, rules-based sorting
- Turning rough notes into clean structure
Stays With You
- The final read before anything leaves
- Anything involving reading a person
- Numbers that carry legal or financial weight
- Your actual voice with a client who knows you
- The judgment call about what matters
- Accountability when something goes wrong
The pattern from the room: A recruiting professional on the call described AI ranking applicants well but flagging a strong candidate as unfit because they lived four hundred miles away. The system had no way of knowing the person was open to relocating. That is the line. AI is good at the volume. It is not good at the thing you would have asked in a two-minute phone call.
● Worked example: the multi-source document problem
One attendee walked through a live example on the call. Information arrives from three directions at once: a client questionnaire, emails from a colleague, and phone calls. It all has to end up in one document formatted exactly the way a specific reviewer expects it. AI organizes the information, but not into the shape the reviewer wants, so every submission turns into a back-and-forth. Sometimes the AI invents details along the way.
This problem is not unique to insurance. Swap the words and it is a loan file, a client onboarding packet, a proposal, a board report, or a permit application. The fix is the same in all of them.
The fix has three parts:
- Train the format, not the file. Sit down when you are not under deadline and teach the AI what the finished document looks like. Give it the structure, the section order, the required fields, the language the reviewer expects.
- Use fake data to do it. Have the AI generate its own sample client, sample amounts, sample details. You get a real training session with zero client data exposure. This solves the privacy problem and the training problem at the same time.
- Make it ask you questions. Instead of handing over everything and hoping, tell it to interview you for what it is missing before it produces anything. This is where most hallucinations get caught, because a system that asks is a system that is not guessing.
Realistic time investment: one to two focused hours. That is not a small ask on a busy week. But it is a one-time cost against a task you repeat, and once the format is locked in it is repeatable and adaptable. Save the result as a project or a saved instruction set so you never rebuild it.
The prompt for exactly this is in the Prompting section below, template 02. Copy it and swap in your own document type.
The Adoption Gap
Buying the tool is not the finish line. Building the habit is. Here is the number that made the point on the call.
Using It
Running On It
Source: Goldman Sachs 10,000 Small Businesses Voices, survey of 1,256 small business owners across all 50 states, DC and Puerto Rico, fielded by Babson College and David Binder Research, January 27 to February 4, 2026.
Honest caveat on that 76%: The Goldman Sachs network skews toward higher-growth operators, so it runs ahead of broader samples. The Federal Reserve's 2026 Report on Employer Firms found 46% of small employer firms currently use AI, with another 15% planning to start within twelve months. Both numbers are real. The 14% integration figure is the one that matters here, and it points the same direction regardless of which adoption number you prefer.
The gap between using it and running on it is the whole game.
Why the Gap Exists
Almost nobody in that 62-point gap failed because they picked the wrong model. They failed for one of four reasons, and every one of them showed up in our conversation:
- No documented process. The AI was pointed at something nobody had written down, so it guessed, and the guesses had to be corrected.
- Correction cost exceeded the savings. Once fixing the output takes longer than doing the work, people stop. Rationally.
- It was one person's side project. One enthusiast built something useful, nobody else adopted it, and it died when they got busy.
- No trust framework. In a regulated business, "is this allowed" went unanswered, so the safest move was to not use it at all.
Notice what is not on that list: which AI company you chose. Tool selection is the decision people agonize over and it is the one that matters least.
Choose Your Tool, Fast
Less bake-off, more "which tool for which job." You do not need all five. You need to know what each one is actually for, and then pick one and get good at it.
| Tool | Strongest At | Where It Lives | Worth Knowing |
|---|---|---|---|
| Claude | Writing, reasoning, long documents, nuance | Web, desktop app, mobile | Handles large amounts of information well. Can get expensive if you are a heavy user. No native image generation. |
| ChatGPT | Broadest general-purpose assistant | Web, desktop, mobile | Largest app and plugin ecosystem. Its image generation is among the best available, but it needs very precise instructions, especially about what you do not want. |
| Gemini | Anything already inside Google | Gmail, Docs, Sheets, Drive | The natural fit if your business runs on Google Workspace. Recalling what a client conversation covered months ago is a standout use. |
| Grok | Real-time information and trending topics | Built into X | Strongest when the question is about what is happening right now on social platforms. |
| Copilot | Working inside Microsoft 365 | Outlook, Word, Excel, Teams | If your company already pays for Microsoft 365, you may already have access. See the section directly below before you assume anything about the privacy side. |
All of them can now search the live web. That was not true a couple of years ago. It also means some websites block AI crawlers. If your own site blocks them, you will not show up when a prospect asks an AI for a recommendation in your category. Worth checking.
Copilot Is Probably in Your License. Most People Never Turn It On.
If your organization pays for Microsoft 365, there is a good chance some level of Copilot is already available to you. The gap here is not the purchase. It is the habit.
Sources: Recon Analytics survey of 150,000+ U.S. respondents, January 2026 (35.8% Copilot workplace conversion vs. 83.1% for ChatGPT). Seat count per Microsoft Q3 FY26 earnings disclosure, April 29, 2026.
The Copilot Privacy Question, Answered Properly
On the call the shorthand was that if your company provides Copilot, they are handling the data privacy for you. That is often true, and it is frequently the safest place to start. But it is not automatic, and if you work in a regulated field you should not repeat it to your compliance officer without checking.
The reason is that "Copilot" covers several different products with different data handling. A licensed Microsoft 365 Copilot deployment under a commercial agreement behaves very differently from the free Copilot Chat, and both depend on how your organization has configured its tenant.
Send these four questions to whoever manages your Microsoft environment. The answers take them two minutes and they settle the question permanently.
- Which product do we have? Licensed Microsoft 365 Copilot, or the free Copilot Chat tier?
- Is our data covered by the commercial data protection terms, and does anything I type leave our tenant?
- What is our retention policy on Copilot prompts and responses?
- Are there categories of data our policy says I should not put into it, even inside the tenant?
Why this matters more than it sounds: Several people in the room work in wealth management, tax, payroll, and accounting. In those fields the difference between "our IT department approved a tool" and "our IT department approved this specific tier of this tool for this category of data" is the entire compliance conversation.
● How to pick one in thirty seconds
- Your company runs on Microsoft 365 and you handle sensitive data. Start with Copilot. Ask the four questions above first.
- Your business runs on Google Workspace. Start with Gemini. It already has context you would otherwise have to paste in.
- You write a lot, or work with long documents and contracts. Start with Claude.
- You want one tool for a bit of everything, including images. Start with ChatGPT.
- You need to know what is being said right now. Grok, for that specific job only.
Pick one. Use it for thirty days before you evaluate a second. Switching tools every two weeks is how people spend six months learning nothing about any of them.
Claude
ChatGPT
Microsoft Copilot
Email Is the Proving Ground
When we polled the room, 18 of 29 people named email as their main AI use. That is not a coincidence. It is the one workflow every business in the chamber shares, and it is the best place to build a habit that sticks.
The 28% figure comes from McKinsey Global Institute, "The Social Economy: Unlocking Value and Productivity Through Social Technologies," July 2012. It remains the most-cited workweek benchmark because McKinsey has not republished a comparable study, but it is an older number and worth labeling as such. The volume and interruption figures come from Microsoft's 2025 Work Trend Index, based on analysis of Microsoft 365 signals.
What Changes When the Habit Sticks
Before
- Every reply drafted from a blank page
- Follow-ups depend on you remembering
- Inbox triaged by whatever arrived most recently
- Important messages buried under newsletters
After
- Replies drafted and waiting for your review
- Follow-up sequences that fire on their own
- Inbox sorted by what actually needs you
- Noise routed away before you ever see it
Automation and AI Are Not the Same Thing
Automation is "if this, then that." A rule fires the same way every time with no thinking involved. Routing every message containing the word "unsubscribe" into a holding folder is automation. It is fast, free, completely predictable, and it never hallucinates.
AI is what you add when judgment is required. A message contains the word "unsubscribe," but it came from a client you have forty threads with. A rule sends it to the junk pile. Reasoning keeps it in your inbox.
Most people reach for AI when a rule would have worked better. Set up your rules first. They cost nothing, they run instantly, and they make anything you layer on top dramatically more effective, because the AI is no longer wading through mail you already knew was noise.
Spend twenty minutes in your mail settings building rules before you spend twenty dollars on a tool. Then point the AI at what is left.
Tips From the Room
These came from your peers on the call, not from a vendor. They are free, they work, and several take under fifteen minutes to set up.
- Tom's client folders. Create a folder per client and a rule that routes their mail there automatically. Check those folders throughout the day and treat everything else as lower priority. He described going from scanning twenty or thirty messages to a five-second glance at the left sidebar.
- Cheryl's unsubscribe rule. Any message containing the word "unsubscribe" goes to a separate folder. Scan it quickly, delete almost all of it. Simple, and it removes most of the daily noise without a subscription to anything.
- Brittany's dropped-ball check. Ask Copilot to scan your inbox for anything awaiting your reply that fell through the cracks. One question, and it catches what you missed.
- Leandra's sounding-board approach. Do not hand over the whole email. When a single point is not landing, ask the AI for five ways to say that one thing. You keep your voice and get the help exactly where you needed it.
- Imran's polish pass. Write it yourself, then run it through for cleanup and clarity before sending. The thinking stays yours, the friction goes away.
- Ajo's memory recall. Ask Gemini what you and a client discussed a month ago and get a summary of topics and open items pulled from your own Google mail history. Especially useful for relationships you touch sporadically.
- Doug's named experts. Build a few dedicated assistants with standing instructions, one for time management, one for leadership, and go to the right one instead of re-explaining context every time.
- Matt's benchmarking. Not email, but too good to leave out. He asked AI to analyze top-performing job ads from established agencies in his market and extract the keywords they used. Result was roughly triple the candidates on a three-day ad, at lower spend, from about thirty minutes of work.
● "It doesn't sound like me." How to fix that permanently.
This came up more than once, and it is the most legitimate objection on the list. One attendee put it perfectly: nobody comes to them for a cookie-cutter approach, so sending cookie-cutter writing actively damages the thing they sell. Another mentioned being told by younger colleagues that it is obvious when AI wrote something.
They are right, and the cause is specific. Out of the box, these tools write in the average of everything they have read. Average is exactly what you do not want. The fix is not better prompting in the moment. It is giving the system a permanent reference for how you actually sound.
Do this once and it applies to everything afterward:
- Collect five to ten emails you actually wrote and were happy with. Real ones, not polished ones. Strip out anything client-identifying.
- Have the AI analyze them and write you a description of your own voice: greetings, closings, sentence length, level of formality, what you never say.
- Correct that description. It will get a few things wrong. This step is the whole exercise.
- Save it as standing instructions in a project, gem, or custom assistant so it loads every time instead of being pasted every time.
- Give it an explicit banned list. Phrases you would never use. "I hope this email finds you well" and "at your earliest convenience" are on most people's list. Naming what to avoid does more work than describing what to aim for.
Template 03 in the Prompting section does steps one through four for you.
A tool used once isn't a system. It's a novelty.
Prompting: The RTCROL Framework
How you ask dictates what you get. This is the framework, and then a library of prompts you can copy directly. The first one is the most useful thing on this page.
| Letter | What It Means | What It Sounds Like |
|---|---|---|
| R — Role | Who should it act as? | "You are an experienced commercial insurance underwriter." |
| T — Task | What exactly needs doing? | "Review this submission and flag anything missing." |
| C — Context | The background in your head that it cannot see. | "This goes to a carrier that rejects files without loss runs." |
| R — Reasoning | How should it think it through? | "Work section by section, and check each against the requirements list before moving on." |
| O — Output | The exact format you want back. | "A table with three columns: section, status, what is missing." |
| L — Limitations | The guardrails. What to avoid or never do. | "Do not invent figures. If something is missing, say missing rather than estimating." |
The one exception, and it matters: Do not assign an expert role for anything mathematical. If you tell a model it is an expert mathematician or a CFO and then ask it to calculate, the effect across current models is that it becomes more confident in wrong answers rather than more accurate. For calculations, skip the role entirely. Go straight to the task, give it context, specify the output, and set your limitations. This does not apply to writing, strategy, or outreach, where a role helps considerably.
Not everything needs all six. "Summarize this email" does not need a framework. RTCROL earns its keep when you are at the doorstep of something larger: a repeated process, a document that has to be exactly right, or a workflow you intend to reuse.
The Prompt Library
Click any template to open it. Each one has a copy button. Replace anything in [square brackets] with your own details.
01 The Prompt Builder — start here
This is the one to keep. Instead of learning to write RTCROL prompts yourself, you paste this once and the AI builds them for you. Give it a rough description of what you are trying to do, in whatever messy language comes out, and it interviews you until it has what it needs, then hands you a finished prompt you can use anywhere.
It is deliberately built to ask before it writes. A prompt builder that skips the questions just invents your context, which is exactly the failure mode we are trying to avoid.
02 Train it on a document format, using fake data
For any recurring document that has to be built a specific way for a specific reviewer: submissions, loan files, onboarding packets, proposals, board reports, permit applications. The fake data step means you can do the entire training session without exposing a single real client detail.
03 Teach it to write in your actual voice
The fix for "it doesn't sound like me." Run this once, save the result, and every draft afterward starts from your voice instead of the internet's average.
04 Weekly performance readout from a spreadsheet
For anyone tracking numbers across locations, providers, reps, or branches and distributing a summary. Note there is deliberately no expert role assigned here, for the reason covered above.
05 Turn messy notes into owners and deadlines
Works on meeting notes, call recordings, transcripts, or whatever you scribbled during a conversation.
06 Outreach strategy for a specific target
For business development. This one does use a role, and should. Roles work well for strategy and writing. The math warning does not apply here.
07 Benchmark your listings against the market
Built from the job-ad example shared on the call, but the pattern works for any public listing you compete against: job postings, service pages, property descriptions, event pages.
Inside the Demo
The live walkthrough was in Claude because that is what I use daily, but nearly everything shown has a direct equivalent in ChatGPT and Gemini. Written out here so you can follow along on your own screen, and in case the screen share did not come through clearly on your end.
| What It Does | In Claude | In ChatGPT | In Gemini |
|---|---|---|---|
| Standing instructions for a recurring type of work | Projects | Projects or custom GPTs | Gems |
| Instructions that apply to every conversation | Personal preferences in settings | Personalization in settings | Saved info |
| Working with files on your own computer | Cowork in the desktop app | File upload, connectors | Drive integration |
| Connecting to outside services | Connectors and plugins | Apps and connectors | Workspace extensions |
| Speaking instead of typing | Dictation and voice mode | Voice mode | Voice input |
| Turning web search on or off | Toggle near the message box | Toggle near the message box | Generally automatic |
Projects, Gems, and Custom Assistants
This is the feature most people have never opened, and it is the one that changes the most. A project is a workspace with permanent instructions attached. You tell it once what you are working on, who you are, how you want output formatted, and what to avoid. Every conversation inside that space inherits all of it.
The practical effect is that your prompts get much shorter. Instead of re-explaining your business every time, you walk in and describe the task. Role, context, reasoning, and limitations are already loaded. You supply the task and the output you want.
You can also attach a folder. Drop files into it on your computer and the project sees the current versions without you re-uploading anything.
Create one project this week for the task you do most. Put your standing instructions in it. That single setup does more for output quality than any prompt you will write.
Chat vs. Cowork, and Why the Difference Matters
Chat is what most people use. You bring the information to it, by pasting or uploading. Nothing on your computer is visible unless you hand it over.
Cowork in the desktop app can be pointed at a folder on your own machine and work with what it finds there. Useful when the relevant material is spread across many files. It also means you need to be deliberate, because everything in that folder is in scope.
Set approvals to manual when you start. There are three levels: approve each action individually, auto-approve, or skip approvals entirely. Manual is slower and it is the right setting until you genuinely understand what the tool does with access. If you handle client PII, do not point it at a folder containing that data without a clear internal policy on what is allowed.
Skills, Connectors, and Plugins
- Skills capture something you do repeatedly so the tool does not rebuild the approach from scratch each time. The example from the call: an introduction email skill that always includes both parties' contact details, states plainly why they are being connected, and adds context on each relationship. The prompt becomes two names instead of a paragraph of instructions.
- Connectors link the assistant to an outside service so it can go retrieve what it needs and come back.
- Plugins extend that further and can run multi-step workflows rather than just fetching information.
Do not start here. These are worth setting up after you have a process that works manually. Automating a workflow you have not yet proven is how you end up with a fast machine producing the wrong thing.
You Do Not Need the Most Powerful Model for Everything
Most platforms now offer several models at different capability levels, and most people leave it on the strongest one permanently. That costs more, runs slower, and rarely produces a better answer for ordinary work.
In Claude at the time of the session, the tiers ran from Fable 5 at the top, then Opus 5 for complex work, Sonnet 5 for everyday tasks, and Haiku 4.5 for quick answers. Most people should be living in the everyday tier. Some platforms also expose an effort or thinking-depth setting, which is a second dial worth turning down for routine work.
The fastest tiers work noticeably better inside a project, because the standing instructions do the work that a larger model would otherwise have to infer.
These names will change. Model naming across every platform turns over every few months. The principle outlasts the names: match the model to the difficulty of the task, and check what tier you are on before a long session.
Stop Typing. Start Talking.
Dictation is available in every major platform and almost nobody uses it. Two things happen when you switch.
First, you give far more context, because you are not filtering your thinking down to what you are willing to type. Second, you are no longer limited by typing speed, so the tool gets a fuller picture of what you actually want.
Rambling is fine. A disorganized two-minute explanation beats a tidy two-sentence one, because the model can organize but it cannot read your mind.
● The system prompt I use, and why it is short
Every platform has a place for instructions that apply to every conversation you have. In Claude it is under personal preferences, in ChatGPT it is under personalization, in Gemini it is saved info. Most people leave it empty.
Mine is two sentences. Instructions do not need to be long. They need to be the things that change the model's behavior most.
Write your own rather than copying mine. Two or three sentences about how you want to be treated and what accuracy means in your work will change every conversation you have from that point forward.
Protect Yourself for the Long Game
Data privacy came up more than any other concern on the call, from tax, payroll, wealth management, insurance and accounting. This section is the part of the page most worth forwarding to whoever makes technology decisions where you work.
SOC 2 Is a Starting Point, Not the Answer
On the call, SOC 2 and GDPR came up as the signals to look for when you are evaluating software that has AI built into it. That is genuinely useful and it is the right instinct. It also deserves more precision than a spoken answer allowed, because the two things people most want to know are not what those terms actually cover.
SOC 2 is an audit of security controls. It tells you a third party examined how the vendor protects data. It does not tell you whether your inputs are used to train their models, and it does not tell you how long they keep what you send.
GDPR is a regulation, not a certification. Nobody holds a GDPR certificate. A vendor claiming to be GDPR compliant is making a self-assessment. It signals they have thought about data protection, which is worth something, but it is a claim rather than a credential.
Both are worth checking. Neither answers the question you actually care about, which is what happens to your client's information after you type it.
The Questions That Actually Settle It
Ask these of any vendor whose product touches client information. Every one has a documented answer, and a vendor who cannot produce it quickly has told you something.
| Ask This | What You Are Looking For |
|---|---|
| Do you train on our inputs? | A clear no for business and enterprise tiers, in writing. This is the single most important question and it is often answered differently for consumer versus business plans of the same product. |
| How long do you retain prompts and outputs? | A specific duration, and whether you can shorten or disable it. |
| Will you sign a data processing agreement? | Yes, with terms your counsel can read. For regulated work this is usually non-negotiable. |
| Which tier are we actually on? | Consumer, business, and enterprise tiers of the same product often have materially different data terms. People assume they have the protections of a tier they are not paying for. |
| Where is our data stored? | Geography matters if you have clients or obligations abroad. |
| Do subprocessors see our data? | A named list. Many AI features are built on another company's model underneath. |
| What admin controls do we get? | Whether you can restrict features, see usage, and remove access when someone leaves. |
The pattern to internalize: Paid does not automatically mean private, but the terms attached to business and enterprise tiers are usually meaningfully different from the free version of the same product. If you are handling client PII on a free consumer account because it seemed like the same tool, that is the gap worth closing first.
What Not to Paste Into a General Consumer Chat Window
- Social security numbers, tax IDs, and account numbers
- Full client names paired with financial details
- Medical or health information
- Anything covered by a confidentiality agreement you have signed
- Complete contracts containing identifying party details
- Credentials of any kind
The workaround is almost always simpler than it sounds. Replace real identifiers with placeholders, run the work, and put the real details back yourself. For training the AI on a process, use invented data entirely. You get everything you need without the exposure.
● Running AI on your own hardware, so nothing leaves
Someone on the call asked whether it is possible to run this entirely on internal servers so no client data ever reaches an outside company. It is, and it is more accessible than most people expect.
Smaller open models are lightweight enough to run on ordinary business hardware. They will not match the frontier tools on hard reasoning, but a great deal of routine business work does not need frontier reasoning. Summarizing, reformatting, drafting, extracting, and classifying are all well within reach.
- Ollama is the most common way to download and run models locally. Straightforward to install.
- LM Studio does the same with a friendlier interface, better if you would rather not use a command line.
- Gemma is Google's family of open models, available in sizes that run on modest hardware.
- Google AI Edge Gallery runs a small model on your phone with no internet connection required, which is also handy in areas with poor reception.
Check hardware requirements before you commit. The tradeoff is real: you get complete data control and no per-use cost, at the price of weaker capability, setup effort, and maintenance being yours. For a business with genuine data residency requirements, that trade is often worth making. For everyone else, a properly configured business tier of a commercial tool is usually the better answer.
Tokens, and Why Your Bill May Not Stay This Low
A token is a chunk of text, usually a fragment of a word. These systems break language into tokens to process it, then convert tokens back into the text you read. Everything you send and receive is measured this way.
The useful analogy is cell phone minutes in the early 2000s. New technology is expensive at first, then the cost falls. We are early. Right now the AI companies are absorbing a significant share of what your usage actually costs, because they want you to build the habit.
That subsidy is a business decision, not a permanent feature. If you are building AI into a core operation, model what happens if the per-use cost rises materially. It may not. But a process you cannot afford to run is not a process you own.
Skill Atrophy Is Real, and It Is Measured
One attendee said plainly that he can still write a good personal note and does not want to lose that, so he deliberately reverts to doing some things himself. Several people agreed immediately. He is right, and the data backs him up.
Source: GoTo Pulse of Work 2026, conducted with Workplace Intelligence. Survey of 2,500 employees and IT decision-makers across ten countries, November 2025 through January 2026.
My own practice: two days a week I do not use AI to draft or write emails at all. I am less efficient on those days and I plan around it. It is a deliberate cost, paid to keep a skill I am not willing to lose.
Pick one skill you would be unhappy to lose and schedule regular time doing it manually. Two days a week works. One day works. Zero days is a decision too, just an unexamined one.
Audit Everything. Trust, But Verify.
This was said several times on the call by people in accounting, wealth management, and operations, and it is the point I would most want to survive from the session.
Sources: GoTo Pulse of Work 2026 (87%, survey of 2,500 employees and IT leaders, November 2025 to January 2026). Connext Global 2026 AI Oversight Survey (37%, 17%, 70%, survey of 1,000 U.S. adults who use AI at work, fielded via Pollfish, January 2026).
One attendee described exactly this failure: asking for research, getting an answer with a citation, opening the cited article and finding it said the opposite. That is not a rare edge case. It is the most common way these tools fail, and it is invisible unless you open the source. If a claim matters, click the link.
Trust the tool to draft. Trust yourself to check.
Questions From the Room
These are the actual questions asked on August 12, plus several we ran out of time for. Click any question to open the answer.
I only have time to change one thing. What should it be?
Create one project or gem for the task you repeat most, and put standing instructions in it. Fifteen minutes of setup, and it improves every conversation you have inside that space from then on. It beats learning prompt tricks by a wide margin.
If I sign into Copilot with my Google account instead of Microsoft, do I get different answers?
No. The sign-in method is only there to identify you. Copilot serves whatever model it is configured to use regardless of which account you logged in with.
That said, use your organization's Microsoft sign-in if you have one. Not because the answers differ, but because signing in through your organization is what puts you under your company's data protection terms and admin controls. Signing in personally may leave you on a consumer footing without realizing it.
Wait, you said not to give it a role. But I tell Gemini it's a master connector and it works well. Which is it?
Keep doing exactly what you are doing. Roles work well and often improve results for writing, strategy, outreach, and anything requiring a point of view.
The warning applies only to math. If a task involves calculation or financial figures, do not tell the model it is an expert mathematician or a CFO. What researchers observe across current models is that assigning mathematical expertise makes the model more confident in wrong answers rather than more accurate. For anything computational, skip the role and go straight to the task.
Can I run AI entirely on my own servers so client data never leaves the building?
Yes. Ollama and LM Studio both let you download and run open models locally, and Google's Gemma family includes lightweight versions that run on ordinary hardware. Google AI Edge Gallery does something similar on a phone, with no internet connection needed.
The tradeoff is capability. Local models are meaningfully weaker than frontier tools on hard reasoning, and setup and maintenance become your responsibility. For summarizing, reformatting, drafting, and classification they are perfectly adequate. Full details are in the Protect Yourself section above.
My numbers and dates come back wrong maybe a third of the time. Why, and what do I do?
Because these systems are predicting plausible text, not calculating. A date that looks right is, to the model, as good as a date that is right. This is the failure mode that matters most in accounting, finance, and anything client-facing with figures in it.
Four things reduce it substantially:
- Do not assign a math expert role. Covered above.
- Give it structured data, not prose. A spreadsheet produces far fewer errors than a paragraph describing the same numbers.
- Require it to restate its work. Tell it to show the totals it calculated so you can check them against the source.
- Forbid estimation explicitly. "If a value is missing, write MISSING. Do not estimate or infer." Without that instruction the default behavior is to fill the gap.
Template 04 in the prompt library has all four built in.
Can I get it to pull data from my business system automatically on a schedule?
Yes, and someone on the call described trying and abandoning it after running into more errors than the time savings justified. That was a sensible call, and the current workaround of exporting to a spreadsheet and handing that over is a perfectly good process. Stable and working beats elegant and fragile.
The reason the direct approach struggles is that pointing a general-purpose assistant at a web application and asking it to navigate is brittle. Interfaces change, sessions expire, and pages load unpredictably. Scheduled automation for this normally requires a purpose-built integration through the system's data connection rather than a chat assistant clicking through screens. That is a real project rather than an afternoon, but it is well-established work and it is not exotic.
Is there an out-of-the-box tool for inbox management, or do I have to build it?
There are several. The one I use and can speak to from experience is Fyxer, which handles inbox organization, drafts replies, and now includes a note taker. It takes patience to set up so the drafts sound like you, and you still review everything before it goes out.
Before you pay for anything, spend twenty minutes on mail rules. Tom's and Cheryl's approaches in the Email section are free and solve a real share of the problem on their own.
Disclosure: The link below is my referral link. It gives you $25 off and I receive a referral credit if you sign up through it. You should know that before you click it. If you would rather not use a referral link, go directly to fyxer.com and you will get the same product. I am recommending it because I use it, not because of the credit.
Fyxer, $25 off through my referral link →
Check their current pricing on their site rather than taking a number from me. Software pricing changes and I would rather you see it from the source.
Is the free version good enough, or does paid actually give me more privacy?
Two different questions, and they have different answers.
On capability: Free tiers are genuinely useful and are the right place to start. You will hit usage limits and lack access to the strongest models, but for learning what these tools do, free is enough.
On privacy: Paid is not automatically more private. What matters is which tier you are on, because business and enterprise tiers typically carry different data terms than consumer plans of the same product, including commitments about training on your inputs. A personal paid subscription is not the same thing as a business agreement.
The approach several people on the call already use is the right one: a paid business tool for anything touching client data, free tools for flyers, drafts, and general work. That is a sound policy.
How do I stop it from sounding canned when authenticity is what my clients pay for?
This is the most legitimate objection raised on the call, and it is worth taking seriously rather than arguing with. If your differentiator is that you are not cookie-cutter, sending cookie-cutter writing damages the actual product.
Out of the box these tools write in the average of everything they have read, and average is precisely the enemy. The fix is a one-time voice calibration you save permanently, not better prompting in the moment. Template 03 in the prompt library walks through it.
One more thing worth saying, because someone on the call made the point well: the flood of generic AI content raises the value of anything unmistakably yours. Use the tool for the parts nobody is buying from you, and put the reclaimed time into the parts they are.
Is this going to replace roles on my team?
Some tasks, yes. Whole roles, less often and more slowly than the headlines suggest, and the recruiting professionals in the room had the most grounded view on this.
The point made on the call was that as doors close others open, and this year a law firm hired someone whose entire role is AI, in an industry people assume is the last to change. Demand is rising fastest for people who can review, verify and improve AI output, which is a direct consequence of the 87% revision rate.
The honest framing for your team: the work changes shape before it disappears. People who can direct these tools and catch their mistakes become more valuable, not less. That is a training conversation, not a headcount conversation.
My team is split. Sales uses it constantly, accounting refuses. How do I handle that?
That split is close to universal, and the accounting side is not being difficult. They are correctly reading that their work carries different consequences for being wrong, and they usually sit closest to the most sensitive data.
Three things help:
- Answer the policy question first. Much of the resistance in regulated functions is not about the tool, it is that nobody has said in writing what is allowed. Publish that and a share of the objection disappears.
- Start them where accuracy is not the risk. Communication, organization, and administrative work rather than the numbers themselves. That is where several finance people on the call already use it comfortably.
- Do not force it. A skeptical team member who adopts one workflow that clearly works becomes a better advocate than an enthusiast who over-promises.
What about the environmental impact? Am I killing the polar bears?
Fair question and it got the biggest laugh of the morning, so it deserves a straight answer.
The energy cost is real and it sits mostly in data centers full of processors, which draw significant power and water for cooling. That is a genuine and actively debated issue at industry scale.
Your individual share of it is very small. A day of ordinary chat use is not where this gets decided. If it matters to you, the two levers with actual leverage are using smaller models for routine work, which uses meaningfully less compute per request, and running local models on your own hardware, which is far more efficient for simple tasks than sending them to a data center. Both of those also happen to save you money, which is a rare alignment.
How much time should I actually expect to invest before this pays off?
For a single recurring document or process, one to two focused hours of setup. That is the honest number and I would rather say it than pretend it is fifteen minutes.
The reason it is worth it is that the cost is one-time and the task is not. If a submission packet takes you forty minutes and you do six a week, two hours of setup pays back in the first week and every week after that is profit.
The reason most people do not do it is that the setup hour always has to come out of a week that is already full. That is the real barrier. It is not technical.
What is an AI agent, and is it different from what I am doing now?
Yes, meaningfully different. Using Claude, ChatGPT or Gemini in a chat window means you ask and it answers. An agent is configured to carry out work and make decisions along the way, across multiple steps, without you approving each one.
Agents need guardrails: explicit boundaries on what they can access, what they can act on, and where they must stop and ask. They also need oversight, especially early, and the same 87% revision figure applies to their output.
Do not start here. Build a manual process that works, then a semi-automated one, and only then consider handing over the steering wheel.
Let's Keep the Conversation Going
Thank you to Michelle and the North Jersey Chamber of Commerce for hosting, and to everyone who shared openly on the call. This page exists because the room made it worth writing. If something here raised a question, reach out. No pressure and no obligation.
Introductory Call
A conversation about where you are and where you want to be with AI. No preparation needed and nothing to buy. Start here if you are not sure what you need.
Book an intro callAI Clarity Meeting
A working session, not a consultation. We build something together on the call and you leave with it finished. Good for a single process you want solved.
Book one hourExtended Clarity Meeting
The same working format with room to go deeper. Better suited to document training, a multi-step workflow, or getting a team set up properly.
Book two hours