Eight ideas for using AI on purpose, in your shop, under your rules. Captions are built into the picture. The chapters, the full transcript, and the AI voice disclosure are on the video gallery.
The Owner's Clinic • Debrief
Watch first
Topics
Keep these handy
Analog plan B
AI sovereignty
Four commitments, forged in
01
02
03
04
Honorable mention
Honorable mention
Follow-up survey
Downloads
Next steps
Owner's Clinic debrief
Your Owner's Clinic debrief from Liberty Forge AI: the clinic video, a short summary and next step for each topic, the follow-up survey, and the take-home PDFs.
The clinic video, in about four and a half minutes.
Pick a topic. Get the short version and a next step.
Click a topic card for a short summary and a next step. Use “Expand for more” when you want the details and sources.
Your AI constitution
From the Owner's Clinic follow-up packet
An AI constitution lists the only jobs your AI or agents may do. Anything not on the list stays with a person. To give the AI a new job, you add it to the list on purpose, like an amendment, and a short never-list sits on top as a backstop. This is an analogy for writing rules for a tool you hired, not legal advice.
Next stepThe matching handout, “Enumerated powers: how to write an AI constitution,” with the companion “Built in our image,” comes in your follow-up packet. Fill out the survey to pick a follow-up date (weeks of Nov 30 and Dec 7), or call (605) 954-4834 or email zachary@liberty-forge.ai.
Be specific: “draft a customer email for me to review” is a job, and “help with the business” is not. Start the never-list with: do not move money; do not put client payment details, account numbers, or card numbers into AI; do not send, post, or file anything without a person checking it; do not invent sources, prices, or legal conclusions. Name the human in the loop, and amend in writing before a new use creeps in through a chat.
A companion handout, “Built in our image,” explains why this enumerated-powers approach fits when AI picks up human habits such as wanting to please, sounding sure when it is not, and filling any gap you leave it.
Source. Prairie Liberty Forge AI, “Enumerated powers: how to write an AI constitution,” updated November 16, 2026 (America/Chicago). See also Forge Paper IV, Write the rules before you need them, and the follow-up handout “Built in our image: why AI can fail when it carries human tendencies.”
AI sovereignty
Official four from liberty-forge.ai; honorable mentions from the clinic take-home
Your constitution says what AI may do; sovereignty is who owns and controls it. Liberty Forge's official floor is four commitments, as published on liberty-forge.ai: data sovereignty, zero blue-collar layoffs, human-in-the-lead, and in-house, client-specific models. Individual consent and Seen and unseen (Bastiat's lens) stay as honorable mentions from the clinic take-home. See the sovereignty card and the commitments below.
Next stepAsk Zach about a Context Audit or a Forge Assessment. The pre-assessment self-check (PDF) and the one-page process printout (PDF) are ready to download. Fill out the survey to pick a follow-up date (weeks of Nov 30 and Dec 7), or call (605) 954-4834 or email zachary@liberty-forge.ai.
On the Liberty Forge site, sovereignty is a place: records sit on a physical machine, and the question is whether that machine is yours. The firm sets up open-weight models on hardware you own, with local storage and keys you hold, and can also connect hosted models through accounts you own when you decide what may leave. The Local Forge, a productized local buildout with Watertown-area IT partners, is described on the site as in progress.
Three questions from Forge Paper III, to ask of every tool you already use: Where does this actually live, not “the cloud,” but which company, building, and country? What are they allowed to do with it? If they vanished tomorrow, what would you still have? If the answer is nothing, you do not own it. You are renting your own records.
- Private on-premises AI for small business
- Your Data Is Property (Forge Paper III)
- Local LLM vs ChatGPT for sensitive data
- AI integrations: Claude and the toolbox
- Family farm AI without giving up data
Source. Official four as published on liberty-forge.ai and liberty-forge.ai/private-on-prem-ai, read November 16, 2026. Honorable mentions quoted as published in the Owner's Clinic take-home (Handout A) and at liberty-forge.ai/heartland.html.
AI tools comparison
From the Owner's Clinic follow-up packet
Teach the setting, not the slogan: business plans usually do not train on your chats. Personal plans often do, unless you turn that off. The comparison sheet covers ChatGPT, Claude, Gemini, Microsoft Copilot, Grok, and Perplexity, plus voice and video tools. Plans change, so confirm the live site before you buy.
Next stepThe matching handout, “AI tools comparison,” comes in your follow-up packet. Fill out the survey to pick a follow-up date (weeks of Nov 30 and Dec 7), or call (605) 954-4834 or email zachary@liberty-forge.ai.
| Tool | Best for | Training, in brief |
|---|---|---|
| ChatGPT | Everyday writing, brainstorming, images, and a wide tool set in one chat. | Personal plans: training on by default. Turn off Settings, then Data controls, then Improve the model for everyone. Business, Enterprise, Edu, and Healthcare: not used for training by default. |
| Claude | Careful writing, long documents, and work you want to brief like a new hire. | Personal plans: opt-out. Turn off Settings, then Privacy, then Help Improve our AI models. Team, Enterprise, and API: no training by default. |
| Gemini | Work already living in Gmail, Docs, Drive, and Search, plus image and video features. | Keep Activity is typically on for adults. Review or turn it off in Gemini Settings, then Activity. |
| Microsoft Copilot | Work already living in Word, Excel, PowerPoint, Outlook, Teams, and SharePoint. | Microsoft says prompts, responses, and file contents are not used to train foundation models, for Copilot in Microsoft 365 apps for home and for work or school accounts. |
| Grok | A second opinion, current X/web lookup, plus image and video in the same subscription. | Logged-in users can turn training off: Settings, then Data Controls (app) or Data (website), then Improve the model. |
| Perplexity | Questions that should come back with sources, especially current-web research. | Consumer training and retention settings could not be re-read on the official help page in the November 15, 2026 check (unverified). Confirm on the live site. |
The full sheet also covers voice and meeting tools (including ElevenLabs, ChatGPT Voice, Wispr Flow, and Granola) and video tools. Sora (OpenAI) is no longer a current product.
Source. Prairie Liberty Forge AI, “AI tools comparison,” plan names and privacy settings checked November 15, 2026 (America/Chicago) on official vendor pages, including Data controls in ChatGPT, Claude model improvement settings, and the Gemini Apps Privacy Hub.
Voice cloning safety
From the Owner's Clinic follow-up packet
A short, clean recording of your voice is now enough for someone to fake a call that sounds like you. Treat your voice the way you already treat a signature or a PIN, and keep a physical safe phrase agreed on in person. A good clone takes a few clean minutes, and a near-perfect clone takes half an hour to a few hours, but a scammer does not need a perfect clone to fool someone on a bad phone line. If the caller will not use the phrase, hang up and call back on a number you already have.
Next stepThe matching handout, “Protect your voice from cloning,” plus the longer voice cloning guide, comes in your follow-up packet. Fill out the survey to pick a follow-up date (weeks of Nov 30 and Dec 7), or call (605) 954-4834 or email zachary@liberty-forge.ai.
The longer voice guide, which goes with the one-page safety tip, follows ElevenLabs' documentation for the tool shown at the clinic. Other tools differ, so check theirs. An instant clone uses about 1 to 2 minutes of clean audio. A professional, high fidelity clone asks for at least 30 minutes, with best results near 1 to 3 hours, and requires voice verification of your own voice only.
Quality beats length. One clean minute beats ten noisy ones. If you cannot get high-fidelity audio, gather the cleanest existing recordings, isolate the single speaker, and set expectations: an instant clone from old audio may sound close but not right. To defend your own voice, set a safe phrase in person (one inside the family, a different one for close friends and key business contacts), hang up and call back, slow down on pressure, and add a second channel at work before anyone moves money or changes bank details.
Source. Prairie Liberty Forge AI, “Protect your voice from cloning,” and “Voice cloning: make your own on purpose, and defend against the fake,” sources checked November 16, 2026, including ElevenLabs voice-cloning docs and FBI IC3 Alert I-120324-PSA, December 3, 2024: ic3.gov/PSA/2024/PSA241203.
Bad data and bad users
From the Owner's Clinic follow-up packet
AI makes your process faster, including your mistakes. Bad inputs come out as confident, well-written, wrong outputs, and whatever a person can reach, an AI working for that person can reach too. Bad data can be stale, duplicate, wrong, missing, overshared, or pasted into the wrong tool. Most “bad users” are busy people, so stress-test the process and name who checks before you roll it out.
Next stepThe matching handout, “What could go wrong? Bad data, bad users, and who checks,” comes in your follow-up packet. Fill out the survey to pick a follow-up date (weeks of Nov 30 and Dec 7), or call (605) 954-4834 or email zachary@liberty-forge.ai.
Watch for the well-meaning shortcut taker, the overtruster, the insider who goes looking, the outsider hiding in a document (prompt injection), and the over-powered agent given access “just in case.” Ask the tool to list ramifications, assume some of the source data is wrong, and assume both a careless user and a dishonest user will touch it. Name a person who checks totals, names and addresses, sources, file versions, access, and the AI's list of allowed jobs. “Somebody” is not a name.
Source. Prairie Liberty Forge AI, “What could go wrong? Bad data, bad users, and who checks.” Prompt injection: OWASP Top 10 for LLM Applications, LLM01. Copilot surfaces only what each user can already view: Microsoft Copilot privacy. AI tuned on human ratings tends to agree with the user: Sharma et al., arXiv:2310.13548.
Microsoft: Automate, SharePoint, Copilot
From the Owner's Clinic follow-up packet
Power Automate moves the work. To Do, Planner, and Lists hold the work. SharePoint permissions decide who can open it, and Copilot can see whatever each person can already open. Fix permissions before you turn Copilot loose, map the work before you automate it, and start with a Context Audit if you have shared files, several users, and client data.
Next stepThe matching handout, “Power Automate, SharePoint, and Copilot: who does the work, and who can see what,” comes in your follow-up packet. Ask Zach about a Context Audit or a Forge Assessment; the pre-assessment self-check (PDF) and the one-page process printout (PDF) are ready to download. Fill out the survey to pick a follow-up date (weeks of Nov 30 and Dec 7), or call (605) 954-4834 or email zachary@liberty-forge.ai.
A flow is a trigger plus steps. It does not hold your checklist. You check items off in To Do, Planner, or a list, and a flow can react when you do. Write the process out by hand first, including what happens when the internet or power is out, then simplify, then automate. SharePoint permissions flow downhill from the site. Add people to groups, prefer Specific people or People with existing access on sensitive sites, and keep HR, payroll, finance, and client files in their own place with a small group.
Microsoft says its work Copilot only surfaces organizational data to which individual users have at least view permissions. It does not break your permissions. It uses them, quickly, so oversharing that was hidden by clutter becomes visible. After a Context Audit or Forge Assessment, the next pieces of work are an automation audit and a Copilot permissions review, then a hands-on build.
Source. Prairie Liberty Forge AI, “Power Automate, SharePoint, and Copilot: who does the work, and who can see what.” Microsoft sources, checked November 16, 2026, including Copilot data and permissions, permission levels and inheritance, and sharing link types.
Analog backups, plan B
From the Owner's Clinic follow-up packet
AI can make the front office and your automations faster, but the work still has to move when the internet, the power, or a cloud tool is down. For every automated workflow, keep a one-page paper fallback: who decides, who does the work, the form to use, and how you catch up later. Print a phone tree and the lists, schedules, and customer contacts you actually need, and keep them current. Lock the kit, label it, keep one copy offsite, and run a short outage drill once a quarter.
Next stepDownload the handout, “Analog plan B: a dead-simple human process for when systems go down” (PDF), and see the plan B card below. Fill out the survey to pick a follow-up date (weeks of Nov 30 and Dec 7), or call (605) 954-4834 or email zachary@liberty-forge.ai.
If you cannot do the job with a printed page, a phone, a pen, and a named person, you do not have a plan B. You have a hope. Keep the one-page fallback where the work happens, not only in a folder on a computer. Name a person and a backup. “Somebody” is not a name.
Date the printed lists, and write the next refresh date on the packet. A monthly review is enough for most shops. A weekly one is better for lists and schedules that change often. Client payment details, account numbers, and card numbers stay out of AI tools, and they stay off loose printouts too. If a number must be on paper, lock it up.
During an outage, write every sale, order, or promise on paper, give the customer a paper receipt, and drop your copy in a tray labeled “Not entered.” When systems return, one named person re-enters those items, checks totals against the paper, and moves each sheet to “Entered.” Do not skip the check.
Once a quarter, run one real workflow as if the usual software is down, even for 20 minutes. Fix whatever was missing or stale, and set the next drill date on the packet. Lock the closet or drawer, label every binder or drive, and keep one current copy offsite so a fire or a break-in does not take the only packet.
Source. Prairie Liberty Forge AI, “Analog plan B: a dead-simple human process for when systems go down,” November 16, 2026 (America/Chicago). This is practical guidance for a shop or office, not legal, insurance, or cybersecurity advice.
Make AI not a yes-man
From the Owner's Clinic follow-up packet
You don't want it to agree with you. You want it to test your ideas. Put stress-test rules in your master instructions, which apply to every chat, and put learning methods like Socratic questioning or the Feynman technique in a project, not in those standing instructions. The handout shows where to find training settings and master instructions in ChatGPT, Claude, Gemini, Microsoft Copilot, and Grok.
Next stepThe matching handout, “Make AI not a yes-man,” comes in your follow-up packet. Fill out the survey to pick a follow-up date (weeks of Nov 30 and Dec 7), or call (605) 954-4834 or email zachary@liberty-forge.ai.
If this is a work or school account, your organization's admin controls training and data retention, so ask IT first. On a personal account, turn off training on your chats. Off does not mean private. Then paste the stress-test prompt into master instructions. Test it: state an opinion, then say you were joking. If it flips, tighten the instructions. The full wording for Prompt A (“Be my sparring partner, not my cheerleader”), Prompt B (Socratic tutor), and Prompt C (the Feynman technique) is in the handout. Always: no client payment details, account numbers, or card numbers, in any tool, with any setting.
Source. Prairie Liberty Forge AI, “Make AI not a yes-man,” sources checked November 16, 2026, including OpenAI, Anthropic, Google, Microsoft, and the xAI Consumer FAQ.
ChatGPT ads and training
From the Owner's Clinic follow-up packet
Training, ad targeting, and retention are three different switches. Turning off “Improve the model for everyone” controls whether your chats help train OpenAI's models; it does not turn off ads. On the Free and Go plans, ads can still be picked from the current chat, and if ad personalization is on, from past chats and memory too. OpenAI says advertisers never see your chats, and the Plus and Pro plans, plus Business, Enterprise, and Edu, have no ads.
Next stepThe matching handout, “ChatGPT ads: training, ad targeting, and retention are three different switches,” comes in your follow-up packet. Fill out the survey to pick a follow-up date (weeks of Nov 30 and Dec 7), or call (605) 954-4834 or email zachary@liberty-forge.ai.
Turning one switch off does not turn the others off. “Not shared with advertisers” is not the same as “not used for ads.” For business use, use a business plan. On a personal Free or Go account, turn off training and ad personalization, and review memory now and then. Use Temporary Chat for one-off sensitive questions. Recheck every few months, because ads are new and the pages say availability may evolve.
Source. Prairie Liberty Forge AI, “ChatGPT ads: training, ad targeting, and retention are three different switches.” OpenAI pages read November 16, 2026: Ads in ChatGPT, Our approach to advertising, and the Privacy Center. This describes OpenAI's published statements. It is not legal advice.
Plan B on paper. Records you own.
If the internet, the power, or a cloud tool goes down
Every automated system needs a dead-simple human process. If you cannot do the job with a printed page, a phone, a pen, and a named person, you do not have a plan B. You have a hope.
- One-page fallback per workflow: who decides, who does the work, the form to use, how you catch up.
- Print and date the phone tree, the schedule, and the customer list you actually need.
- Write every sale or promise on paper; a named person re-enters it later and checks totals.
- Lock the kit, label it, keep one current copy offsite.
- Run a 20-minute outage drill once a quarter.
Who owns and controls it?
Your constitution says what AI may do. Sovereignty is who owns and controls it. Ask three questions of every tool you already use:
- Where does this actually live: which company, building, and country?
- What are they allowed to do with it?
- If they vanished tomorrow, what would you still have?
If the answer is nothing, you do not own it. You are renting your own records. Read more in Your Data Is Property (Forge Paper III) and Private on-premises AI.
The floor under every engagement. Plus two honorable mentions.
The four official commitments, as published on liberty-forge.ai.
Data sovereignty
On-premises, owner-controlled systems. Your operational data never leaves your custody.
Zero blue-collar layoffs
In the engagement terms, in writing. AI goes after office overhead, never the production floor.
Human-in-the-lead
The AI drafts, the owner decides, and the important stuff gets checked.
In-house, client-specific models
Models tuned on your operations, running on your hardware.
Honorable mentions, quoted as published in the clinic take-home.
Individual consent
Every customer, every record, informed and opted in. No surveillance dressed as service.
Seen and unseen
Bastiat's lens applied to every recommendation. The cost no vendor shows you, we name out loud.
Tell us how it went. Pick a follow-up week.
A few minutes. Questions 1, 2, 3, 4, 5, and 8 are required. Your answers go straight to Zach.
Take-home PDFs. Print them, share them in your shop.
- Clinic take-home (Handout A)The forum short version, the four commitments and honorable mentions, and a plain-English AI glossary. PDF, 3 pages, 0.5 MB.
- Pre-assessment self-checkA short questionnaire on data, tools, people, permissions, and goals. PDF, 2 pages, 0.3 MB.
- What the Liberty Forge assessment process looks likeA one-page printout of the steps below. PDF, 1 page, 0.3 MB.
- Analog plan B: a dead-simple human process for when systems go downPaper fallbacks, outage drills, and physical security for backups. PDF, 2 pages, 0.4 MB.
- Self-check. You fill out the short self-check: data, tools, people, permissions, and goals.
- Discovery visit. We sit with you and your people, walk the operation, and map where hours actually go.
- Findings. We write down what we saw: where the work is, where the data lives, and where the risk is.
- Written recommendations. You get a plan you own, including what to do, what to leave alone, and what a later build would look like if you want one.
- You decide whether to build. Nothing is built until you say so. The recommendations stay yours whether you hire us or hand them to your own people.
A person stays in the loop at every step: the AI drafts, the human decides. To ask about a Context Audit or a Forge Assessment, use the survey or call (605) 954-4834 or email zachary@liberty-forge.ai.
What happens now
You will get two follow-up emails. One has your tailored next steps. The other is a content and citations pack, so the materials above are easier to find in one place. You can also fill out the survey and pick dates in the weeks of Nov 30 and Dec 7. More to watch: the Heartland AI Ethics Forum livestream and Tyler Fischella's “Don't Take the Gold Coins” are on the video gallery.