Using AI in Your Business:
A Practical Study Guide

John opens by framing his AI use as a 30-year tax attorney. He's not naturally technical — he's a talker, which turns out to be a feature, not a bug, when working with conversational AI. The real lesson is uncomfortable: the broader the question, the worse the output. AI doesn't replace tax expertise; it accelerates the work of someone who already knows what to ask. He's stopped trusting the IRS as an AI source for nuanced research and prefers Claude for its willingness to admit when it's wrong.

• 01
  Knowledge gates AI value. “You have to know the question to ask it. Sometimes you even have to know the answer.” If your domain is shallow, AI just amplifies the shallowness.
• 02
  Narrow first, then broaden. John starts every research session with very specific questions and only widens the aperture once the model has shown it understands the constraints.
• 03
  Source curation matters. He explicitly excludes the IRS as a primary source for nuanced questions and tells the model to weight authors he trusts (Bradford Tax Institute, Kitsis, etc). Output quality jumps.
• 04
  Tell it to stop being sycophantic. Put it in the base instructions. Models lean into telling you what you want to hear — the attention economy applies to AI too.
• 05
  Let it do internal memos. Sign hybrid output. John lets Claude draft long memos he'll edit, but signs anything mixed as “John Claude Von Hyre” for transparency. Authenticity gets more valuable, not less, in the AI era.

Tax returns are a worst-case data type — names, SSNs, addresses, employer info, account numbers. John refuses to feed any of it to a hosted LLM. He's building, with Ben, a tax-return redactor (the “Redactotron”) that runs completely offline on a Mac Mini. No internet connection. Documents in via thumb drive, redacted output back. The bar is paranoid by design — and that's the point.

• 01
  Air-gapped means literally no network. “There is a gap of error between it and any internet connection.” Data crosses on a sanitized USB or not at all.
• 02
  Redact before you process, even on-prem. Even if the AI itself is local, redacting first creates a defense-in-depth layer if anything ever did leak.
• 03
  Dual-output redaction for human review. Two summaries — one in original return order for page-by-page eyeball verification, one re-arranged into a working analysis layout.
• 04
  Two kinds of IP to protect. Client data is one. The other is your own proprietary methodology — which you do not want training someone else's model.

John introduces Ben, drawing a sharp line between someone who's prompted their way into AI confidence and a software engineer who's been building real systems for 20 years. The reason matters: AI-generated code accomplishes things in “the easiest and flimsiest way possible.” Production systems need redundancies, error handling, and judgment calls about when AI is the wrong tool. That's a different job.

• 01
  Vibe coding is useful but fragile. Fine for a prototype. Dangerous as the foundation of a system handling real client data.
• 02
  “Made in China” vs. designed for failure. Production code anticipates edge cases the AI never saw. There need to be redundancies.
• 03
  Background gates judgment. Knowing how systems break tells you which AI suggestions to act on and which to push back against.

Ben takes the floor and lays out what AI is genuinely good at today: pattern-matching across long documents, drafting from templates, summarization, structured extraction, and routing between specialized agents. What it's still bad at: math without scaffolding, anything where source quality determines correctness, and decisions that require contextual judgment about a specific business. The honest answer to “should I use AI for X” is often no — and that's the most valuable thing a real engineer can tell you.

• 01
  AI is great for first drafts, structured extraction, and summarization. Repetitive language work, where a human still reviews the output, is the sweet spot.
• 02
  AI is bad at arithmetic without tools. If precision matters, hand the model a calculator (tool use), don't trust raw inference.
• 03
  If your sources are wrong, AI makes the wrong answer prettier. Source curation is more important than prompt cleverness.
• 04
  The wrong tool is still the wrong tool. Some problems are better solved by a regex, a spreadsheet, or a phone call.

A practical pattern: when accuracy matters, run the same question through two different model families and reconcile differences. Models have distinct training distributions and failure modes — Claude tends to be better at admitting uncertainty, ChatGPT better at certain creative tasks, Gemini stronger at images. The verification cost is low because inference prices have collapsed. The judgment cost — knowing when the disagreement is real vs. cosmetic — stays human.

• 01
  Different models, different blind spots. Running the same prompt through Claude + ChatGPT + Gemini surfaces disagreements you'd never catch with one model.
• 02
  Cross-examine like a deposition. Ask the same question several ways. If the model contradicts itself, you've found the soft spot.
• 03
  Automate the cross-check. One AI prompts another, scores agreement, and flags the disagreements for you. Cheap insurance.
• 04
  Claude is better at admitting it's wrong. ChatGPT defended bad answers like an 8-year-old. That's a real differentiator for high-stakes work.

The most consistent piece of advice in the webinar: before you reach for a tool, audit your own process. The single best way to get value from AI is to first know exactly which 5 things you do every week that you wish someone else could handle. The conversation drives the discovery — Ben's job, in John's words, is to start with “what does your business do?” and only then think about what could be automated.

• 01
  List the friction first. Sit down for 30 minutes and write the 10 things that take time, cause errors, or get put off. That's your roadmap.
• 02
  Conversation drives discovery. A good engineer starts with your business, not their tech stack. If they pitch tools before they understand the work, leave.
• 03
  The larger goal beats the specific request. “Maybe there's a different way to get to that goal.” Don't anchor on a feature — anchor on outcomes.
• 04
  Free intake is supposed to be free. If a vendor won't talk to you for 30 minutes without a contract, you've already learned something about the relationship.

John steps back to the macro: AI's trajectory isn't unbounded. Energy and capital are real constraints. Every wave of cheap inference is partially subsidized — providers are taking losses to win market share. That won't last forever. He also gets personal: the danger isn't that AI replaces you, it's that you let it atrophy the skill that makes you valuable. Phone numbers went first. Spelling went next. Don't let analysis be third.

• 01
  Today's prices are subsidized. Plan for the world where inference costs more, not less, and price your usage with that in mind.
• 02
  Energy and capital are the moat. The companies who own GPU and energy supply will determine which AI you can use and at what price.
• 03
  Don't lose the skill of analysis. AI handles the writing. Your judgment about what to write is the part that compounds. Practice it deliberately.
• 04
  Authenticity goes up in value. When everyone's first draft is AI-shaped, the human voice becomes the differentiator.

John closes the call by openly recommending Ben to the audience and explaining how to engage. The tone is honest: there's a limit to capacity (“there's one Ben”), pricing won't be abusive but won't be a giveaway either, and the right way to start is with a conversation. He pitches Oaken AI's free intake call directly, including the phone number — which is exactly the assessment we offer to anyone reading this page.

• 01
  Start with a conversation, not a contract. The intake call is free and outcome-focused; nothing technical required from you.
• 02
  Capacity is real. “There's one Ben” — books fill quickly when client demand from referrals like this lands at once.
• 03
  Bring an outcome, not a feature. “I want to redact returns offline” works. “I want to use AI” doesn't.

• 01
  AI accelerates expertise, it doesn't replace it. Get good at narrow questions.
• 02
  Curate sources before you tune prompts. Garbage in, prettier garbage out.
• 03
  Air-gap anything sensitive. Even on-prem AI should see redacted data.
• 04
  Use multiple models to check each other. Cheap insurance against confident errors.
• 05
  Audit your process before you buy any tool. The conversation comes first.
• 06
  Don't let AI atrophy your judgment. The skill of analysis is what compounds.
• 07
  An engineering background matters. Vibe coding is fragile under load.