AI Opportunities for Small Business

A couple of hours after my own talk, How I Run My Business with AI, I was back at Launch SA for the last event of San Antonio Startup + Tech Week 2026: an informal panel and open Q&A moderated by Michael Espinoza. It was the fourth AI event of the day, and Michael's premise was simple. Put four business owners on stage and you'll see four completely different ways of using AI. He was right. What follows is the recap, organized by the questions the room actually asked.

Meet the panel

Panelist	Business	How they run it
Paul Wouters	Founder, Audit Logic	Local, on-premises AI for businesses that can't share their data, such as medical, financial and legal firms. A team of two: Paul on user experience and a developer on the technology. Previously spent ten years building a team at Shell.
Michelle Gimbel	Founder, 3E AI Solutions	Helps small companies implement AI, from lead systems to full CRMs. A solopreneur supported by virtual AI employees, after 16 years in business development for global IT firms.
Silvana Pimentel	Founder & Host, AI Your Superpower	Helps business owners use AI to win customers and scale. A solopreneur with a full staff of AI employees, two podcasts, and 20+ years in sales and business development.
Matt Nelson	Founder, Peachtree Rose Marketing	Boutique San Antonio agency focused on Answer Engine Optimization and video podcasting, with 15 to 20 recurring clients. Five humans plus a suite of virtual AI executives.

How each of us uses AI day to day

Paul doesn't plan his AI use. He reaches for the model that fits the job: Gemini for general work, Claude Code for technical and code work, ChatGPT for images, and specialized models for video. When one tool isn't working he switches, and he often uses one model to check another's work. His core business is the opposite of the cloud: keeping sensitive data on a local model so it never becomes someone else's training data. His one takeaway for the room was simple: don't upload anything you wouldn't want shared.

Michelle starts every day with a briefing from her virtual chief operating officer: priorities, overnight email and the day's first meeting. Her desk runs four screens of AI work at once: a client build, proposals being scored by an evaluator she built, and email, while n8n publishes her social posts at the click of a button. She builds on Gemini and deploys to Google Cloud, partly because its federal security assessment matters to clients in the federal space. For research she asks Claude, Gemini and ChatGPT the same question, then has Perplexity combine the best of each. She calls herself "people first AI." The goal isn't to replace anyone. It's to take the spreadsheets off people's plates so they can focus on strategy.

Silvana treats every task the same way: should I do this myself, or delegate it? More often than not, AI gets it, and she stays in the loop on strategy and on judging the result. Then she showed us instead of telling us, calling up Charlene, her bilingual ChatGPT chief of staff, live on stage to check event registrations and confirm the free resource she'd prepared for the audience.

For me, the day starts and ends with briefs from a set of AI executives. Each one sees only its own slice of the business, such as finance or marketing. I also get a synthesized feed from trusted podcasting and industry sources, so if Spotify ships a new feature tomorrow, we can be the first to publish about it for our clients. What excites me most, though, is getting rid of mundane work for my team. I once worked bilingual tech support for a cable company, and I never want anyone on my team stuck doing copy-and-paste work like that. If our editors repeat the same processing steps on every mic track, that's the first thing to automate. It frees them for the storytelling only a human can do. I build most of it with Claude Code.

What AI actually costs a small business

An audience member asked whether we use free or paid versions. Every panelist pays for at least one tool, and the advice for a beginner was consistent:

• Silvana: pick one and pay for it. Most basic plans cost $20 a month, and it will be the best $20 you spend. She now pays for ChatGPT, Claude and Gemini (Gemini comes bundled with many Google Workspace accounts) and adds specialty tools such as HeyGen and Riverside. Her AI video clone is good enough that it fooled her mom. Her rule: don't chase the tools, chase the solution.
• Michelle: exhaust the free tier first, then pay about $20 each for the deeper work. She pays for Gemini because that's where she builds.
• Paul: keep subscriptions cheap and try new tools for a month instead of a year. You'll learn what's real and what's just a good demo.

Here's our actual stack. Claude runs us roughly $200 to $300 a month including API usage. Gemini, through Google Workspace for the whole team, is about $250 to $300. I keep a ChatGPT subscription mostly to keep up with how the models compete. On top of that, we use specialty platforms such as Descript, Captions.ai and Artlist, often for realistic b-roll and on-brand imagery that doesn't look AI-generated. Our overall software spend has dropped sharply because we build so much ourselves.

Someone asked how we decide what to keep. AI tracks AI: our virtual COO and CFO report who logs in to each platform, how much we use it, and whether my build time would pay off. For scale, between my 2:00 talk and this panel we'd probably spent $10 to $15 on API calls, running agentically on everything from afternoon briefs to auto-drafted blog posts waiting for human review.

Don't trust it blindly

Paul's warning: AI makes mistakes, and pasting in a script you don't understand can cause real damage. He's had people come to him nearly in tears after locking themselves out of their own laptops. Michelle added the subtler problem: AI is built to make you happy, because validation keeps you engaged. So ask it the uncomfortable questions. What am I missing? Where are the gaps? What could go wrong with this workflow? She told the cautionary story of a man who spent a year convincing his AI, and himself, that he was Secret Service, and who ended up in prison. Paul's practical fix got a laugh and some nods: tell the AI you're asking for a friend. Its feedback gets a lot more honest.

How to keep up without drowning

• Build yourself a finger on the pulse. My suggestion is a simple automation that emails you the changes that matter in your industry. For a Texas personal injury attorney, a change to the state's minimum auto liability coverage, which hasn't moved in years, is exactly the headline you want delivered instead of hunted for.
• Ask the tool about itself. When something new ships, Silvana asks that same tool how to get more out of it, then asks it to think like a top 1% strategist and improve its own plan. If the to-do list it hands back runs past five to eight steps, she asks it to do more of the work itself.
• Batch it. Every Friday afternoon, Michelle's AI summarizes the week's changes for the platforms she builds on and flags which ones matter to her clients. That's about an hour, once a week.
• Pick a couple of sources you enjoy. Paul reads one newsletter in the morning and follows one YouTube channel, then checks a story against multiple sources before he believes it.

Vibe coding, CRMs and when to call a developer

Michael asked whether we build only internal workflows or real tools, and whether any of us are trained coders. None of us are. Michelle vibe codes about 80% of the time and brings in a junior developer for the other 20% on larger builds. My view: if it transacts or holds customer data, get a human expert to look at it. You don't need a full-time W-2 hire. Plenty of experienced software engineers and cybersecurity professionals now work as contractors. Separate high-risk from low-risk too. An internal tool that drafts social content from our brand dossiers is a very different animal from a customer relationship manager.

An audience member asked why you can't just have one AI cross-check another's code. Paul explained context limits: once a codebase runs to thousands of lines, the model can't hold it all at once, and you get spaghetti code that doesn't connect. Michelle added that vulnerabilities can hide where you'll never see them. And there's a liability angle. When a data leak happens, they won't come after the AI, they'll come after a human. Having a subject-matter expert review the work is the due diligence that limits the damage.

What counts as client data?

For Michelle, it's a spectrum. A name and email for a free workshop is lower risk. A full name, phone number, email and home address together is high risk. Silvana pointed to the rules on collecting it: clear consent language for email sign-ups, phone numbers as voluntary fields, and sensitive data like Social Security or account numbers never going into an AI tool at all. Paul treats even email addresses as personal data and keeps everything local. In marketing, I said, it depends. A podcast transcript is client data, but it's client data we're meant to publish. So have a crystal-clear internal conversation about what counts as client data and how bad it would be if each kind went public. Then wall it off, with separate systems for separate things.

Should you connect AI to your email?

Silvana's case for yes: the more context you give it, the better the results, from summarizing threads to drafting replies to suggesting agents. But do it on a paid account. Free accounts don't carry the same protections, and employees pasting company data into free tools is a leak many owners don't know they have. That's why she recommends a written AI policy. Michelle's household is split: she uses paid accounts for the business, and her husband won't touch AI at all. A separate Google account is one middle path, though as she noted, AI is clever enough to connect the dots anyway. Paul's answer was a local, air-gapped model that can still pull in outside sources through connectors. I asked him whether a local agent could anonymize email before handing it to a cloud model, for the best of both worlds. His answer: possible, but he'd advise against it, because you'd have to trust that layer never to make a mistake. I agree. Not yet.

How long does it take to build AI agents?

A three-month-old business asked how long it took us to reach virtual CFOs and COOs. Paul said you can start with something off the shelf in a day and tweak it from there. Silvana has been refining her system for three years, but today's tools let you shortcut most of that. All of us came back to garbage in, garbage out. My advice is to start with the outcome you want, not the steps to get there. Nine times out of ten, the AI will find a better path than you would. If you want a structure, use CRAFT: Context, Role, Action, Format and Target audience. Tell it what right looks like, what wrong looks like and what done looks like.

SOPs vs. prompts

Silvana's distinction: a standard operating procedure is the process, and a prompt is the specific instruction to the tool. Prompt engineering matters less than it used to, but explaining your goal still matters a lot. Michelle connected it to AI governance. Build your guardrails, such as SEC rules for a financial firm, into the SOP and the prompt, just as you would for any employee. Paul's small-business tip: practice the one-shot prompt. Think hard enough to get the result in one go, because the good models keep getting more expensive. My short answer: an SOP is human-facing, start to finish. Its AI-facing counterpart is an instruction file, usually Markdown. Same idea, but it reads differently.

Are customers asking for AI?

Paul's experience: most customers don't know yet what they can do with it. Give them a safe local setup and the first question is, "What do I type?" A week later they call back with ideas. It's a good reminder that those of us on stage work with AI every day, and most of the room is still deciding where to start.

Parting advice

• Matt: stop asking AI "what is X?" and start asking "how can I ___?" with your real destination in the blank. I want to start a podcast; how can I succeed in this niche? Then keep the conversation going.
• Silvana: personalize your tool, and move up the three levels: assistant, then mentor, then thinking partner. Above all, take action. Implement one thing from today and you'll see the difference.
• Michelle: AI is the great equalizer for education and for small business. Start now. Competitors who implement it a year before you are years ahead. Small businesses are the tugboats that pull the big ships into harbor.
• Paul: AI is not magic, and it won't make a business successful by itself. But it gives you skills that used to cost money, like programming. He built a game good enough for the App Store. Try something fun, and know what you want out of it.

Try this tomorrow

1. Pay for one AI tool and personalize it with your business context before adding any others.
2. Set up a weekly "what changed" brief for your industry and for the AI tools you rely on.
3. Write down what counts as client data in your business, and get a human expert to review anything that touches it.

Video chapters

1. 00:00Welcome from Launch SA
2. 02:30Paul Wouters, Audit Logic
3. 05:06Michelle Gimbel, 3E AI Solutions
4. 08:20Silvana Pimentel, AI Your Superpower
5. 10:50Matt Nelson, Peachtree Rose Marketing
6. 12:46How do you use AI in your business today?
7. 20:37Live demo: an AI chief of staff
8. 30:36Free vs. paid AI tools
9. 33:28What AI actually costs a small business
10. 45:49How to keep up with AI
11. 53:25Vibe coding, CRMs and when to call a developer
12. 1:02:22What counts as client data?
13. 1:07:13Should you connect AI to your email?
14. 1:14:23How long does it take to build AI agents?
15. 1:16:06SOPs vs. prompts
16. 1:21:54Are customers asking for AI?
17. 1:23:07Parting advice for business owners

Frequently asked questions

Special thanks to Michael Espinoza and Launch SA, to my fellow panelists Paul Wouters, Michelle Gimbel and Silvana Pimentel, and to everyone who stayed for the last event of the week with questions. Recorded and produced by Peachtree Rose Marketing.