AI for the Rest of Us: What We Learned at Boston AI Week

On September 29, we hosted AI for the Rest of Us, a free three-hour virtual session for Boston AI Week. It was for owners and operators who want AI to actually work in their business, not for people who already have a data team.

I was joined by co-host Joe Khoury, who put real dollar figures on AI wins and failures from the buyer's side of the table, and Katie Cook, publisher of Salem & Haverhill Life, who showed what one person can run with a well-trained agent. (She named hers Walter Cronkite. And that's the way it is.)

Here is what we covered, and what came out of the questions.

The difference is management, not the model

MIT's 2025 GenAI Divide report found that 95% of organizations saw no measurable return from generative AI. The study has fair critics: it used a small sample and wasn't peer reviewed. But its main finding matches what we see in practice. The companies that got results built AI into specific workflows and measured outcomes. The ones that didn't bought tools and hoped.

In a small business, the owner is senior leadership. That is an advantage: you can decide on Monday and start on Tuesday, without legacy systems or a six-month approval chain.

Five questions to answer before you buy anything

• A specific, painful problem. What task costs us hours every week, and who feels it?
• One named owner. Whose name is on this, and do they use it every day?
• A number, set up front. What is the baseline, and what is the target?
• Rules before rollout. What may it read, what may it draft, and what must it never touch?
• Someone keeps building. Who improves it after the consultant leaves?

Treat AI like a new hire

Nobody hands a new hire the company checkbook on day one. They start on drafts. As they prove themselves, they get more responsibility. AI should earn trust the same way.

Most of the value for a small business today comes from the first two steps:

• Assistants: chat tools that answer questions and draft. They're here now.
• Supervised agents: they use your tools and draft for your approval. This is the sweet spot.
• Unattended agents: they run a whole workflow with no one checking. These are emerging and fragile, and they're risky for anything involving money or customers.

Anyone who sells you certainty about artificial general intelligence is selling.

The one rule

If it commits, posts, pays, or sends, a human approves it.

• Commits: contracts, the books, code going to production. AI drafts and checks; a person signs.
• Posts: social media, the website, public replies. AI writes and queues; a person clicks publish.
• Pays: invoices, payroll, refunds. AI prepares and reconciles; a person releases the money.
• Sends: any email or text to someone outside the company. AI drafts in your outbox; a person presses send.

Everything else (reading, sorting, summarizing, drafting, flagging) can run on its own. Draft-only isn't a limitation. It's the control.

The dangers that are real today

The scary risks aren't sentient robots. They're more ordinary:

• Confident and wrong: plausible answers with nothing behind them.
• Prompt injection: hidden instructions in an email or web page hijack an AI that reads it.
• Silent failure: the tool says "Done!" and nothing happened.
• Copied examples: AI repeats the examples you give it, almost word for word.
• Deepfake fraud: in 2024, an employee at the engineering firm Arup wired about $25 million after a faked video call with "executives."
• Data leakage: client data pasted into consumer tools, or agents running on the owner's login.

We shared three of our own mistakes, because they earn more trust than any client story. An agent called its work "verified" when it had only skimmed for keywords. An email tool reported success twice when the draft never saved. And our finance agent kept recommending a $500-a-month retainer that didn't exist, because it was copying a teaching example in its memory. Each one became a rule. "Verified" needs primary evidence. Check the result, not the receipt. Examples are instructions.

AI doesn't clean up messy data

It moves the mess faster. Gartner predicts that through 2026, organizations will abandon 60% of AI projects that aren't supported by AI-ready data. AI also exposes broken human processes: if the steps live in one person's head, written on sticky notes, writing them down is half the value, with or without AI.

Joe walked through the ways good projects die: a vision five to ten times the budget, a foundation that isn't ready, tools that add daily friction until people stop using them, vendor permissions you don't control, knowledge locked in one person's head, and too many priorities at once.

Your questions

A few of the questions registrants sent in, and the short answers:

"AI is only as useful as the access you give it. How do I balance that with security and hallucination risk?" Give it read access before write access, one system at a time, through its own account with the narrowest permissions it needs, never the owner's login. Broad access is safe when the output is a draft. Make every fact point to the record it came from, so a claim with no source gets flagged instead of trusted. And don't roll out "AI." Roll out one workflow, draft-only, for two weeks, then widen.

"What does AI mean for traditional, even old-fashioned, companies?" Start where the paper is: email, quotes, invoices, scheduling. AI reads PDFs and scans, so you don't have to go digital first. The knowledge in your long-timers' heads is the real asset. AI handles the back office so your people can keep doing the work customers pay for.

"How do you get enterprise-grade security and traceability at small-business size?" Give agents scoped accounts, keep keys in a vault instead of in the prompt, and log every run: what it read, what it decided, and who approved it. Enterprise features aren't a product tier. They're the shape of the system.

Your one-page AI project charter

Copy this and fill it in before your first project:

• The problem: one task, measured in hours per week
• Who feels it
• Owner: a name, someone who uses it weekly
• Baseline and target: hours, dollars, errors, or days
• Rules: what AI may read, what it may draft, what it must never touch
• Human approves: anything that commits, posts, pays, or sends
• Review cadence: weekly, 15 minutes, and who attends
• Two-week first step: draft-only
• Kill criterion: the result that means you stop

The state will help pay for training

Most businesses don't know this. Massachusetts employers with up to 100 employees in the state can be reimbursed for eligible training through the Workforce Training Fund Express program: up to $3,000 per person per course and $15,000 per company per year. Support Forge's AI courses are approved for the program, and many other states run similar training funds. See how the grants work.

What's next

The session became a habit. The Thrills, the Chills & the Pitfalls is our free monthly webinar on real-world AI implementation: one hour, one topic, live Q&A. The first one, on how to finance your team's AI training, is Tuesday, October 27.

Thank you to Boston AI Week, to everyone who spent a Tuesday morning with us, and to Joe and Katie. If you want to talk through your own charter, take the readiness assessment or reach out.