Where to Start With AI Without Handing Over the Keys: November Session Recap
On November 19, we ran the second Thrills, the Chills & the Pitfalls session, our free monthly hour on putting AI to work. The topic came straight from the questions we hear most: where do we start, what do we buy, and how do we do it without giving a tool the keys to everything?
Thank you to everyone who registered and to everyone who shared it. If you couldn't make it, here are the highlights.
1. What AI gets to see
The same vendor can follow two different rulebooks. The split that matters is usually personal account versus business account, not which brand you pick.
- Consumer Gemini trains on chat content by default, and a subset of chats is read by human reviewers. Reviewed chats are kept for up to three years, even after you delete them. Gemini in Google Workspace doesn't train on your data without permission and isn't reviewed by humans outside your organization.
- Claude consumer accounts ask each user to choose whether chats are used for training. Business plans and the API aren't used for training by default.
- Microsoft 365 Copilot on a work account is covered by Microsoft's data protection terms, but web searches it sends to Bing fall outside those terms and outside its HIPAA agreement. That's the kind of detail that only shows up in the fine print.
Meanwhile, 80% of AI users at small and mid-sized companies bring their own AI tools to work (Microsoft and LinkedIn Work Trend Index, 2024). If you haven't picked a tool for your team, your team has probably picked one for you, on personal accounts.
If you're a tax, accounting or lending firm, this isn't optional. Under the FTC Safeguards Rule, tax preparers count as financial institutions, and you're expected to choose service providers that can protect customer data and to spell out your security expectations in the contract.
Questions to ask before the first upload
1. Is our data used for training by default? Show me the contract clause, not the FAQ. 2. How long do you keep prompts, files and logs? Is there a zero-retention option? 3. Will you sign a data processing agreement? A BAA if we handle health data, and which features does it cover? 4. Do humans ever review our content? When, and who? 5. Which connected features, like web search, fall outside those agreements? 6. Can we see a SOC 2 Type II report or ISO certification? 7. What happens if you get a subpoena for our data, and will you tell us? 8. Does it respect our existing permissions and audit logs?
Answers that should end the conversation: "we may use your data to improve our models" with no opt-out in the contract, no data processing agreement, retention "as long as needed," no independent audit report, or staff having to sign in with personal accounts.
2. What AI gets to do
Give AI access in layers, the same way you'd onboard a new hire:
1. Read-only. It can look things up and summarize, but it can't change anything. 2. Draft-and-approve. It writes the email, the reply, the update. A person clicks send. 3. Acting on its own. Only for narrow, low-stakes jobs that have earned it.
This isn't just our opinion. OWASP, the security community's standard reference, names "excessive functionality, permissions, and autonomy" as the root causes of one of the top risks for AI applications. Its fix is the same: limit permissions to the minimum necessary, and require a human to approve high-impact actions.
It also explains why an assistant that reads your inbox shouldn't hold the master key. Text hidden in an email, web page or file can carry instructions the AI may follow. OWASP calls this indirect prompt injection.
The cautionary tale: in July 2025, during an explicit code freeze, an AI coding agent deleted a company's live database. The fix wasn't a better prompt. It was separating the agent's access from production.
3. Who checks the work
You own what your AI says. When Air Canada's website chatbot invented a refund policy, a Canadian tribunal held the airline responsible for it: "It should be obvious to Air Canada that it is responsible for all the information on its website."
This isn't rare. One researcher's public database now tracks more than 2,300 court cases involving AI-generated hallucinations. Even legal research tools built on real case law got answers wrong 17-33% of the time in a Stanford study. Grounding the AI in your own documents helps, but it doesn't replace checking.
The habit we recommend: click every citation, and check every number, before it goes to a customer.
4. Where to start
You're not behind. About 1 in 5 U.S. businesses use AI to produce goods or services (U.S. Census Bureau, May 2026). Among those that do, sales and marketing is the most common use.
Three first projects we recommend, because all of them stay at layer one or two:
1. Marketing and sales drafts. Posts, emails and proposals that a person edits and approves. 2. Routine customer questions. AI drafts the reply, and a person approves it before it goes out. 3. Search over your own documents. Read-only access to your policies, procedures and past work.
Pick one, give it to one named person, and decide what success looks like before you buy anything.
Getting the training paid for
In Massachusetts, the Workforce Training Fund Express program covers up to 100% of an approved course for companies with 100 or fewer Massachusetts employees, up to $3,000 per person per course and $15,000 per company per year. Three details worth knowing:
- Reimbursement is capped at $300 per hour of instruction, so a $3,000 course is fully covered only if it runs at least 10 hours.
- Express covers tuition, not the wages you pay people while they train.
- You pay first and get reimbursed after the training, so plan for the cash.
For the full breakdown, see our October session recap or how the grants work.
Take the checklist with you
The vendor questions are also on a one-page AI vendor checklist you can print and bring to your next demo. If you'd like help working out where your business should start, book 30 minutes.
Next session
December's session will be announced at support-forge.com/events/monthly. What should we cover? Reply to any of our emails with the AI question your team is stuck on.
*Sources, checked November 2026: Google Gemini and Workspace privacy documentation, Anthropic consumer terms and privacy center, Microsoft enterprise data protection documentation, Microsoft and LinkedIn Work Trend Index 2024, FTC Safeguards Rule guidance, OWASP Top 10 for LLM Applications 2025 (LLM01, LLM06), AI Incident Database, Moffatt v. Air Canada (2024 BCCRT 149), Damien Charlotin's AI hallucination cases database, Stanford RegLab/HAI (arXiv 2405.20362), U.S. Census Bureau Business Trends and Outlook Survey, and Commonwealth Corporation Express program guidelines.*
Related Articles
How Much Does AI Consulting Cost? What Actually Drives the Price
Typical AI consulting costs in 2026, from hourly rates to a first project, what moves the number up or down, and the costs left off the quote.
How to Choose an AI Consultant: 14 Questions and the Red Flags
The questions that separate an AI consultant who will leave you with something that works from one who will leave you with a slide deck, plus the red flags worth walking away over.
How to Finance Your Team's AI Training: October Session Recap
The short version of our first monthly session: who the Massachusetts Workforce Training Fund pays, how much, the timing rules that trip people up, and a few facts most employers have never heard.
Subscribe to The Forge
Practical takes on AI, automation, and IT for small and mid-sized businesses. New posts, delivered when they're ready.
Double opt-in — we'll email you a confirmation link first. We only use your address to send The Forge. Unsubscribe any time.