Goldman Sachs Is Backing AI Training. What Should You Learn?

What has actually been announced

On October 22, Goldman Sachs announced its Small Business AI Institute with Anthropic, Google, Meta and OpenAI as partners. It sits within a new $500 million commitment covering the next chapter of 10,000 Small Businesses, including education, capital access and support. That figure is not a grant offer to each business. [1] (opens in a new tab)

Planned activities include an Anthropic learning series and workshops for program alumni, Google training, Meta tools and training, and OpenAI workshops with community college partners. These are different offerings. An announcement about the institute does not mean every course is open to every owner today. [1] (opens in a new tab)

Use the institute’s official page as a starting point, then check the specific offering: who can attend, where or online, when it starts, what account is required, what is included and what might cost money afterward. The page checked for this article does not establish a complete public workshop schedule or universal eligibility. Don’t reorganize staff time around a place you have not confirmed. [2] (opens in a new tab)

There is another route to watch. The Public Library Association and Microsoft announced an AI skills collaboration on October 22, with free webinars and virtual workshops beginning in December 2026. It centers on library workers and supporting patrons; it does not establish a business class at your local branch. Ask the library what it actually offers. [3] (opens in a new tab)

The practical opportunity is to learn something useful without assuming you need to buy a custom system first. Here is one exercise an owner could prepare for whichever suitable course or workshop they choose.

Bring the inquiry that makes everyone pause

Imagine a small print shop. A customer asks for “the same flyers as last time, but a little bigger, for next Friday.” An employee searches old messages, asks about quantity and tries to establish which Friday the customer means. The estimator cannot quote until those gaps are resolved.

For this illustrative shop, the learning goal is modest: can an employee use AI to organize an inquiry into supplied details, unresolved questions and a short draft for the estimator to review? The exercise does not authorize the software to set a price, accept artwork or promise a production date.

Prepare a tiny fictional order history and three fictional inquiries. In one, all the requested details are present. In another, quantity is missing. In the third, the message says Friday but an attached note says Thursday. Write the expected questions yourself before showing the exercise to a tool.

That last step changes what you are learning. You now have a reference for judging the response. A tidy paragraph is not enough: the employee should be able to point to where each detail came from and explain which question still needs a person’s answer.

Use that exercise to choose a lesson. A course on supplying context and checking answers is relevant here. A course about generating promotional videos might be interesting, but it does not address this particular delay. If the owner cannot name the task yet, a broad introductory session can help them explore possibilities before choosing one.

The most valuable moment may be spotting a bad answer

NIST describes how generative AI can confidently produce false or inconsistent content. It also identifies privacy risks. Those are reasons to teach judgment alongside tool operation: a polished response and a familiar interface do not establish that an answer is correct or that information is appropriate to upload. [4] (opens in a new tab)

Keep the practice material entirely fictional. Recreating the shape of a difficult inquiry is enough; there is no need to paste a customer’s email, artwork, private contact information or order history into a classroom tool. Before real work begins, the owner needs to approve the tool, account and permitted information based on the actual service’s data handling.

During the lesson, ask what happens when a fact is missing, when two details disagree and when the request falls outside the shop’s instructions. Watch whether the employee notices the problem and can explain the next step. A prompt that asks the tool not to guess is useful instruction, but it is not a guarantee.

The shop might discover that the best response is simply to ask two questions. That is a worthwhile lesson: the handoff to the estimator becomes clearer because someone noticed what was missing.

A trainer can explain a feature. The shop still has to decide who owns the estimate, which information is authoritative and when the customer hears back. Bring the person who actually handles the inquiries into that conversation. Otherwise, a beautiful classroom demonstration may depend on rules nobody at work has agreed on.

What comes back to the shop on Monday?

Ask the employee to walk a colleague through a fresh fictional inquiry without the instructor taking over. Can they find the supplied facts, catch a contradiction, explain what they changed and stop before making a promise? This is a proposed practice exercise, not a validated assessment or a certificate of competence.

Keep a short shared example showing what an acceptable handoff looks like, plus an example that should be referred to the estimator. Give someone responsibility for updating it when the shop’s process changes. A saved prompt alone will not tell the next employee why a particular answer was rejected.

Then ask what the exercise revealed about the website. If almost every inquiry lacks quantity, size or a required date, a clearer request form may solve more of the problem. If messages contain useful but scattered details, staff might try assisted organization. If the same confirmed details are retyped into several systems, the next question may be integration rather than more AI training.

Those options have different costs. A form change needs thoughtful questions and testing. Assisted work needs practice, review and an approved account. Connecting tools adds permissions, mapping between fields, error handling and ongoing maintenance. Start with the smallest change that addresses the delay you actually observed.

For an initial trial, record how many inquiries arrive ready for the estimator, how many require clarification and how much staff time the complete handoff takes. Include review and corrections. Compare similar inquiries before and after the change, and keep notes on differences in complexity. A handful of successful examples is a reason to investigate, not proof of a lasting productivity gain.

The Kundan HQ project in this portfolio connects jewelry inventory, customer claims and fulfillment records. It illustrates the broader design question: what information needs to travel with a piece of work? The fictional print-shop exercise is a proposed learning approach, not an AI feature delivered in that project.

Before signing up for your next AI session, write one sentence: “After this lesson, we want to be able to…” Finish it with an observable task. Bring an ordinary example, an incomplete one and one that should make someone stop. You will have something concrete to learn—and a better basis for deciding what to improve afterward.