Enterprise Adoption2 min read

Schwab Starts Its AI With Jobs That Are Easy to Undo

By , Senior AI ConsultantPublished

Schwab's AI assistant can change a client's contact details but cannot move money, and the order in which it hands AI jobs, from cheapest mistake to costliest, is a pattern any team can copy.

Nikhil Sudan, Schwab's chief strategy and innovation officer, describes the firm's approach as testing AI in a controlled setting before clients see it, and waiting when there is no clear use that works at scale. The clearest result is Charley, the assistant that began reaching eligible U.S. retail clients in October. It answers questions and makes a short list of small changes, and it does not trade.

The order of jobs follows the cost of a mistake

Each step in Schwab's AI puts a bigger error in play. A wrong summary of why a portfolio fell wastes a minute of a client's time. A wrong watchlist or contact detail is annoying, and the client can fix it. A wrong trade loses money that nobody can give back.

So the first two steps run on Schwab's own assistant, and the third does not. Trading goes through the client's own AI, and the client approves each trade, which leaves the last decision with the person whose money it is. Schwab has also made clear that it blocks outside AI agents it has not approved. That leaves two doors, and Schwab built both of them.

The doors are meant to widen slowly. Schwab plans to give Charley more to do as it grows comfortable with the quality of the model and with its own ability to govern what the AI does. Capital One also set its AI rules before it deployed agents, so this looks like how regulated firms now work, not one company's habit.

Two questions decide whether a pilot is worth running

At the same conference, executives at New York Life and CBRE Investment Management set their own tests for AI pilots. New York Life wants every pilot designed so that it could reach all of its clients, even if it starts small. CBRE wants the pilot tied to revenue or profit, and it drops one that only automates a step in a workflow without improving the process underneath.

Those tests work in a small business too. Ask whether the thing could serve everyone who does this job, and whether you can put a number of hours or money on it. A script that helps one manager with one report usually fails both. An assistant that answers every applicant's first question passes both.

The same ladder in a hospital finance team

For example, a finance team at a hospital group could start with AI that only reads. It goes through the month's supplier invoices and flags the ones that look wrong: a price that jumped, a duplicate, a bill with no purchase order. If the team handles 600 invoices a month at about two minutes each, the first pass alone takes about 20 hours of reading off a person's desk.

The second step lets AI draft the query emails to suppliers, with a person pressing send. The third lets it schedule payments, with a person approving each batch. At every step the team has watched the AI be right for a while before it is allowed to be wrong somewhere that costs money.

That is the question to ask of any AI at work: what is the worst thing it can do on its own, and can that be undone? Where the answer is a wrong summary or a wrong draft, give it the job now. Where the answer is a payment that cannot be recalled, keep a person's approval in front of it until the AI has earned the next step.

Share this

STAY INFORMED

Get AI intelligence like this delivered to your inbox.

Free forever · Unsubscribe anytime


You May Also Find Valuable