AI WORLD HQ
When you hear about AI "agents" that can book your flights, send emails, or move money between accounts, the natural question is: how do we trust them not to mess up? Salesforce's Kathy Baxter, in the company's own announcement, argues the answer isn't more AI — it's less, in the right places.
The problem with full autonomy
An AI agent is an AI that can take real actions in the world, not just answer questions. Most agents today are built on a probabilistic model — the same kind of large language model (think of it as the engine behind ChatGPT) that powers chatbots. "Probabilistic" means the AI makes its best guess based on patterns it has seen.
That guess is usually good. But "usually" isn't good enough when the agent is about to spend your money or email your boss. Without something checking the AI's choices, a small mistake can snowball into a big one.
The two-logic approach
Baxter's core idea, as Salesforce describes it: pair the probabilistic AI with deterministic logic.
- Probabilistic AI does the creative, flexible part. It understands what you want, writes a reply, picks a tool, and handles situations it's never seen before — because it thinks in probabilities, not rules.
- Deterministic logic does the rule-following part. Given a clear input, it always returns the same output. If an agent is about to send a payment over a set limit, the deterministic check blocks it. No guessing. No "maybe."
The AI does the thinking; the rules do the checking. Think of it like a self-driving car: the AI steers, but the brakes are still mechanical.
Why this matters for trust
Salesforce describes the result as more trustworthy, explainable, and accountable. That's the company's own framing — but the reasoning is straightforward:
- Trustworthy because the rules catch the edge cases the AI might miss.
- Explainable because you can point to the exact rule that fired (for example, "payments over a set limit require human approval").
- Accountable because every decision has a clear paper trail.
For a user, that translates into something simple: you can let an agent act, and you don't have to hover over it.
Wrap-up
The big idea from Baxter and Salesforce is simple: let the AI be creative where it should be, and let hard rules handle the decisions where guessing isn't safe. You don't need to choose between smart AI and safe AI — you just need both, doing different jobs.
A practical next step today: the next time you set up an AI tool that takes actions for you, write down one rule you wish it had to follow. That's the deterministic layer — and it's the part that turns "letting the AI run" into something you can actually trust.