Andrew Ng keeps citing one data point this week: 90% of enterprise AI Agent projects stall at deployment. The problem isn't that models aren't smart enough—it's "tool calling." The AI tells the program "I want to do this thing," but parameters get filled wrong, requests time out, permissions don't match, and the entire chain collapses.
What this is
Tool calling (Function Calling / Tool Calling) is the "action" foundation of AI Agents. The model outputs structured "function call intent" in conversation—it doesn't actually execute anything, just fills in parameters and names which function to call. The program then runs it.
Analogy: the AI is like a department manager who writes work tickets. It receives "check A123 order logistics," writes up the ticket (order number A123, call the "query logistics" function), and hands it to IT for execution. When results come back, the AI tells the user in natural language.
Three core steps: use JSON Schema (a universal data format spec) to "tell" the model which tools exist and how parameters should be filled; the model output passes three checkpoints—"required field validation + error retry + permission whitelist"—before anything actually executes.
Industry view
Mainstream consensus: this is the infrastructure layer of the Agent era, and engineering quality directly determines call accuracy. The front-line developer community has three iron rules—write tool descriptions clearly enough to cover "when to use," give examples in parameter descriptions, and use enums (lock down the possible options) instead of letting the model freestyle.
But there are dissenting voices. One camp argues: this "plumbing" layer will eventually be standardized and abstracted away by cloud vendors—enterprises shouldn't bother investing in researching it. Traditional IT leaders counter: their internal system APIs are messy, field naming is non-standard—this is exactly where 80% of integration costs go. The older the system, the less anyone else can substitute for this plumbing work.
Worth flagging: parameters the model fills are "untrusted by default." Hallucinated parameters—AI fabricating field values out of thin air—is a high-frequency failure, ranked by the author as Agent's #1 engineering risk.
Impact on regular people
For enterprise IT: when evaluating AI vendors, don't just chase benchmark scores. Ask "what's your tool calling failure rate on our systems, and how do you handle fallbacks." That's the real moat.
For individual careers: when your AI assistant gives irrelevant answers or pulls the wrong data, it's probably not that the AI is dumb—it's that the tool calling chain behind it broke at some link. Understanding this cuts down on misjudging the tool.
For the consumer market: next year you'll see AI customer service and sales products continue to split apart—plenty of demos will dazzle, but very few will survive peak sales events reliably. The gap sits in this "plumbing" layer.