This deep dive expands on the Loop Asia conversation with Keith Carter, Founding Partner at KDA Capabilities.

Keith Carter says something in passing that most AI conversations skip: "I can also decide whether I need an agent at all. Maybe it's not an agent, it's just a database lookup and clear crisp black and white response." For a CTO, CDO, or Head of Architecture sitting on a board mandate to "do something with AI," that line is close to the whole job.

Carter's starting point is unglamorous: know your workflow, and know the data it actually needs — not a data center, a data warehouse, or a lake house. "These are technical things," he says, the kind of infrastructure a vendor sells before anyone's answered what the customer is actually going to ask. He calls the answer actionable intelligence: not all your company's data, just what answers the questions coming in.

Three tools, one decision

From there, Carter picks the tool — one of three, not one. Automation for the deterministic case: a rule fires, a lookup runs. Analytics where a human still has to interpret a pattern before deciding. An AI agent only for the genuinely ambiguous or judgment-dependent case. "We need to pick out when we use technology, whether it's automation or AI... or analytics," he says. "Let's put all three of those on the table."

That's a governance decision, and it's the one most organizations skip. Under board pressure the default is the agent — visible, fundable, demo-able. Carter inverts that: the agent is the expensive, highest-risk option, appropriate only when automation and analytics genuinely can't do the job. An order-status lookup needs a fast API call, not an LLM reasoning over it. Spending agent-grade compute on a database-lookup problem is a sign nobody did the workflow analysis first.

Why the wrong tool is what makes people angry

"Here's where people get angry about AI," Carter says, "it's not because of the technology, it's because of the way it's handled." An agent deployed where automation belonged will hallucinate, hedge, or take three exchanges to say what a lookup could say in one — and "AI" gets blamed for what was really an architecture decision. The reverse happens too: rigid automation where genuine ambiguity needed an agent's judgment leaves the customer stuck in a decision tree with no way out.

This is precisely the decision Blue Connector gets pulled into with clients carrying a board-level AI mandate and no framework for making it — vendor pitches arriving, nobody senior having mapped the workflow against the three tools before money moves. Carter's test is the first filter: what's the workflow, what data does it need, which of the three does this step deserve. Getting that sequence right is cheaper than fixing it after the wrong tool has already angered a customer.