Bounded AI agents: what we'll switch on, and what we won't

Every AI vendor pitch right now sounds broadly the same. "Agents will run your business autonomously." Cue rousing music, dashboards full of green ticks, testimonials from businesses that look nothing like yours.

We're going to draw a line publicly, so you know where we sit.

What we'll switch on

• AI agents for defined, bounded workflows — lead qualification triage, customer service email routing, invoice data extraction, quote generation, meeting summarisation.
• Agents with a clear scope — this data, this action, this checkpoint.
• Agents with a human approval step before anything client-facing or business-critical is sent.
• Agents with a kill switch you can flip.
• Agents with full logging so anyone auditing them can see what happened and why.

What we won't switch on

• Fully autonomous agents on business-critical or client-facing decisions. Not because we can't — because we shouldn't. Yet.
• Agents making financial decisions without human approval.
• Agents interacting directly with clients without a human in the loop.
• Agents changing access, permissions or configurations outside a defined change window.
• Anything a competent regulator would want to see the audit trail of, and we can't produce one.

Why the line sits where it sits

Because agentic AI amplifies whatever it touches — order or chaos, with equal efficiency. In an environment where the workflows are clean, the data is trustworthy, and the checkpoints are real, agents produce real leverage. In an environment where they aren't, agents produce fast, confident, hard-to-unwind mistakes.

The businesses winning at agentic AI right now are the ones that spent six months tidying processes before they turned an agent on. Not the ones who bought "AI transformation" from a slide deck.

The rule we hold ourselves to

If we can't tell you, in plain English, what an agent will do, when it will do it, what it will pause on, and who is accountable — we won't switch it on. Not on your environment, not on ours.

Where the line will move

Over time. As tooling matures. As regulators clarify. As we build more evidence.

We'd rather be the provider that moves the line slowly and deliberately than the one that moves it fast and then apologises to a client whose contract got auto-cancelled by an over-eager bot.