A little over two years ago, I sat down with Aarish Shah to discuss Devang’s and my vision for Glyphic.

At the time, the AI conversation was still dominated by chatbots and generated marketing content. I argued that AI would be more valuable as a decision-support system: not because it always gives you the answer, but because it expands the search space. It helps teams reason over incomplete information, compare tradeoffs, and make better calls faster.

I still believe that. But decision support is one point on a longer curve:

augmentation -> bounded execution -> autonomy

The progression is less about models becoming more articulate and more about systems becoming trustworthy.

An augmented system helps a person inspect evidence or prepare a decision. A bounded agent takes a well-defined action under explicit constraints. An autonomous system plans, acts, observes what happened, and adapts toward a goal.

The demos make this progression look smooth. Production does not. Every step increases the surface area for failure. The system needs the right context, permission boundaries, tools, memory, evaluation criteria, and a way to recover when reality differs from the plan.

This is why trust and evals, not raw model capability, are usually the bottleneck.

Revenue organizations make the problem especially clear. Their highest-leverage raw material is customer-facing conversation data: calls, emails, support requests, meeting notes, and the changes buyers make or refuse to make. It contains the ground truth about pain, intent, objections, and momentum.

Most of it is unstructured, ephemeral, and lost to the organization as soon as the conversation ends.

The first unlock is to turn that stream into operational context:

  • structured signals rather than isolated transcripts
  • shared memory rather than private recollection
  • feedback into strategy and execution rather than another dashboard

Only then does useful autonomy become possible. An agent cannot pursue an outcome reliably if it does not understand what happened, why it mattered, what has already been tried, and how success will be judged.

Mustafa Suleyman’s “Modern Turing Test” captured this better than another conversational benchmark: give a system a goal, resources, and the ability to act in the real world, then evaluate whether it achieves the outcome.

The important milestone is not an AI system that can talk. It is one that can plan, act, receive feedback from reality, and iterate without quietly optimizing the wrong thing.

Autonomy will arrive workflow by workflow, not as one magical switch. Some work will remain advisory. Some will become bounded execution with human approval. A smaller set will become genuinely autonomous.

The useful question is not “Do we have agents?” It is: “For this workflow, what evidence would justify moving one step further along the curve?”