AI’s trillion-dollar opportunity (thoughts)

I really agree with this piece: https://www.linkedin.com/pulse/ais-trillion-dollar-opportunity-context-graphs-jaya-gupta-cobue by Jaya Gupta and Ashu Garg .

In fact, as I was reading it, I had a strong sense of déjà vu - in a good way.

What Jaya and Ashu are describing with decision traces, context graphs, and agents sitting in the execution path echoes a set of ideas many of us wrestled with in the BPM era… just updated for a world where software is finally capable of acting, not just recording.

Back then, BPM vendors made a provocative (and mostly correct) claim:

org charts are a terrible way to understand how work actually gets done.

Systems of record mirrored organizational structure. But customers experienced end-to-end processes that cut across sales, finance, support, legal, ops. BPM tried to elevate process as the real unit of truth - not departments, not objects. The goal was the same as it is now: support real customer outcomes, not internal abstractions.

Where BPM largely failed wasn’t the philosophy. It was the substrate.

I saw this firsthand early in my career while working on a project at PwC . We tried to reverse-engineer business process from SAP data - the theory being that if systems of record truly reflected reality, we should be able to infer how decisions were made by looking at the data exhaust. We found almost no meaningful correlation beyond the most obvious paths. The interesting decisions - the exceptions, the overrides, the “we did this because…” moments - simply weren’t there.

Which is exactly the authors’ point.

The missing layer isn’t better schemas or more governance. It’s that decision context was never treated as data in the first place.

Where I think our thinking converges even more closely is around agents as the interface - and as operators of deterministic workflows. I’ve written previously (From Salesforce Flow, to Boomi, to Slack Platform 2.0) about “agents as operators,” where agents don’t magically replace human judgment, but instead orchestrate well-defined workflows, gather context, route decisions, and act as a new interaction layer over existing systems.

Crucially: this takes time.

Human organizations rely on tribal knowledge that is accumulated slowly - through precedent, mistakes, escalation paths, and social learning. You don’t replace that overnight. You start with the boring stuff. The deterministic stuff. The workflows where humans mostly act as routers and coordinators anyway. Then you work your way up the complexity curve.

What this article does well is extend that thinking into architecture: if agents are going to operate workflows, then the act of operating them must emit durable decision artifacts. Not chain-of-thought, not opaque reasoning - but replayable, auditable traces of how context turned into action.

That said, I do have one lingering concern.

When decision context meets LLMs, there’s an implicit assumption that more context equals better decisions. In practice, my teams’ experience has been that LLMs don’t actually “understand” decisions in the way humans do. They approximate decisions via similarity - vector embeddings that are far closer to search than reasoning. Apple and others have documented this well (great summary here). Which means consistency - especially over time - is still a fundamental weakness.

Ironically, that reinforces the argument here rather than undermining it.

If LLMs struggle with consistency, then capturing structured decision traces becomes even more important, not less. Precedent has to be explicit. Exceptions have to be durable. Humans still need to be in the loop - not just to approve, but to teach the system what “this kind of situation” actually means.

So yes, systems of record will survive. But the real opportunity isn’t adding agents on top of existing data planes. It’s recognizing that decisions themselves are first-class business artifacts - and that whoever captures them, at the moment of execution, will define the next generation of enterprise platforms.

This feels less like a revolution than a long-overdue correction - one that BPM pointed at years ago, but never had the tooling to realize.

Agents might finally make it real.

Image courtesy of Welsh Labs (https://welchlabs.com)