The Problems We Couldn’t Justify

AI is changing the economics of automation – and the problems we can afford to solve.

By Li Pan

When we first started looking at the email problem across CIB Operations, the number was staggering: more than 100 million emails from clients a year. But the volume wasn't actually the hardest part. The challenge was the vast variation of unstructured data and requests. 

Each request carries its own context. Different lines of business have different processes, terminology, and systems. Even small distinctions matter. An abbreviation. A marking on a document. Whether a loan is "expired" or "terminated." 

Change the context, and the process changes with it. 

That variation has historically made email difficult to automate at scale. We could solve individual workflows, digitize processes, and create self-service experiences. But there was no generic solution that could absorb the sheer range of work arriving through an inbox. 

So the limit was never capability. Each workflow is solvable, but separately. Once a problem requires its own build, the question stops being whether we can automate it and becomes whether we can justify automating it. 

The most underrated thing AI does is not make hard problems solvable. It makes what's already solvable much cheaper to solve. 

Calculating the Economic Threshold 

Historically, automation had a relatively clear economic threshold. If a process happened millions of times each year, assigning engineering and data science resources to it was easy to justify. If it happened 10,000 times, or even 1,000, the pain could be very real for the team doing the work, but the use case was rarely a priority. 

Between those extremes sits a vast middle: smaller processes that are individually difficult to justify and collectively enormous. They went untouched not because they were intractable, but because each one required specialized effort, and no single one of them could carry the cost. 

Inside 100 million emails, that middle is most of the surface area. 

What changes the calculation is where the expensive part now sits. LLMs can handle a remarkable range of tasks without requiring a bespoke build for each one. But general intelligence is not the same thing as institutional knowledge. 

A model doesn't inherently understand how our firm works. 

That's central to how we approach Email AI. We're developing a common agentic framework that can support many processes while the surrounding context changes: the instructions, examples, tools, systems of record, and workflow connections that tell the agent how a process works. The specialized build once required for a single process becomes a shared capability across many. 

The technology becomes more general as the context becomes more specific. 

That shifts the bottleneck. A problem that might once have required an engineer to build a specialized solution increasingly becomes a question of operational knowledge. What does the AI need to know? Which examples represent ground truth? Which distinctions matter? What systems does it need to interact with? 

Those are not questions an engineer or a data scientist can answer alone. 

Lowering the Barrier to Building 

With AOPs, we can increasingly capture process knowledge in natural language and turn it into instructions an agent can execute. Someone doesn't need to know how to train a model to explain what should happen when an email arrives, what information matters, and what outcome the process requires. Nobody needs a data scientist to explain that an expired loan and a terminated loan require different handling. They need an operations colleague who has worked the process. 

That matters economically, not just culturally. 

Engineering and data science capacity is limited and must be prioritized across a portfolio. But the expertise that defines how a process should work already sits with the operations teams who run it every day. When those teams can become Citizen Developers and turn what they know into AOPs, they no longer have to wait for a transformation project to clear the economic threshold. 

That is what makes automating medium-volume processes affordable. 

That doesn't make the work easy. In some ways, it raises the bar for our Citizen Developers. They need to understand when more context helps AI and when it simply creates noise. They still need support from technologists and data scientists to ensure solutions connect properly to the systems around them. 

Efforts like the AOP Academy and AOP Studio aim to accelerate that upskilling and help more teams create AI agents that can automate their processes. 

Capturing What Matters 

It's tempting to say AI makes work easier. It's more accurate to say AI is changing where the hardest work sits. As the underlying technology handles a wider variety of tasks, more of the work shifts toward capturing the institutional knowledge that tells AI what actually matters in a specific environment. 

Enormous amounts of knowledge live in procedures, learning aids, policies, and the heads of people who know that one abbreviation or exception changes what happens next. AI gives us a new reason to capture that knowledge, while forcing us to decide what deserves to survive: preserving the expertise that matters without teaching AI every workaround we've accumulated along the way. 

The cost of a bespoke build is the constraint AI is lifting. And the same shift that makes medium-volume problems affordable also puts them within reach of the people closest to them, the ones who could always describe exactly what needed to happen and who now have a way to express it in a form a machine can act on. 

No single one of these processes is impressive. Together, they may be the largest opportunity in Operations. 

The question is no longer what we can automate. It's what we can finally justify.