Financial work should carry itself forward

Financial software has historically been built around a simple assumption: people carry the work.

A system records a transaction. A model produces a score. A rule determines whether a condition has been met. A workflow sends something to the next person.

The system performs its function, but a person usually understands where the work stands, determines what should happen next, and keeps the process moving.

AI agents make another model possible.

Instead of designing systems only to complete individual tasks, we can begin designing financial work so that the work itself continues—across actions, changing conditions, systems, and people.

That is a larger shift than automating another step.

From doing a task to carrying the work

Most financial automation is task-oriented.

Extract information from a document. Calculate a ratio. Detect an anomaly. Apply a rule. Generate a summary. Route an application.

Each task can be valuable. But completing a task is different from carrying responsibility for what happens next.

An agent changes that relationship when it can pursue an objective rather than wait for every individual instruction.

It can understand the current situation, determine the next appropriate action, use the tools available to it, observe what happens, and continue until the work is complete or something requires another actor.

The important change is not simply that AI performs more tasks.

The unit of automation begins to move from the task toward the work itself.

Financial work changes while it is happening

Financial processes rarely unfold against a frozen set of facts.

Balances change. Transactions arrive. Documents become outdated. A customer provides new information. A policy condition is triggered. Another participant acts. An approval is granted or withheld.

The right next action can change because the situation changed.

Traditional workflows handle this by anticipating paths in advance: if this happens, do that. When the number of possible conditions grows, the workflow grows with it.

Agentic systems introduce another possibility.

The system can evaluate where the work stands now and determine what should happen next within the boundaries it has been given.

That doesn't eliminate rules. It doesn't eliminate deterministic systems. And it doesn't mean an agent should be free to improvise wherever it wants.

It means the path through the work no longer has to be completely specified before the work begins.

The actor can change without losing the work

Once work can continue over time, another problem appears.

The same actor will not always be responsible for the next step.

An agent may carry the work until a deterministic check is required. A policy may require human approval. An external institution may need to respond. A person may resolve an exception and return the work to the system.

The important thing is that changing the actor should not mean starting over.

The next actor needs enough context to understand where the work stands, what has already happened, what remains unresolved, and what authority it has.

That makes continuity a property of the system rather than a property of whichever person or agent happens to be acting at the moment.

The process belongs to the infrastructure. The task belongs to the actor.

The actor can change.

The work should remain coherent.

When the system carries the work, infrastructure has a larger job

A system that produces an answer can be relatively narrow.

A system expected to carry work forward needs more.

It needs current information rather than only historical inputs. It needs to know where information came from and whether it is still valid. It needs durable state so work can pause and resume. It needs permissions that determine which actions are allowed. It needs tools through which authorized actions can actually happen.

It also needs a record of what occurred.

That becomes especially important in financial environments, where the ability to explain an action can matter as much as the ability to perform it.

The intelligence of the model is only one part of the system.

The surrounding infrastructure determines whether that intelligence can participate in real work reliably.

Humans don't disappear. Their place becomes explicit.

Designing work for agents does not require designing people out of it.

Some decisions belong to people because the relevant authority belongs to them. Some situations require judgment that the system should not make. Some exceptions carry enough consequence that intervention is appropriate. Some work depends on relationships, negotiation, or institutional accountability.

The difference is that human participation can become intentional rather than assumed.

Instead of beginning with a process carried by people and asking which steps AI can automate, we can begin with the work itself.

What must happen?

What information is required?

What can be determined reliably?

What authority is necessary?

Which actor should perform each part?

Sometimes that actor will be an agent. Sometimes it will be deterministic software. Sometimes it will be a person.

The important thing is that the work can continue across them.

The architecture is arriving before full autonomy

Today's agents are not reliable enough to own every financial process from beginning to end.

Long-running tasks remain difficult. Errors can compound. Context can become stale. Tools can fail. An agent can misinterpret a situation or conclude that work is finished when it is not.

Those limitations matter.

But they do not require waiting for perfect autonomy before designing systems differently.

The architectural requirements appear earlier.

Durable state matters before an agent can operate indefinitely. Permissions matter before an agent can take high-consequence actions. Provenance matters before every decision can be trusted to a model. Human intervention needs a defined place before humans become exceptions rather than default operators.

The infrastructure for agentic financial work can mature while the boundary of agent responsibility expands gradually.

Design for the work, not only the task

The first generation of AI in financial systems has largely been about improving individual activities: extracting information, generating analysis, assisting decisions, answering questions.

The next design problem is broader.

How should the work itself operate when the actor performing the next task might be an agent, a deterministic system, a person, or another institution?

That question changes what the surrounding infrastructure has to provide.

Financial work does not become AI-native simply because AI performs more of its tasks.

The deeper change happens when the surrounding system is designed so that context, state, authority, action, and continuity allow the work itself to keep moving.

Humans can remain wherever their judgment or authority is required. Deterministic systems can remain wherever certainty is preferable. Agents can perform the work they are capable and authorized to perform.

But the thread connecting those actors no longer has to live primarily in someone's head.

The work can carry itself forward.