Capital
Most bank AI projects decorate the branch. The cost sits in the packet.
Credit memos, KYC files, exception queues, and customer follow-up still move through email, shared drives, and people who know which system is lying. A chatbot on the website does not change that cost.
How I see it
Banking operations still run on packets and exceptions
Miami banking and specialty finance are full of high-volume judgment work: assembling a credit file, interpreting documents, clearing an exception, updating the system of record, and chasing the next approver. That is where headcount grows when volume grows.
The official process is usually clean. The expensive path is the unofficial one: a side spreadsheet, a missing ID, a covenant question, a customer who replied in the wrong inbox. AI is useful when it completes the packet and writes back. It is theater when it only summarizes the file a person still has to rebuild.
Controls are not optional. Anything that creates credit, compliance, or customer exposure needs a named human control point. The design error is either rubber-stamping every case or making a person redo the work the software already did.
Start with one operations workflow that already has volume and an owner. Prove a conservative improvement in cycle time or cost per case. Do not launch an enterprise banking transformation.
Common mistakes
What teams usually get wrong.
Starting with a customer-facing assistant
A public chatbot is visible. It is rarely the $540K surface area. Look at credit ops, onboarding, and exception queues first.
Leaving the core system to a human paste step
If someone still re-keys the result into the LOS or core, you automated a side task.
Treating every case as too risky for software
Some fields can be assembled without review. Some actions cannot. Name the boundary. Do not freeze the whole workflow.
A useful diagnostic
Five questions before you fund the work.
Which packet takes the most calendar time: credit, KYC, exception, or servicing?
If you cannot name it, you are not ready to fund an industry AI program.Where does the case wait, and who is waiting on a complete file?
Wait time is often larger than analysis time.What must a person still approve because it creates credit or compliance exposure?
Those steps stay human. Everything else is a candidate to move.Can you baseline cost and cycle time per completed case today?
If not, start with discovery, not a vendor bake-off.Will operations retire the side spreadsheet after a new path works?
If both paths stay, you added cost.
Economic model
Cost of a completed packetwork time + wait time + rework time, valued at loaded cost, plus error cost when the case is wrong
Compare that number to the mixed path, including the review you will still pay for.
Three credible paths
How far should you go?
Do not force one solution. Choose the path the economics, the risk, and the organization can support.
Improve the work you already have
Keep the process mostly intact and use AI on the bottlenecks that create delay, rework, or follow-up.
The workflow is already sound and a few steps create most of the friction.
Gains are usually incremental. The operating economics do not change much.
Redesign the workflow around AI
Question every handoff, queue, and duplicate step, then rebuild the process around what AI can now do.
The process grew over years and coordination now costs more than the work itself.
Requires process change, clearer ownership, and a willingness to retire old steps.
Put an agent on a high-value outcome
Give an agent responsibility for one valuable result across systems, with humans at the control points that require judgment, authority, or risk acceptance.
The workflow is high-value, variable, multi-step, and worth engineering for production.
Needs stronger architecture, evaluation, controls, and monitoring. This is not a prompt project.
When this is the wrong next step
Do not start with a sector program.
- The only acceptable project is a customer-facing demo for a board meeting.
- Risk will not name any action software may complete.
- There is no operations owner, only a digital committee.
- Volume is too low for the economics to matter.
A useful next step
Bring one expensive workflow from this industry.
We guide the implementation, go deep on the technical path, and stay hands-on through operations — or tell you when a simpler answer is better.
Discuss an AI opportunity