Workflow AI vs agentic AI
Workflow AI assists a human inside an existing process; agentic AI runs the multistep process end-to-end. Both sit above a deterministic core; neither substitutes for it.
- AI drafts for a person
- Or acts within policy
- Exceptions go to a person
A borrower emails to say a salary has been delayed. Workflow AI drafts a reply and suggests a seven-day extension; the servicing officer reads, edits and sends. Agentic AI checks the account history, confirms the extension sits inside the hardship policy, applies it, sends the confirmation, updates the mandate date and logs every step, and routes to a person only if the borrower has used the allowance twice already.
Workflow AI drafts a dunning letter, summarises a borrower file, suggests a hardship category, explains a calculation: the human approves and moves on. Agentic AI picks up a defaulted loan, contacts the borrower, negotiates a plan within policy, escalates exceptions, and logs every step for audit. Workflow AI is a productivity layer; agentic AI is a labour layer. They serve different hand-offs and require different governance. The reasoning behind the split is in our AI philosophy in lending.
The line between them is drawn by who is accountable for the outcome of a step. In workflow AI the person is: the model produces something, the person judges it, and the person's approval is the event on the record. In agentic AI the platform is: the agent's proposal, the core's decision and the posting are the events, and the person appears only where the policy says an exception needs a human. A lender can run both on the post-approval book. The mistake is to govern the second as if it were the first.