AI Agent (Agentic AI)
Last updated October 1, 2026
An AI agent is software that can pursue a goal with limited supervision. Give it a task, and it plans the steps, uses the tools it has been granted, such as databases, documents, email or other software, and takes actions until the task is done. It checks its own progress and adjusts when a step fails. Agentic AI is the wider term for systems built this way, sometimes with several agents working together on one job.
Most AI agents run on a large language model (LLM), the same kind of model behind AI chat tools. The difference is what surrounds the model. A chatbot answers a question. An agent can look up a loan, fill in a form, send a message or start a process.
Agents also differ from lending workflow automation. A workflow rule always does the same thing when the same event happens. An agent decides what to do based on the situation, so two similar requests can lead to different steps.
Why it matters for lenders
Agents can take on the tedious parts of lending: chasing missing documents, reading files and listing gaps, drafting borrower messages, preparing a file summary or watching a portfolio for warning signs. That is real time saved for loan officers and servicing staff.
The same power creates risk. An agent with access to borrower data and payment tools can make a costly mistake quickly. Lenders should:
- Limit each agent to the specific actions its job needs.
- Keep credit decisions, fee waivers and money movement with a named person, a practice called human in the loop.
- Log every output and action, with a time stamp and the tool's version.
- Make sure outputs can be explained, especially anything that touches a credit decision.
US guidance is still catching up. In April 2026, the Federal Reserve, the Office of the Comptroller of the Currency and the Federal Deposit Insurance Corporation replaced their model risk guidance, and the new version says generative and agentic AI are outside its scope. For now, lenders must set their own controls.
How it works in practice
A small business lender sets up an intake agent. When an application arrives, the agent reads the uploaded documents, compares them with the product's checklist and drafts a message asking the borrower for the two missing items. It also writes a short summary of the request for the loan officer.
The loan officer reviews the draft, edits one line and sends it. The agent cannot approve, decline or change loan terms. Those actions stay with people, and the agent's work is saved on the file for review.
Related terms
Agentforce
Agentforce is Salesforce's platform for building and running AI agents that answer questions and take actions inside Salesforce. Agents use a company's Salesforce data, follow instructions set by administrators and act through approved tools, such as flows and code. Launched in late 2024, it replaced Einstein Copilot. Agentforce remains the name of Salesforce's AI agent platform as of 2026.
Human-in-the-Loop
Human-in-the-loop means a person reviews and approves what an AI system prepared before it takes effect. The AI can gather data, draft a summary or suggest an action, but a named person decides and is accountable. In lending, it keeps credit decisions, adverse action notices and money movement under human control, with each review and override recorded.
AI Underwriting
AI underwriting uses artificial intelligence to help assess a loan. Most often, the AI prepares the file: it reads documents, pulls financial figures into a spread, checks for gaps and mismatches, and drafts a summary for the underwriter. A narrower use lets a model score or decide the loan, which raises fair lending and explanation questions that most lenders must answer first.
Explainable AI
Explainable AI is AI whose outputs come with reasons a person can understand and check. It shows which factors drove a score, flag or recommendation, and the explanation must truly reflect how the system reached it. In lending, explainability lets a lender give accurate reasons for a credit denial, test for unfair treatment and answer examiners and funders.
Lending Workflow Automation
Lending workflow automation uses set rules to move loan work forward without manual handoffs. When something happens, such as a document arriving or a loan being approved, the system takes the next step: it assigns a task, sends a reminder, routes an approval or updates a status. The rules are written by the lender, so the same step happens the same way every time.
Further reading
See how lenders put this into practice
FUNDINGO runs origination, servicing and reporting on one loan record inside Salesforce.