Agentic AI is software that is given a goal, then plans steps, uses tools and acts on its own until the goal is met or it gets stuck. A chatbot answers a question; an agent does a job. In finance and audit that is useful for routine, well-defined work, and risky wherever a wrong action cannot be undone, so the controls matter more than the model.
Last reviewed: 17 October 2026. Facts checked. We searched for how Wolters Kluwer, KPMG and Deloitte describe agentic AI. The Wolters Kluwer article was read through search results only, and the KPMG and Deloitte descriptions reached us second-hand (a Wolters Kluwer summary and a vendor blog), so we do not quote either firm. There is no single agreed definition. We did not open any regulator or ICAI page on agents.
What "agentic" means in plain words
Most people know AI as a chat box. You ask, it replies, you decide. An agent adds three things.
- A goal. "Reconcile this bank statement to the ledger" rather than "what is a bank reconciliation?"
- Tools. It can open files, query a database, run a calculation or send a request to another system.
- A loop. It plans, acts, looks at the result and tries again.
Wolters Kluwer describes agentic AI as systems that perceive context, reason and act, and contrasts them with rule-based automation that repeats fixed steps. Definitions vary between vendors, so treat any one definition as a working description.
Chatbot, automation script or agent?
| Chatbot | Rule-based automation | Agent | |
|---|---|---|---|
| Starts from | A question | A fixed trigger | A goal |
| Steps | One reply | Fixed, written in advance | Chosen as it goes |
| Can it act? | No | Yes, in set ways | Yes, with tools |
| Predictable? | Mostly | Highly | Less so |
| Main risk | Wrong answer | Wrong rule | Wrong action |
The last row is the point. A chatbot's mistake is a bad paragraph you can read and reject. An agent's mistake may already be a posted entry or a sent email.
What agents can do in finance and audit
Tasks that suit an agent have a clear goal, structured inputs and a result a person can check.
- Match bank lines to ledger lines and list the unmatched ones.
- Pull invoices from a folder, extract fields and flag mismatches.
- Draft a working paper from tested data for a reviewer to edit.
- Chase missing documents by preparing a request list.
- Run a set of agreed analytics and assemble the exceptions.
Our companion piece on what AI agents can and cannot do in internal audit goes through audit procedures one by one. This post focuses on the control side.
What agents do badly
- Judgement. Whether a provision is reasonable, or whether management is credible, stays with a person.
- Long chains. Each step can add a small error, and later steps build on it.
- Made-up detail. Language models can state wrong things confidently. See AI hallucinations in audit.
- Hidden steps. If you cannot see what the agent did, you cannot review it.
What to control: seven checks before any agent touches finance data
- Scope. Write down the one job, the data it may read and the systems it may change.
- Read-only first. Start with access that cannot post, delete or send.
- Human approval. Require a named person to approve anything that changes a record or leaves the firm.
- A log of every step. Prompts, tool calls, data used and outputs, kept so a reviewer can replay them.
- Test on known answers. Run it on a file where you already know the result, and compare.
- Stop rule. Define when it must halt and ask, such as low confidence or a mismatch above a limit.
- Data protection. Know where client data goes. India's Digital Personal Data Protection Rules were notified in November 2025 with an eighteen-month phase-in, so build habits now. Our DPDP audit readiness checklist covers it.
Where determinism fits
For audit conclusions, many teams want the same input to give the same answer every time. That is hard with a model that generates fresh text on each run. One practical split is to let AI read and organise documents, while fixed rules do the testing and calculation. CORAA takes this approach, explained in why CORAA uses deterministic AI.
A simple way to decide
Ask two questions. Can a person check the result quickly? Can the action be reversed? Two yeses mean a good first candidate. Two noes mean keep a person in the loop, or do not use an agent. The AI audit tool evaluation checklist helps when a vendor claims "agentic" features.
Frequently Asked Questions
What is agentic AI in simple terms?
It is AI that works toward a goal by planning steps and using tools, instead of only replying to a prompt. It can take actions, which is why it needs controls.
Is agentic AI the same as RPA?
No. Robotic process automation follows steps someone wrote in advance. An agent chooses its steps, so it can handle variation, and can also go wrong in new ways.
Can an agent replace an auditor?
No. Professional judgement, scepticism and sign-off stay with the auditor. An agent can prepare and test; a person concludes.
What is the first control to put in place?
Limit access. Give the agent read-only access to a small data set, require approval for any action, and keep a full log.
Do agents need to be disclosed to clients?
Check your engagement terms, your confidentiality duties and your firm's policy. We did not find a rule on this that we could cite, so take advice for your own case.
Related Articles
Statutory facts on this page are checked against their sources, and the page says where it relied on secondary reporting. How we verify · Report an error