AI Models, Copilots and Agents: What's the Difference
A plain guide for CFOs being pitched AI agents in finance and payments operations.
- An AI model takes text in and writes text out. On its own it can't see your systems or change anything in them.
- An agent is a model that the firm has given tools to reach specific systems, a loop to work a case step by step, and limits on what it may change.
- A copilot suggests and a person does the work. An agent does the work and a person approves it.
- When a vendor says "agent", ask which systems it can reach, what it can change without approval, and what it costs per case.
Every finance and payments software vendor now says it has an AI agent. The word covers everything from a chat box in the corner of the screen to software that works a case from start to finish. This piece sets out what the terms mean, using one process, reconciliation, as the example throughout.
What an AI model is
An AI model, such as GPT, Claude or Gemini, is a program that takes text in and writes text out. You give it instructions and some material, called the prompt, and it writes a reply. It's priced by the amount of text it reads and writes, counted in tokens of roughly three quarters of a word each.
On its own, a model can't see your ledger, fetch a settlement file or post a journal. I find it easiest to think of a very well-read analyst sitting in a closed room. You pass a note under the door and get an answer back, and the analyst only knows what's written in the note. If you paste in a statement extract and ask why it doesn't add up, you'll often get a good answer. You still do the copying in and out yourself.
Models also state wrong things with confidence, which is called hallucination, and they're weaker at arithmetic than at reading. Both points shape how a good agent is built.
What an agent adds
An agent is the same model with three things added by the firm that runs it.
- Tools. Small pieces of code the model can ask to run, such as "get the settlement lines for this batch" or "search the ledger for this reference". The firm writes the tools and decides what each one can do. This is called tool use or function calling.
- A loop. The model calls a tool, reads what comes back, decides what to look up next and keeps going until it has an answer.
- Limits. Tools are read-only unless the firm decides otherwise. Anything that would move money or change a record goes to a person to approve, and every step is logged.
In the closed-room picture, the agent is the same analyst with logins to a few specific systems, a queue of cases and a supervisor who signs off before anything leaves the room.
Ordinary code still does a lot of the work in a well-built agent. It does the sums and the matching, because models make arithmetic mistakes. It fetches the data, so the model only sees what the tools return. And it makes the model cite the record behind each number, so a reviewer can check the answer quickly.
Four kinds of automation, side by side
| Kind | What it does | Who acts | In reconciliation |
|---|---|---|---|
| Rules software, including RPA bots | Follows fixed rules written in advance | The software, within its rules. Anything that doesn't fit goes to a person | Matches most lines on amount and reference and leaves the rest as breaks |
| Model on its own | Answers what you paste in, using only what you paste in | A person, who copies data in and the answer out | An analyst pastes a statement extract and asks why it doesn't add up |
| Copilot | A model built into a product, suggesting an answer or a draft for the case on screen | A person, who checks the suggestion and does the work | Summarises a break and suggests a cause. The analyst looks up the records and books the fix |
| Agent | Works a case from start to finish: looks things up, decides the next step, explains the result and drafts the fix | The agent does the work. A person approves the fix before anything is posted or sent | Traces a break to its cause, values it and drafts the journal and the claim |
Vendors use "copilot" and "agent" loosely, and plenty of products sold as agents today are copilots. In June 2025 Gartner estimated that of the thousands of vendors selling "agentic AI", about 130 were the real thing. In the ScalePoint Research vendor tables I only call a product an agent in production when paying customers use it to work cases start to finish.
What to ask when a vendor says "agent"
- Which of our systems can it reach, and is that access read-only?
- What can it change without a person approving it?
- Does it work a case to the end, or does it suggest and leave the rest to my team?
- Which model does it use, and can we switch to another one?
- What does it cost per case at our volumes?
- Where's the audit trail, and could our auditor follow a case from it?
I'd push hardest on cost per case. An agent calls the model several times for each case, so the cost per case decides whether the business case holds at scale. The second piece in this series follows one reconciliation break through an agent step by step and shows where those calls come from.
Terms you'll hear in a pitch
- LLM. Large language model, the common type of AI model behind GPT, Claude, Gemini and Mistral.
- Prompt. The instructions and material given to the model for a task.
- Token. The unit models are priced in, roughly three quarters of a word. A long settlement file can run to tens of thousands of tokens.
- Function calling, or tool use. The model asking a piece of code to run, such as "search the ledger".
- MCP. Model Context Protocol, a standard way to connect a model to a system's data and tools, published by Anthropic in November 2024. Payments and finance vendors including Form3, GoCardless, Embat and Payhawk now offer MCP access to their data.
- Human in the loop. A person reviews and approves the agent's work at set points, usually before money moves.
- Audit trail. The log of every lookup, result and draft the agent made.
This is one of four pieces on AI agents in finance and payments operations. The series, and the research map behind it: