Learn · Compare

A chat tool or an AI agent: which does this job need?

Published 31 October 2026

Short answer. For work a person does once, or does differently each time, a chat tool such as ChatGPT, Microsoft Copilot Chat, Claude or Gemini is usually enough. An AI agent is worth setting up when the same job comes back, needs your company’s own systems, has to act in them, needs a person’s sign-off, and must leave a record. If the data the job needs is scattered or does not agree, neither will do the job well until the data is fixed.

Run the five questions below on one job, and the verdict table tells you which of the four it points to.

What a chat tool can do today

A chat tool is an AI assistant a person types into: you ask, it answers, and you decide the next step. An AI agent is a system that takes a goal, breaks it into steps and uses tools to carry them out.

Chat tools do more than answer questions today. With the right plan and settings, each of the four can read files you give it and reach some of your company’s sources:

Chat tool	What it can connect to, depending on plan and settings	Source
ChatGPT	Apps such as Google Drive, SharePoint and Slack; some apps also let it take actions, such as updating a file, often after you approve. Agent mode carries out tasks on a virtual computer and asks before actions of consequence.	OpenAI Help Center; OpenAI, Introducing ChatGPT agent
Claude	Connectors to Google Drive, Gmail, Slack and other services, and custom connectors to your own MCP servers. On Team and Enterprise plans, an owner can limit which actions a connector may take, for example read but not write.	Claude Help Center
Microsoft Copilot Chat	The web and files you upload. A Microsoft Copilot licence (until recently Microsoft 365 Copilot), or agents your organisation sets up, add your emails, files and SharePoint content.	Microsoft Learn
Gemini	Gmail, Docs, Drive and Calendar for a Google Workspace account, when the administrator allows it.	Google Workspace Admin Help; Gemini Apps Help

An MCP server is a program that gives an AI assistant a defined way to read your data and, where you allow it, act in your systems.

In every case a person drives the work: they start each task, check each answer and decide what happens next. The dictionary puts it this way: “When an assistant makes your slides or builds a budget sheet, the work is still yours: you prompt every step.” That is the line this page draws.

Five questions to ask about the job

Take one job and answer yes or no:

Does the job come back? It runs every day, week or month, in the same way.
Does it need your company’s own data and systems? More than one person can paste or upload: the accounting system, the warehouse, the inbox.
Must it act in a system? It posts, files, updates or sends something, beyond writing text for a person to use.
Must a person approve the result before it counts? For example before an entry is posted or an email goes outside.
Must you be able to show afterwards what was done? Which source it used, which rule it applied, and who approved it.

And one about the data:

Is the data the job needs in systems you can reach, and does it agree? If the same figure differs between the accounting system and the spreadsheets, answer no.

The verdict

Your answers	Verdict	What it means
Mostly no to questions 1 to 5	A chat tool is enough	Use the chat tool you have. Give it the files and context for each task, and check its answers.
Yes to 1 and 2, no to 3	A chat tool connected to your data	The job is answering questions from your own figures. Connect the chat tool to your data through the data foundation, so each answer shows its source and each person sees only what they have access to. The Ask your data job works this way.
Yes to 1, 2 and 3, and usually 4 and 5	An agent is worth it	Set up an AI agent for the job, with access only to what the job needs, rules, guardrails, a logbook and an operator who signs off.
No to the data question	Neither yet: fix the data first	Build the data foundation, then ask the five questions again.

The operator is the person in the company who gives the agent its instructions, reviews its work and signs off. Which steps the operator approves is decided step by step with three sign-off zones: “acts alone”, “asks first” and “the operator signs off”.

Four jobs run through the test

Job	Answers	Verdict
Draft a reply to a supplier about a late delivery	Comes back, but differs each time; one person has the facts; no system action	A chat tool is enough
Answer “why did costs rise in that shop last quarter?” whenever a manager asks	Comes back; needs the warehouse; no system action; inside the company	A chat tool connected to your data
Read every supplier invoice and prepare the entry in the accounting system	Comes back daily; needs the accounting system; acts in it; the operator approves; every entry must show its source	An agent is worth it
Draft the month-end report when the accounting system and three spreadsheets disagree	Comes back; needs several systems; the data does not agree	Fix the data first

When a chat tool is enough

A chat tool is enough for most writing, summarising, analysing a file someone hands you, and one-off questions. It is also enough for a job that repeats, if a person is happy to do each step and check each answer.

If that is where your job lands, use the chat tool you already have, and learn to give it good context. You may not need anyone’s help.

When an agent is worth it

An agent earns its setup when the job repeats and has to be done in your systems, with a person approving what matters and a record of every step. An agentic knowledge worker, Gathena’s name for an AI agent set up to do a job in your company, takes the goal once and comes back when the job is done, or with the one question it could not answer alone.

What it needs to do that work, and why a chat tool on its own stops short, is explained in What an AI agent needs to do real work. Which steps it may take without a person is covered in When may an AI agent act on its own?

When the data comes first

An agent or a connected chat tool reads what your systems hold. If the systems disagree, it reports the disagreement, or it picks one figure without saying why. The data foundation is your connected, cleaned data, where every figure can be traced to its record. Building it is the first thing worth doing, whatever you decide about AI. The data foundation, explained

Checked against the vendors’ documentation on 31 October 2026.

Sources

Product names are trademarks of their owners, used to identify the products; no endorsement is implied.