The short answer
When you use ChatGPT or Claude in a chat window, you ask, it answers, and you do everything after that. An AI agent is given an ongoing job, access to the tools that job needs, and rules about when to stop and check with you. It does the work and reports back.
Which AI is smarter matters far less than who ends up doing the work. That shift decides what you can hand over, what can go wrong and what you need to set up first, and much of the confusion about AI in professional firms comes from treating chat and agents as the same thing.
Where do most firms stand with AI today?
Law firms, accounting practices, agencies and property businesses tend to sit in one of three places.
Individual chat accounts. A few people pay for ChatGPT or Claude and paste in emails to answer or documents to summarise. Each of them saves some time, but nothing is shared, nobody measures the gain, and the rest of the team learns nothing from it.
Saved prompts and templates. Someone writes instructions that work well and saves them in a shared document or a custom assistant. Quality goes up, but a person still has to open the tool, paste in the material and copy the result back to where it belongs.
Rule-based automation. A web form creates a client record, a template fills itself in, a message goes out when a status changes. These tools are reliable, and they're no use for anything that needs reading or judgement.
In all three, a person carries the job from start to finish and AI speeds up one step of it. There's a side effect, too. When ten people each keep a private setup, one person's discoveries never reach the others, their prompts leave the firm when they do, and when a client asks how their information is being handled, nobody has a clear answer.
What makes something an AI agent?
Four ingredients turn a chat model into an agent.
An ongoing job. Instead of a one-off question, it gets a standing task, such as "every Monday, list the clients still missing year-end documents and draft a reminder to each one for me to send." It works toward that every time the trigger comes round.
Access to tools. It can read and write where your work lives: a shared folder, a task list, your practice-management system, the chat app you use. Without access, all it can do is suggest what someone else should do.
Knowledge of your firm. Your templates, tone, fee structure, client history and the exceptions to your usual rules, which is roughly what a new colleague would need. Firms underestimate this part more than any other, and an agent without it turns out generic work faster.
Limits. What it must never do, and when to stop and ask a person. Limits are what make it safe to hand work over in the first place; How AI agent safety works explains how to set them.
Add a trigger (a set time each morning, a new file landing in a folder, a matter changing status) and the agent starts work without anyone opening an app.
How does an agent differ from chat day to day?
Who carries the work
With chat, the AI writes a draft and you handle everything after it: sending, filing, updating the records. With an agent, the steps around the draft become part of its job. It prepares the reply, the record update and a reminder to follow up, and anything that goes to a client or changes a client record waits for your approval. Most of the other differences follow from this one.
Where the output lands
Chat lives in a browser tab, so whatever it produces has to be copied, reformatted or retyped into your systems, and every hand-off is a chance for something to slip. An agent works inside those systems, and its output ends up where it belongs.
How consistent the results are
Chat starts from nothing each time. You supply the background and the instructions, and the result depends on who's typing and how busy they are that afternoon. An agent is set up once and follows the same steps with the same knowledge, so the twentieth file gets the same treatment as the first.
Who holds the context
With chat, you're the memory. You know the client's history, the firm's position on fees and the awkward detail from last year, and you type it in again every time. An agent keeps that knowledge in its setup and can look things up in your records, which is what lets it handle a task you never personally walked it through.
Who answers for it
With chat, you check the answer and you take the action, so responsibility is obvious. An agent does some things while you aren't watching, which means you need a record of what it did, what information it used and why. If you can't see what an agent did, you can't safely hand it work, however good that work looks.
Example: chasing missing documents before a tax deadline
Picture a small accounting practice in early December. The online Self Assessment deadline is 31 January, and thirty clients still owe receipts, bank statements or P60s.
By hand. The senior bookkeeper keeps a spreadsheet of who has sent what, works through it between other jobs and writes reminders one at a time. By mid-January four clients have gone quiet, and nobody has chased them a second time.
With chat. She pastes the spreadsheet into ChatGPT and asks for reminder emails. The drafts are fine. She still has to check the list, personalise each email, send them and update the spreadsheet, all as separate jobs. The writing got quicker and the process stayed the same.
With an agent. Every Monday the agent compares the checklist against what has arrived in the shared folder, updates the tracker and drafts a reminder for each client who is still outstanding, naming exactly what's missing. The drafts wait for her to read and send. Anyone who has been silent for two weeks goes to the top of her list, and every step is logged with the time and the file the agent looked at.
What stays with her: how to handle a client who is late every year, whether to charge for a rushed return, and the relationship itself. The agent pulled together what she needed for all three, and she made the decisions.
When is chat the better choice?
Agents don't replace chat, and for plenty of work chat is the right tool. It suits:
- thinking through a new problem, where a back-and-forth conversation helps you explore
- first drafts you'll rewrite anyway, where speed of starting is what you want
- one-off analysis that isn't worth building anything for
- learning about a subject by asking questions
- work that's different every time, where a person should lead
Chat costs little, needs no setup and works straight away. It's also perfectly reasonable to decide a task isn't worth automating. If it happens a few times a month, a mistake would cost nothing, or nobody would notice whether it got done, an agent will cost more to set up and supervise than it saves.
When does an agent earn its place?
Agents do better with:
- regular work that follows a pattern, such as forty bank reconciliations a month that look broadly alike
- tasks with several steps across systems: read the file, check the records, update the tracker, tell the right person
- work that has to happen when you're not there, at 7am, on a Sunday or during a meeting
- deadlines that are easy to miss
- jobs you keep meaning to do and never get to, like chasing clients for missing documents
A rough rule: if a task repeats and mistakes or delays matter, look at it. If it repeats but doesn't matter much, leave it. If it doesn't repeat, keep it with a person.
What goes wrong is different too
When chat gets something wrong, you see a bad sentence, fix it and lose a minute.
When an agent gets something wrong, it may act on the mistake: send the wrong letter to the wrong client, update the wrong record, or chase a matter that closed last month. That kind of error can lose you a client.
So most of the effort in setting up an agent goes into limits: what it can reach, what it may do without asking, and how quickly you can stop it. Firms that treat setup as "writing better instructions" tend to run into trouble, and the ones that think carefully about permissions and approval steps usually do fine. Our guide to AI agent safety covers how, including ten questions to put to any agent provider.
How do the costs compare?
Chat is a monthly subscription for each person. It's easy to approve and hard to justify, because nothing links the bill to results.
An agent has three kinds of cost:
- Setup, mostly at the start, when you define the job and give it the knowledge it needs.
- Running, which grows with the amount of work it handles.
- Supervision, the time people spend checking its work. This falls as the setup improves, but it never reaches zero.
The fair comparison is twelve separate subscriptions, each helping one person, against one system that does a defined job for the whole firm and improves as it learns how you work.
Should you go straight to an agent?
Usually not. Firms tend to move through these stages, and the ones that skip ahead often decide AI doesn't work for them:
- No AI: everything is done by hand.
- Individual chat: a few people experiment on their own.
- Shared templates and prompts: repeatable, but a person still does each step.
- Rule-based automation: fixed rules move information around, with no judgement involved.
- An agent with approval steps: it reads, decides within limits, prepares or acts, and asks when it's unsure.
- Several agents working together, each handling a different stage of a process.
Most professional firms are at stage 1 or 2. Jumping to stage 4 tends to fail because the agent inherits a process nobody has written down, and you can't hand over a job you haven't defined. Stage 2 is where that writing-down happens, and doing it is how you find out what the job involves.
A four-question test for any task
- Does it repeat? At least weekly, and in a recognisable form.
- Does it touch more than one system? For example email, client records and documents.
- Does it need doing when you're not available? A deadline that won't wait for your diary.
- Is the output similar each time? The same kind of result, case after case.
Four yeses: an agent is worth designing. Two or three: start with better templates and a written process. One or none: use chat, or leave the task as it is.
Plenty of tasks fail question 4. That usually means the task rests on judgement, and judgement belongs with a person anyway.
Common questions
Is ChatGPT an AI agent?
A normal ChatGPT conversation works as an assistant: it answers and you act on the answer. Some chat apps now add agent-style features that can browse or carry out steps for you, but a one-off conversation without a standing job, tool access and limits is still chat. What makes something an agent is the ongoing job and the permission to act within rules.
Can an AI agent work without anyone checking it?
For low-risk work that's easy to undo, such as internal summaries or tagging files, it can run with occasional spot checks. Anything that goes to a client, changes a client record or touches money should wait for a person to approve it. The amount of checking drops as the setup improves, but it never disappears.
Do I need technical skills to use an AI agent?
Not for the part that matters most. Defining the job, gathering your documents and listing the exceptions takes knowledge of your firm, not code. Connecting the agent to your systems is technical, and a provider or consultant usually handles it; with Able°, we set it up with you.
What's a good first job for an agent?
Something that repeats weekly, has a clear end point, is cheap to fix if it goes wrong, and that someone on the team dislikes doing. Chasing missing documents or preparing a weekly status summary are typical starting points. How to get the best out of an AI agent goes through choosing and briefing that first job.
Next in this series: How to get the best out of an AI agent and How AI agent safety works.
We run a 30-minute diagnostic for professional firms. We look at what you do by hand, show you what's worth automating, and tell you what to leave alone. Book a diagnostic.