A 14-person e-commerce team I work with handed ChatGPT’s agent mode a job nobody wanted: pull the week’s refund requests from their help desk, match each one against the order record, and flag anything over £200 for a human to approve. The first run took eleven minutes and needed one correction. The same task used to eat half a morning of a customer service manager’s week.
That’s the realistic version of a chatgpt agent for business. Not a digital employee running the company, and not a party trick that invents a spreadsheet. Something in between: a fast, slightly overconfident junior who never gets bored of the boring parts.
The gap between teams that get real value here and teams that burn a quarter on it usually comes down to two things. Picking the right first job, and knowing exactly where to keep your hands on the wheel.
What changes when ChatGPT stops chatting and starts doing
In a normal chat window you ask a question and get an answer. Agent mode is different: it plans a sequence of steps, opens a browser, clicks, scrolls, downloads files, fills in forms, and reports back when it thinks it’s finished. You hand it an outcome rather than an instruction.
The difference shows up in chores that span systems. “Find the three suppliers who haven’t confirmed Q3 pricing and draft a chase email for each” means logging into a portal, reading a status column, comparing it to a list, then writing three genuinely different emails. A chat reply can’t do that. An agent can attempt it, which is not the same as succeeding, but it’s a much better starting point than a blank page.
It’s a browser with a to-do list
Most of what a ChatGPT agent does in a business setting is web work. If your process lives in a desktop application with no web interface, it can’t touch it. If your data sits behind single sign-on with a hardware key, expect friction. If everything you need is a URL and a login, you’re in the sweet spot.
Five jobs where it pays for itself quickly
The pattern that repeats across every team I’ve watched try this: the agent is excellent at work that is multi-step, text-heavy, and dull, with a clear definition of done.
- Weekly reporting assembly. Pulling numbers from three dashboards into one summary, then writing the two-paragraph commentary. Roughly 90 minutes of human time, 10 to 15 minutes of agent time plus a review.
- Supplier and invoice chasing. Checking a shared inbox for unanswered threads older than five days and drafting follow-ups. The agent writes worse emails than you do, but it writes them at 7am.
- Competitor and market scans. Visiting 20 competitor pages, noting pricing changes, and summarising what moved since last month. Tedious, verifiable, and easy to spot-check.
- Spreadsheet tidying. Standardising company names, merging duplicates, fixing date formats on a 4,000-row CRM export. This is where agents quietly earn their keep.
- Form and journey testing. Signing up as a customer would, walking the checkout, and reporting where it breaks. Cheaper than most QA tools and surprisingly thorough.
If you’re still choosing between tools, this roundup of the best AI agents for productivity covers which ones are worth the seat licence and which are shelfware.
The failure modes that actually cost money
The problems are rarely dramatic. There’s no smoke, no error message. The agent tells you it finished, and you believe it.
Three patterns repeat. First, partial completion reported as full completion: it handled 34 of 50 records and cheerfully wrote “all records updated.” Second, confident arithmetic with no traceable source, where a total looks plausible and is off by 8%. Third, session timeouts mid-task, leaving a half-edited spreadsheet and no clear record of which rows were touched.
There’s a fourth, and it’s organisational rather than technical. People plug in a personal login, the agent inherits permissions nobody audited, and now a tool with a generic password can see payroll. If you want a fuller teardown of where to let go and where to intervene, this piece on what ChatGPT agent mode can actually do and where you should step in is the unglamorous version.
Guardrails worth setting before the first login
Use a dedicated account, not a person’s
Create an account that belongs to the process, not to Sarah in accounts. Give it read access first. Write access comes after two weeks of clean runs, and never to the system of record without an approval step.
Put a value threshold on autonomy
Drafting a £40 refund unsupervised is fine. Issuing it is not. Set the number where a human signs off, write it down, and put it in the prompt rather than trusting memory.
Log the review time, not just the run time
Every business case I’ve seen collapse did the same thing: counted the 12 minutes the agent took and ignored the 25 minutes a person spent checking it. Track both for the first month. If review is longer than the task, you’ve automated nothing.
The money question, with actual numbers
ChatGPT Business runs roughly $25 to $30 per seat per month on annual billing. Pro is $200 a month for heavy individual users. Set that against an operations contractor at £35 to £45 an hour and the maths is not subtle: a single task that saves two hours a week clears £280 a month in saved time against a £30 licence, before you count the review overhead.
The honest calculation includes a failure tax. If 15% of runs need redoing and each redo costs 20 minutes, that’s real money coming out of the savings. Budget for it and the whole thing still works. Pretend it doesn’t exist and your first bad month will kill the project.
For leaner teams trying to decide which subscriptions survive the card statement, this breakdown of which AI agents are worth paying for is a useful sanity check before you commit to a dozen seats.
How to tell whether the output is any good
Run a blind audit. Take ten outputs the agent produced and ten a person produced, strip the labels, and ask a colleague who doesn’t know the source to grade them on accuracy and usefulness. It takes an hour and settles arguments faster than any demo.
For research-heavy work, add a second test: pick three factual claims at random and verify every one against a primary source. Agents are at their most convincing precisely when they’re wrong, and the tone doesn’t waver. The technique in this guide to spotting AI agents that make things up transfers directly to business reporting.
When a single ChatGPT plan stops being enough
Around the point where three departments depend on the same agent runs, consumer tooling starts to chafe. No shared audit trail, no role-based access, no service-level agreement, and no way to answer an auditor who asks what the agent did on the 14th. That’s the moment to evaluate platforms built for it, and this comparison of enterprise AI agents under real production load is blunt about which ones buckle.
Until then, a £30 seat and a well-chosen task will outperform a six-figure deployment that’s still in procurement.
Pick the job you’d give a smart intern in their second week
Not the first week, because they don’t know the systems yet. Not the sixth month, because by then you’d be asking them about strategy and they’d have opinions you’d want to hear.
Second week looks like this: the task is clearly defined, the inputs are findable without tribal knowledge, and a mistake is annoying rather than catastrophic. Refund triage. Supplier chasing. Competitor price checks. Cleaning a CRM export. Onboarding paperwork.
Start with one of those, log the review time for two weeks, and widen the scope only when the checking is cheap and the errors are boring. That’s the whole method. The teams getting real leverage from a ChatGPT agent for business aren’t the ones with the cleverest prompts. They’re the ones who picked a dull process, measured it honestly, and didn’t hand over the keys to anything that touches money on day one.

