The "Which AI Tool" Question Is Nearly Settled. The One After It Isn't.

Almost every organisation I speak to opens with the same question, which is some version of "so which one should we be using", and it is a reasonable thing to ask and an increasingly unimportant thing to settle, because the honest answer has become short enough to give in a sentence.

Ethan Mollick, who writes what is probably the most useful running guide to this for non-specialists, updated it in July with what amounts to a two-line recommendation: pick ChatGPT or Claude, pay the twenty US dollars a month, and give an agent a real task from your real work. His summer 2026 edition of "An opinionated guide to which AI to use to do stuff" is worth reading in full, but the part that matters for a business deciding what to do is that the choice between the two leading systems is no longer where the difficulty sits. For low-stakes work, he notes, even the free models are good enough that you should simply use whichever you prefer, and the case for paying is not that one brand is smarter but that a paid account is what gives you an agent at all.

What changed is not the model, it is what the model is holding

Until fairly recently, using AI meant typing into a chat window and reading what came back, which made the tool comparison a fair question, since you were essentially comparing the quality of two conversations. That is no longer the main way this technology gets used for work. Mollick's framing is that an agentic system gives the AI a computer, either one the AI company provides or, if you install the desktop application and switch it on, your own, and the practical consequence is that a single instruction can now consume the better part of an afternoon's worth of work before it reports back.

His own demonstration is mundane enough to be persuasive. He told both systems to connect to his Gmail, prepare for a seminar he was giving on a particular date, build presentation material, and answer any outstanding messages on the topic. Both worked out what he meant, including which month the date fell in, researched, drafted, and came back roughly ten minutes later with teaching materials and a reply to a colleague, which he estimates would have taken a couple of hours to do by hand. That is the capability people are actually buying now, and it is not meaningfully different between the two products.

The difference showed up in what happened next

One system produced a draft email for him to look at. The other sent it. Mollick is clear that this was his own doing, since he had previously granted ChatGPT permission to send mail on his behalf while Claude had been told to check first, and the point he draws from it is the one I would underline for any business here: when these systems are used for real work, the permissions matter a great deal more than the brand on the login screen. Both companies let you decide whether the agent must ask before it sends, spends, changes a file, or buys something, and both default to asking, which is the setting to leave alone until you have watched the system make its own particular kind of mistake a few times.

There is a second reason to be careful with what an agent can reach, which is prompt injection. An agent that reads your inbox and browses the web will sooner or later encounter text somebody else wrote for it to find, phrased as an instruction rather than as content, and while the models have become noticeably more resistant to this, nobody serious is claiming it is solved. An agent with wide access and standing approval to act is, in that light, less a productivity setting than an exposure, and the reasonable response is not to avoid agents but to be deliberate about the surface you give them.

Why this matters for Malaysian SMEs specifically

A twenty-dollar subscription does not go through procurement, does not reach the board, and in most of the businesses I work with is expensed by whoever signed up first, which means the decision that actually carries risk is being made by an individual rather than by the organisation. That is the same accountability gap I wrote about when the Securities Commission began consulting on named AI oversight at listed-company boards, showing up here in a much smaller and more immediate form. Nobody needs a policy document to decide which chatbot to open. Somebody does need to have decided, on the record, whether the agent connected to the company inbox is allowed to reply to a client without a person reading it first.

The cost picture deserves one line of caution as well. Mollick notes in a footnote that the twenty-dollar tiers include real but limited agent usage, and that agents consume those limits quickly, so the more expensive plans are mostly buying more hours of AI labour rather than a smarter AI. For a small business budgeting this properly, that is the useful way to think about the upgrade: you are buying capacity, not capability, and you should know which one you are short of before paying for the other.

What This Means for You

If you have been holding off on a decision because you were waiting to work out which AI product is best, you can stop waiting, since either of the two obvious answers will do and the difference between them will not determine whether this works for your business. What will determine it is whether you have handed an agent a real piece of your actual work, looked carefully at what came back, and asked for changes rather than simply accepting or discarding the result, which is a managing skill rather than a technical one. And before any of that, whether you have decided, deliberately and in writing, what the thing is allowed to do without asking. Running in manual mode longer than strictly necessary costs you a little speed and buys you the only thing the agent cannot supply, which is your own judgment about whether the work is any good. The tool question has an answer now. The permissions question is still yours to answer, and it is the one that will matter in a year.