Something changed in the tools your team already uses, and most people have not noticed yet. The chat window in Slack, the search box in your CRM, the assistant in your spreadsheet — they have quietly started doing things, not just answering. This month it stopped being a demo and became the default.
That is genuinely useful. It is also the moment your data stops being a back-office concern and becomes the thing standing between “the agent helped” and “the agent made a confident mess.”
The tools grew hands
In July, Salesforce made a set of Agentforce commerce agents generally available — and wired them to work inside ChatGPT and Google Search, not just Salesforce. Anthropic’s Claude now works alongside people in Slack and is expanding into Agentforce. Microsoft announced a multi-billion-dollar push to weave agents through its products, and AWS is spending to embed AI engineers directly into customer teams.
Strip away the branding and it is one shift: the assistant that used to suggest is now allowed to act. It can update a record, send a message, place an order, kick off a workflow. You give it a goal; it uses your tools to reach it. That is a promotion — and, like any promotion, it is only a good idea if the person is ready.
An agent is only as good as its boss’s filing
Here is the uncomfortable bit, said plainly. An agent acts on your data. If a customer exists three times in your system under three slightly different names, a human notices and shrugs. An agent does not shrug — it emails all three, or picks one at random, quickly, at scale. The speed that makes agents valuable is exactly what makes messy data dangerous.
One report this year put a number on the gap: only around one in ten organisations rate their data and technology as truly ready for autonomous AI. Everyone else is handing car keys to a very fast driver and hoping the road is clear.
The calm, winnable version
You do not need to slow the agents down or ban them. You need to be a good boss to them, and that is ordinary, achievable work. One clear view of each customer instead of five scattered records. Simple rules about what an agent may do on its own and what needs a human to approve. Data named and organised well enough for a machine to follow without guessing. Consent you actually trust, so the agent only ever acts on data you are allowed to use.
None of that is glamorous, and none of it depends on which agent you pick. That is the reassuring part. The companies that win with agents in 2026 are not the ones who adopted first — they are the ones whose data was ready to be acted on. Pick one small task you would happily hand to a careful new starter, get the data behind it right, and let the agent earn its keep there first. Then do the next one.