The industry stopped selling "assistants." We stopped calling them that a while ago.

Most of what's launching in AI automation this month isn't being pitched as a chat tool anymore. It's being pitched as something that finishes a job: runs a market scan, drafts the follow-up, updates the CRM, keeps going while you're in a different meeting. The framing has shifted from "ask it anything" to "hand it a task and check the result later." That's a real change, and it's the right one — it's also exactly the shift we've built Cove's projects around from the start, so it's less a pivot for us than a confirmation.

"Assistant" was always the wrong word for what clients wanted

Nobody hires an automation to have a nice conversation with them. They want the follow-up sent, the booking confirmed, the report waiting in their inbox Monday morning. Framing these systems as chat assistants set the wrong expectation — that a person stays in the loop for every step — when the actual value is in the steps a person doesn't have to touch at all. The tools catching up to that now are just catching up to what the work always needed.

Longer jobs mean more places to go quietly wrong

An agent that only answers one question in isolation has one place to fail. An agent that scans, decides, drafts, and updates a system of record across ten minutes has ten. Multimodal, multi-step agents make good demos precisely because the audience doesn't see the retries, the fallback prompts, or the place where it almost updated the wrong record. Anyone shipping this for a client needs to see all of that before the client does.

Supervised, not automatic, is the difference that matters

The useful distinction we're seeing isn't "AI agent" vs. "AI assistant" — it's supervised vs. unattended. A repeated, well-scoped job (weekly reporting, lead follow-up, call summaries) can run unattended once it's been watched closely enough times to trust. A judgment call — anything touching money, a customer complaint, or a decision that's hard to reverse — stays supervised, with a human reviewing before it acts, not after. Getting that boundary right matters more than how many tools the agent can reach.

Start one workflow, not the whole operating team

The pitch this month is basically "run your business with a small AI operating team." Maybe that's where this goes eventually. It's not where anyone should start. One workflow, watched closely, with a clear rollback if it's wrong, teaches you more about where an agent actually breaks than five workflows launched at once ever will — and it's the only way to find the failure modes before a client's customer does.

None of this makes the news cycle less noisy. It just means the question worth asking about any new agent launch isn't "what can it do" — it's "who's checking its work, and how fast do they find out when it's wrong."