Every AI roadmap conversation eventually arrives at the same fork: do we build something that answers, or something that acts? The words get used interchangeably in vendor decks, but the engineering, risk profile, and budget behind them are entirely different. Getting this call wrong in either direction is costly — an agent where an assistant would do means months of unnecessary orchestration work; an assistant where you needed an agent means your team still does all the clicking.
The distinction that matters
An assistant responds within a conversation: it drafts, summarizes, answers, and recommends, but a human executes every consequence. An agent plans and acts across systems: it reads a queue, calls APIs, writes records, and verifies its own work — the human supervises outcomes rather than performing steps. The dividing line isn't intelligence. It's who holds the pen when something changes in a system of record.
Four questions that settle it
- Reversibility — if the action is wrong, can it be undone in one step? Drafts are reversible; sent wire transfers are not. Irreversible actions demand either human approval gates or a much longer trust-building period.
- Frequency — does the task happen ten times a day or ten thousand? Below a threshold, a human-with-assistant is cheaper and safer. Agent economics only work at volume.
- Judgment density — how much of the task is mechanical versus judgment? Tasks that are 90% mechanical with clear rules suit agents; tasks where every case is an exception don't.
- Blast radius — what's the worst plausible outcome of a wrong action? A mislabeled ticket embarrasses nobody; a wrong compliance filing has a lawyer's phone number attached.
Rule of thumb: start with an assistant, instrument what humans do with its output, and promote the workflow to an agent only where the data shows the human step adds no judgment — just latency.
The promotion path
The teams that succeed rarely jump straight to autonomy. They ship an assistant, log every human correction, and after a few hundred runs they know precisely which steps the human never changes. Those steps get automated first, with the human moved to an approval gate. Eventually the gate opens only for low-confidence cases. Autonomy is earned in increments, and each increment pays for itself before the next one ships.
What this means for your first build
If you're choosing your first AI system, bias toward the assistant unless all four questions point the other way. You'll ship in weeks instead of months, learn what your data actually supports, and build the operational muscle — evals, monitoring, escalation — that an agent will demand later. The assistant isn't the compromise option; it's the reconnaissance.