Two systems can file the same expense report and behave nothing alike. The first runs a fixed script: a receipt lands in the inbox, the amount gets read, anything over $75 gets flagged, the row goes into the accounting sheet. It fires about 40,000 times a month and never gets bored. The second opens the same email, notices the receipt is in Portuguese, works out the currency is reais, hunts for a matching line in the ledger the script never knew about, and decides for itself to hold the claim because it looks like a duplicate of one filed in March.
Both are software. Both save someone an afternoon. But only one of them is guessing, and that guess is the whole difference between automation and an AI agent.
Automation is a contract. An agent is a judgement call.
Traditional automation, whether it’s a Zapier zap or a UiPath bot someone spent six weeks building in 2021, encodes a promise: if X happens, do Y. The rules live in code, so behaviour is predictable, testable and cheap. An invoice bot chewing through 10,000 documents a day at a fraction of a cent each is a genuinely beautiful thing, and no agent will beat it on cost or reliability.
The weakness sits at the edges. Change a field label on a supplier portal and the bot that scraped it by position breaks at 2am. Nobody notices until the invoices pile up. Rule-based systems don’t adapt; they fail loudly or, worse, keep running and quietly write garbage into your ERP.
Inside an agent’s decision loop
An AI agent wraps a language model in three things: tools it can call, memory of what it has already tried, and a loop that keeps going until the goal is met or it gives up. Roughly:
- Perceive the current state (a page, an API response, an inbox)
- Plan the next step toward the goal
- Act by calling a tool: search, click, query, send
- Check what came back and decide whether to continue
That third step is where things get interesting. Because the model picks the tool at runtime rather than following a hardcoded path, it can handle work nobody wrote a script for. Ask it to reconcile two spreadsheets with different column names and it will work out the mapping. A traditional bot needs that mapping specified in advance.
The scope question matters more than the label
Some vendors slap “agent” on anything with a chat window. The distinction that actually matters is how much freedom the system has to act without asking. It’s a real spectrum, and understanding how AI agents differ from assistants that wait for permission is worth ten minutes before you sign a renewal. An assistant drafts the reply. An agent sends it. That gap is where the risk lives.
Where agents fall apart
Non-determinism cuts both ways. The same prompt can produce two different trajectories on two different days, which makes testing a nightmare. Ask a browser agent to book a hotel and it might succeed in 12 steps, fail in 9, or wander into a loyalty sign-up page and start filling in a form you never asked for.
Accuracy compounds, too, and not in your favour. At 95% per step, a 20-step task lands around 36% end to end. Push per-step accuracy to 99% and the same task clears 82%. That arithmetic explains why most agent work in production is narrow, short-horizon and supervised.
Cost is the quieter killer. A rule-based bot run costs practically nothing. An agent task burning 30,000 tokens across several model calls might land between $0.05 and $0.30, so 50,000 tasks a month gets real. Inference prices keep falling, but the comparison still favours scripts for high-volume, low-variance work.
Where automation quietly gives up
Plenty of jobs are too messy for rules and too expensive for humans. Web work is the obvious case: portals with no API, forms that change, sites behind a login. That gap is what browser agents such as Skyvern, a browser agent that automates the web like a human, are built for. Systems like Browser Use teach agents to click, type and browse by reading the page visually rather than relying on brittle selectors.
Enterprise knowledge work follows the same shape. Searching across 30 SaaS tools is a rules problem. Deciding what to do with the answer, then drafting the renewal summary, updating the CRM and pinging the account owner, is an agent problem. That’s the shift behind Glean Agents moving enterprise search from finding to doing.
The hybrid is where most teams land
Pure automation and pure agency are both rare in practice. What ships looks like a deterministic spine with an agentic brain bolted on: the model handles interpretation and exceptions, then hands off to a tested, auditable script for the actual transaction.
Customer support is a decent illustration. A conversational layer can carry intent and tone, but you probably don’t want it freelancing on refunds. Platforms like Yellow.ai pair natural language with rule-bounded actions, which is the split that survives a compliance review.
Choosing between them without wasting a quarter
Start with the shape of the task, not the vendor demo.
- Same input, same output, every time — script it. Rules are cheaper and easier to audit.
- Input varies and judgement is needed — agent territory, usually a few hundred to a few thousand tasks a day.
- Payments, deletions, anything regulated — the agent proposes, a human or deterministic system disposes.
- Interface likely to change — agents survive redesigns better than selector-based bots.
- Sub-second latency required — keep model calls out of the critical path.
What to ask before you buy either
Where does the audit trail live? If the system takes an action and nobody can reconstruct why, you have a liability rather than a tool. Logs of every tool call and model output are non-negotiable in finance and healthcare.
What happens on step failure? Good systems retry twice, then escalate to a human queue with context attached. Bad ones loop, burn tokens and mark the job complete anyway.
Who owns the fallback? If your model provider has an outage at 9am on a Monday, does the invoice pipeline stop? Hybrid designs keep the deterministic path running while the agent is down.
And what does success cost per completed task, human review included? A $0.30 agent run that saves 90 seconds of a $40-an-hour employee’s day is a bargain. The same run sitting behind a mandatory four-minute review is not, however good the demo looked.