You ship a support bot, and it answers well in the demo. Then a customer asks it to change a billing date. The bot explains how, and the ticket still lands on a person.
A chatbot answers. An agent acts, with tools, on your systems. This guide gives you five signs that a workflow has crossed that line, the costs an agent adds, and a checklist you can run before you fund one.
The Mechanism: Answering vs Acting With Tools
A chatbot takes a question and returns text. It may read your docs first, but the output is still words a person has to act on. If it's wrong, someone reads a bad answer.
An agent takes a goal and calls tools until the goal is met or it stops. It reads a record, decides, writes to a system, checks the result, and moves on. If it's wrong, something changes in your stack.
Anthropic's Building effective agents guide draws the same line.
It calls workflows "systems where LLMs and tools are orchestrated through predefined code paths." Agents are systems where LLMs "dynamically direct their own processes and tool usage."
That gives you three options, not two:
- Chatbot: the model answers, and a person acts.
- Workflow automation: code decides the steps, and the model fills in one or two of them.
- Agent: the model picks the steps and calls the tools, inside a boundary you set.
The same guide recommends "finding the simplest solution possible, and only increasing complexity when needed." We agree. Start with the smallest option that does the job, and move up only when the work forces you to.
5 Signs a Workflow Needs an Agent
1. The answer is only half the job
If every good answer ends with "now go update the CRM," the work is the action, not the reply. A chatbot leaves that step to a person every time.
2. You can't write the steps down in advance
Some tasks branch on what the system finds. A refund might need an order lookup, a policy check, and a carrier query, or none of those. Anthropic notes agents fit open-ended problems where you can't predict the number of steps.
If you try to script that branching, you end up with a flow chart nobody can maintain. Each new edge case adds a branch, and the script gets harder to change than the work it saves.
3. The work spans several tools
One lookup in one system is a query. Reading a ticket, checking an account, drafting a reply, and logging a note across four tools is a job. Agents earn their cost when they connect systems a person would otherwise tab between.
4. A person is the glue between systems today
Watch who copies data between screens all day. If a teammate spends hours moving context between tools and making small calls along the way, that's the pattern an agent replaces.
5. You can name the actions and the owner
This is the sign that makes the other four buildable. You can list every tool the agent may call, which ones are irreversible, and the person who owns the outcome. If you can't name those yet, you aren't ready for an agent.
The Costs an Agent Adds
An agent is more software than a chatbot. On our AI agent development page, every build ships four things a chatbot can skip. Budget for them up front.
A boundary
You define the trigger, the context the agent reads, and the exact tools it may call. Irreversible actions, like sending money, deleting records, or emailing customers, are listed and gated. You find them on a whiteboard, not in production.
A log
Every tool call, input, and result gets written down. When the agent does something odd, you need to replay what it saw and what it chose. Without a log, you can't debug it. You can't defend it either.
Evaluation cases
You collect real examples: acceptable, incorrect, unsafe, and ambiguous. Each change to the prompt, model, or tools runs against them before it ships. Anthropic recommends "extensive testing in sandboxed environments, along with the appropriate guardrails."
A takeover point
The workflow names the person who owns it and the moment the agent stops and hands back. Anthropic describes agents that "pause for human feedback at checkpoints or when encountering blockers." That pause is part of the build, not a runbook you write later.
There's a running cost too. The guide also warns that agentic systems "often trade latency and cost for better task performance," with "the potential for compounding errors." In plain terms, more steps mean more model calls and more places to go wrong.
When a Chatbot or Plain Automation Is the Right Call
Agents are not the default. Our service page lists two clear non-fits.
Retrieval alone answers the need
If people need answers from your docs, policies, or past tickets, and a person still makes the change, build retrieval with citations, which is a chatbot grounded in governed sources. Our RAG development page covers that build. It's cheaper to test because a wrong answer is visible before anything changes.
Deterministic code can express every step
If you can draw the flow as a fixed chart, write it as code. A model can still classify an email or pull fields from a PDF inside that flow. But code should own the order of steps.
B-Pay (historical case study) is a payment, team management and invoicing platform built on fixed logic. Payments and invoices want predictable steps, not a model choosing its own path.
The stakes are high and the cases are thin
If a wrong action is costly and you have few real examples to test against, wait. Collect cases first. Run a chatbot or a human-in-the-loop draft mode until you have enough data to set thresholds.
A Decision Checklist Before You Fund an Agent
Run this list with the person who owns the workflow. If you answer "no" to any of the first four, start with a chatbot or plain automation.
AGENT READINESS CHECKLIST
1. Action Does a good outcome require changing a system, not only replying?
2. Branching Do the steps depend on what the system finds at run time?
3. Tools Can you list every tool the agent may call, by name?
4. Owner Is one named person accountable for the outcome?
5. Gates Are irreversible actions listed, with an approval step for each?
6. Log Will every tool call, input, and result be stored and replayable?
7. Cases Do you have real acceptable, incorrect, unsafe, and ambiguous examples?
8. Thresholds Have you set targets for task success, abstention, latency, and cost?
9. Takeover Is the hand-back moment defined, and does the owner get alerted?
10. Off switch Can you stop the agent in one step without a deploy?Items 5 to 10 are build work, not blockers. They're what the agent costs. If you can answer 1 to 4 now, an agent is in scope and the rest becomes the spec.
For one example of the pattern, read how Hermes split sales and marketing work across agents and kept people on the judgment calls. Running several agents at once? See Mission Control.
FAQ
What is the difference between an AI agent and a chatbot?
A chatbot answers questions, and a person acts on the answer. An AI agent calls tools to take actions on your systems until a goal is met or it hands back. The difference shows up when it's wrong: a chatbot gives a bad reply, while an agent can make a bad change.
Is an AI agent more expensive than a chatbot?
Yes, in both build and running cost. An agent needs a permission boundary, a full log, evaluation cases, and a human takeover point. It also makes more model calls per task, which adds latency and spend.
Can I start with a chatbot and upgrade to an agent later?
Yes, and it's often the smartest path. A chatbot or draft mode collects the real cases and edge cases you'll need to test an agent. When the same follow-up action shows up again and again, that action is your first agent tool.
How do I know an agent is safe enough to ship?
Run it against your saved cases and check that it meets the thresholds you set for task success, abstention, latency, and cost. Confirm every irreversible action has a gate and every run is logged. If it fails any of those, keep it in draft mode.
Start With One Workflow
Pick the workflow that eats your week and run the checklist on it. If it passes the first four questions, it's a candidate for an agent. If it doesn't, a chatbot or a plain script will serve you better, and we'll tell you that.
Our Production Build engagement runs $10K to $40K against written acceptance criteria. If you'd rather scope first, the Architecture Sprint runs $3K to $5K. Tell us about the workflow and we'll reply with a fit call.
