AI automation
AI agent, chatbot or workflow automation: which one does your business actually need?
Pick by who decides the next step. In a workflow your rules decide, so it is the cheapest to run and the easiest to audit. A chatbot talks to people but changes nothing by itself. An agent lets a model choose its steps and tools; pay for that only when the steps cannot be written down in advance.
What is the real difference between an AI agent, a chatbot and a workflow automation?
Compare them on one question: who decides what happens next? In a workflow, a script decides; a model, if there is one, only reads or drafts inside a step. A chatbot is a conversation window: unless it is wired to your systems, it changes nothing. In an agent, a model chooses the next step and the tool to call.
Model providers draw the same line. Anthropic describes workflows as models and tools following predefined code paths, and agents as models directing their own process and tool use. OpenAI's agent guide says simple chatbots are not agents: the model does not control how the work runs.
| Workflow automation | Chatbot | AI agent | |
|---|---|---|---|
| Who decides the next step | Rules written in advance | The visitor asks, the bot replies | The model, within the tools and limits you allow |
| What it can change | Records along a fixed route | Nothing, unless connected to a system | Whatever its tools allow |
| Running cost per task | Lowest; AI steps add model fees | Model fees per conversation | Highest: several model calls per task |
| What a failure looks like | A stop at a known step | A wrong answer a customer reads | A wrong action, sometimes after several right ones |
| Audit trail | Every step, by design | The transcript | Only as good as its tool-call log |
You were pitched an “AI agent”: how do you check what it really is?
Ask the seller six questions. The answers show which of the three you are buying.
- Which of our systems can it change, and with which permissions? If none, it is a chatbot.
- Who set the order of the steps: your developer or the model? Set in advance means a workflow.
- Which actions wait for a person's approval? Refunds, cancellations and payments should.
- When it is unsure, does it ask, stop or guess?
- Can we read a log of every action with the input it acted on?
- What does one run cost in model fees at our volume, and at twice that?
Fixed steps sold under the agent label are not a scam in themselves; they are often the sounder engineering. The trouble is paying agent prices for a script.
Why should you start with the least autonomy that solves the task?
Autonomy buys flexibility and is paid for in money, speed and predictability. Anthropic advises builders to find the simplest solution that works: agents trade latency and cost for performance, and their errors can compound. A mistake in step two stops a script; in an agent, it can steer the next five steps somewhere nobody planned.
OpenAI lists three signs that a task needs an agent; if none applies, a deterministic solution may be enough:
- decisions full of judgement and exceptions, such as approving a refund;
- rules so tangled that every update has become costly or error-prone;
- heavy reliance on unstructured input: free text, documents, conversations.
Treat the options as a ladder: a rule, a workflow with one AI step, a chatbot that hands over, an agent with one narrow job. Climb only when the rung below has failed on your real cases.
Which one fits six common business tasks?
What decides is whether the path through the task is known before it starts.
| Task | Usually fits | Why | Where a person stays |
|---|---|---|---|
| Repeat questions on the website: delivery, sizes, hours | Chatbot with a handoff | The job is conversation, answered from approved pages | Complaints, refunds, anything the pages do not cover |
| Routing new leads to a salesperson | Workflow with an AI step reading free text | Region, product and budget rules can be written | Owns the rules and leads that match none |
| Supplier invoices into accounting | Workflow with AI pulling out the fields | Receive, read, match, queue: the same every time | The bookkeeper approves before payment |
| A research brief before a sales call | Agent | Which sources to open depends on what turns up | Checks every claim before the brief is used |
| Sorting support tickets | Workflow with AI classification; an agent if solving means searching several systems | Most tickets fit a few known categories | Money questions and complaints |
| A quote from a price list | Workflow; an agent once stock and lead times must be checked in varying order | A price list is a fixed lookup; availability is not | Approves every price and date |
Four of the six are workflows: the route is known, only the input is messy, and messy input calls for an AI step, not an agent.
What does an agent add to the bill beyond the build?
An agent brings four running lines that a workflow mostly avoids. A proposal that leaves them out is incomplete.
- Model usage. In Anthropic's data, agents use about four times the tokens of a chat, multi-agent systems about fifteen. Ask for a monthly estimate at your volume.
- Permissions. OWASP's Top 10 for LLM applications traces excessive agency to three causes: more functions, permissions and autonomy than the task needs. Scope every tool and route high-impact actions to a person.
- An evaluation set: real cases with expected results, including ones where the right move is to stop. Anthropic's research team began with about 20 queries from real usage; rerun the set after every change.
- Supervision. OpenAI names two triggers for handing control to a person: the agent exceeds its limit on retries or actions, or it is about to take a sensitive, irreversible step such as cancelling an order or making a payment.
When is an AI agent the wrong purchase?
- The steps fit on one page. A workflow does the job for less and fails more visibly.
- The key action cannot be undone and nobody is set to approve it.
- Your systems have no API or export, so the agent would work through screens it can misread.
- The task comes up a few times a week; supervising it would take longer than doing it.
- Nobody owns the escalations. Handing over to “the team” means handing over to no one.
- The data may not leave your systems, and no model you may use runs where it lives.
In each case the better answer is smaller: a rule, a workflow, a chatbot that hands over, or no automation for now.
When is a chatbot enough, and when does it turn into an agent?
A chatbot is enough when the job is answering: hours, delivery terms, product questions, a call-back request. It becomes an agent the moment it may change something alone, such as booking a slot or issuing a refund; then it needs permissions, limits and an action log.
Our own assistant shows where the line can sit. In the signed-in VITON ID cabinet, VIT LIVE suggests products and studio packages as priced cards but never puts anything in the bag: the member presses Choose, checks the options and adds the item. Leaving that decision with the person keeps it a chatbot, with a chatbot's smaller risk.
A visitor chat that passes conversations to your team is the AI chatbot package: from $73, 5–8 working days.
How does VITON13 get you from the question to a working system?
Each step ends in something you can use or take to another supplier, and the route stops at the first step that solves the task. The written scope, timeline and final price come back within one working day; nothing starts before you approve them.
| Step | What you get | Starting price | Timeline |
|---|---|---|---|
| AI Use-Case Session | A 90-minute session: one process picked and scoped, a written next-step plan | $53 standard implementation or $73 custom implementation | 1 working day |
| AI Workflow Audit | One process mapped, a KPI baseline, a verdict that may be “not yet” | $73 standard implementation or $93 custom implementation | 5–8 working days |
| n8n workflow | Fixed steps across your tools, with retries, an error alert and a runbook | $73 AI-assisted or $93 human-led | 5–8 working days |
| AI chatbot with handoff | One channel answering from approved content, with a handoff to a named person | $73 AI-assisted or $93 human-led | 5–8 working days |
| AI agent with one job | Approved actions, test cases, escalation rules and an action log | $73 AI-assisted or $93 human-led | 5–8 working days |
| Production Pilot | One live workflow tied to one CRM, onboarding, two weeks of launch support | $213 standard implementation or $313 custom implementation | 5–10 working days |
How VITON13 does it
We build one AI agent for one narrow job, such as qualifying leads or answering order-status questions, and we say so when a workflow or a chatbot would serve you better. Duties, actions, tests and the escalation contact are agreed in writing first.
AI agents for business
- Price
- $73 AI-assisted or $93 human-led
- Timeline
- 5–8 working days
AI Workflow Audit
- Price
- $73 standard implementation or $93 custom implementation
- Timeline
- 5–8 working days
Production Pilot
- Price
- $213 standard implementation or $313 custom implementation
- Timeline
- 5–10 working days
The final price is fixed in writing after a short scope review.
AI-assisted: AI drafts inside defined steps, a person directs and checks every result. Human-led: a specialist does the work and AI assists.
What is included
- 1 written responsibility contract: what the agent does and never does alone
- Up to 3 actions in tools you approve, each with its own permission and limit
- 30 test cases, awkward ones included, passed or correctly escalated before launch
- 1 set of escalation rules: when, and to whom, the agent hands over
- 1 log of every action the agent takes
- Delivery: 5–8 working days once the brief and access are ready
Not included
- Model, hosting and telephony fees, paid by you to the provider
- CRM integration, onboarding and launch support: that is the Production Pilot
- An open-ended chatbot for any question
- Promises of savings or accuracy beyond the tested cases
Revisions: 2 rounds of revisions on the deliverables; a further round is quoted before it starts
Get an estimate in 1 working dayWritten estimate within 1 working day
Example
INTERNALOur own product
VJOURNAL: AI drafts inside a workflow whose checks are code
VJOURNAL, our newsroom at journal.viton13.com, follows this guide's pattern. Drafts are AI-assisted, but before a story is written a script compares its plan with the archive and returns one of four fixed verdicts, from CREATE to REJECT_DUPLICATE. After drafting, a deterministic gate keeps descriptions within 90–155 characters and rejects placeholder text and leftover chatbot phrases. In this guide's terms it is a workflow with AI steps, not an agent.
What this does not prove
- It is our own newsroom, not client work, and shows no agent acting inside a client's tools.
- It measures no hours or money saved and claims nothing about readers or rankings.
- The checks are incomplete: the duplicate report only advises, and on 19 September 2026 one paraphrased cluster still appeared in 54 stories per edition.
Read the case: VJOURNAL: our newsroom in five editions, with quality checks in code
Questions and answers
What happens if the agent takes a wrong action?
Plan for it before launch. Every action has a permission and a limit, every run is logged, and anything irreversible, such as a refund or a payment, waits for a person. Tests include cases where the right behaviour is to stop; each later failure becomes a new test.
Who owns the agent, its prompts and its logs?
At handover you receive the working agent, its action log and the escalation rules. The model account is yours, since you pay the provider directly. Write into the scope that the prompts and the test set are handed over too, so another team can maintain or replace the agent.
How long does it take to get to a working agent?
The audit takes 5–8 working days, a one-job agent 5–8 working days and a Production Pilot 5–10 working days, each counted once the brief and access are in. If the audit shows that a workflow or a chatbot is enough, the agent step and its cost are skipped.
Do I pay anything before the scope is agreed?
No. Nothing is charged until you approve the written scope and price and confirm a payment route, and you can cancel free of charge until then. After work starts, only the completed part is charged and the rest is refunded.
What does choosing AI-assisted or human-led change?
Deliverables and deadline stay the same. AI-assisted: AI drafts the job description, prompts and tests, and a developer directs and checks them. Human-led: a specialist writes and tests them, with AI helping. Choose human-led when the agent touches money, legal wording or an old system without a clean API.