# Chatbot, workflow tool, or AI operator: which job is which? | BYOM blog

URL: https://byom.co/blog/ai-operator-vs-chatbot-vs-workflow-tools  
Markdown: https://byom.co/blog/ai-operator-vs-chatbot-vs-workflow-tools.md  
Last updated: 2026-10-02

> Three kinds of tool compared on published facts: what they cost, what they can change and what the law says when an automated answer misleads.

By Kina (Checked by the BYOM team). Published 2026-10-15. 9 minute read. Series: Getting started.

## Key takeaways

- A McGill study of 500,000 interactions found 71% of customers prefer a human agent over a chatbot.
- Zapier counts successful actions as tasks and Make counts operations as credits, and Flow is a free app, so the same job can bill differently on each.
- The DMCC Act 2024 treats misleading information about a product as an unfair commercial practice, whoever or whatever delivers it.

A shortlist for AI help in a Shopify store usually mixes three kinds of product that answer different questions. A storefront chatbot answers a shopper. A workflow tool runs steps when something happens. An agent tool plans its own steps toward a goal. Each category has published facts about how it works, what it costs and where it fails. This post sets them side by side so you can see which job each one is for, what the law says when an automated answer misleads a customer, and what to check before you buy.

## Storefront chatbots answer the shopper

IBM describes the AI assistant, the category most chatbots fall into, as an application that understands natural language commands and uses a conversational interface to complete tasks for a user, and it calls assistants reactive because they need defined prompts to act. A storefront chatbot is the version that sits on your site and talks to people who are buying. Its job ends when the shopper has an answer.

What shoppers make of that is measured. A study by Vivek Astvansh of McGill University, reported in The Conversation, analysed more than 500,000 customer service interactions between customers and either human agents or chatbots at a large North American retailer. The reported findings are that 71% of customers prefer interacting with a human agent over a chatbot, and that 60% report that chatbots often fail to understand their issue. Most customer questions fell into six types: orders, coupons, products, shipping, accounts and payments. Customers avoided chatbots for the sensitive ones, shipping and payments.

The McGill study covers one retailer in North America, not UK stores, so treat the percentages as one large sample and not a national figure. It still gives a practical rule. Chatbots are suited to the question with a stable answer, such as a delivery window or where the returns policy is. They are a poor fit for the questions customers care about most, where a wrong answer costs money or trust. The study describes customers avoiding or distrusting chatbots in exactly those cases.

## Workflow tools run steps when something happens

Shopify's help centre describes Flow as an ecommerce automation platform for tasks and processes within your store and across your apps. It has three working parts. Triggers monitor your store for events, conditions filter whether the actions should go ahead, and actions are the steps taken in response. The trigger reference lists, among others, a product variant inventory quantity changing, a scheduled time and a product being created. Flow is available as a free app on Basic, Grow, Advanced and Plus plans. The Send HTTP Request action is limited to Grow, Advanced and Plus, and custom partner app tasks are limited to Plus.

The help centre pages for Flow's building blocks do not describe a manual approval step. A workflow can still ask a person something, since an action can send an email or a message to someone. But any approval would be something you build around the step, and nothing in the reference documents a built in point where a person has to say yes before an action changes the store.

General workflow tools, Zapier and Make, price by usage, and the unit matters. Zapier's pricing page says a task is counted whenever Zapier successfully completes a unit of work. Triggers, polling for new data and its built in data tools do not consume tasks, and failed actions do not count. The Free plan includes 100 tasks a month, Professional starts at $19.99 a month on annual billing, and Team starts at $69 a month on annual billing. Annual billing gets 33% off. If you exceed your allowance with pay per task enabled, overage is priced at 2.5 times the base rate on monthly plans and 1.25 times on annual plans.

Make prices in credits. Its documentation says that only features triggered by scenario runs, such as apps, modules and some in app features, or the AI agent's chat, use credits. Most commonly one operation equals one credit for non AI apps and standard third party AI applications. AI features can instead use dynamic rates based on token usage, file size, pages processed or processing time. When a user runs out before the next billing cycle, the documentation offers three options: upgrade the subscription, buy additional credits or enable automatic extra credit purchases. The Zapier figures above are as published when we read the page on 2 October 2026, and in dollars.

Both vendors have started selling AI features too, and the billing unit changes when they do. Zapier's page says its Agents and Chatbots products operate on separate billing units, activities and feature tiers respectively, outside the task allowance. A merchant who adds an AI step to a workflow should find out which meter that step runs on before turning it on across all orders.

A worked example shows how the unit changes the bill. Suppose a workflow with five action steps runs for each of 400 orders a month. On Zapier that is 400 times five, or 2,000 tasks, since the trigger does not count. That is twenty times the free allowance of 100. On Make, if each step is one operation and so one credit, it is also 2,000 credits. The same automation can sit inside a free allowance on one tool and need a paid plan on another, depending on how each counts a step, so price your busiest realistic month, not an average one.

## Agent tools plan their own steps

IBM defines an AI agent as a system or program that can autonomously complete tasks on behalf of users or other systems by planning its own workflow and using available tools. That is the dividing line from a workflow tool. In a workflow you write the path in advance: when this happens, do that. In an agent the software chooses the next step based on what it finds.

The extra freedom is the reason to want one and the reason to be careful. A workflow tool will do the same thing every time, so its failures are the ones you built into it. An agent can handle a case you did not anticipate, and it can also handle one badly. The Model Context Protocol specification, which describes how models call tools, says tools are model controlled, so the model can discover and invoke them automatically, and says there should always be a human in the loop with the ability to deny tool invocations. That is a recommendation in the specification. It is a feature of the product only if the vendor has built it.

A workflow tool that goes wrong repeats the same error on every run until someone notices. A condition written too loosely applies to every product that matches it on the next trigger, which is why the first run of any new workflow is worth watching on a test store. An agent that goes wrong can fail differently each time, which makes the failure harder to reproduce. Both are manageable with approval before effects and a record of each run.

The practical question for any agent tool is where it stops. A product that reads your store and drafts is a different purchase from one that writes to your store unprompted, even if the two look identical in a demo.

## The three side by side

The table compares the categories on the four things that matter after you have bought one: who starts it, what it can change, what record it leaves, and what it costs you when it is wrong. The entries describe the category in general. A particular product can differ, so use the table as a list of questions.

|  | Storefront chatbot | Workflow tool | Agent tool |
| --- | --- | --- | --- |
| Who starts it | The shopper, by asking | An event or a schedule you set | A request, or its own plan toward a goal |
| What it can change | Usually nothing in the store, but it speaks for you | Whatever actions you wire up, such as store data or web requests | Whatever tools it is given, chosen at run time |
| What record it leaves | A transcript of the conversation | A run history in the tool; check its retention | Depends on the product; ask to see one |
| What it costs when wrong | A misleading answer you are responsible for | The same error repeated on every run | A wrong choice with real effects, unless a person approves first |
| How it is priced | Per seat or per conversation | Per task or per credit, from the pages above | Per seat, per outcome or per usage |

Matching the job to the category is then a matter of three short tests, applied to each job on your list one at a time. If a shopper is asking and the answer is stable, use a chatbot and keep its sources tidy. If a known event should always lead to the same step, such as a stock quantity change leading to a message to your buyer, use a workflow tool, because Flow's inventory trigger exists for exactly that. If the work needs judgement across many records, such as checking 300 product descriptions for the same mistake, an agent tool fits, as long as it drafts and a person approves before anything is published.

The pricing row also shows how each scales. Chatbots and per outcome tools scale with your volume of conversations. Workflow tools scale with the number of steps your automations take. A cheap plan on one measure can be expensive on another once your store gets busy.

## What UK law says when an automated answer misleads

A chatbot speaks for your business, and UK consumer law does not have a separate rule that excuses an automated speaker. Part 4 of the Digital Markets, Competition and Consumers Act 2024 prohibits unfair commercial practices. Section 226 says a misleading action includes the provision of false or misleading information relating to a product or a trader. It also covers an overall presentation designed to deceive, and it treats even true information as misleading if it is presented in a deceptive way. The Act says these provisions came into force on 6 April 2025.

Enforcement sits with local weights and measures authorities, which are the trading standards services, and with the Competition and Markets Authority. On conviction on indictment, an offender faces up to two years in prison or a fine, or both. Law firm Pinsent Masons reports that the CMA can also fine businesses directly, at up to £300,000 or 10% of global annual turnover, whichever is higher, for substantive infringements. Those are ceilings, and most disputes with a customer will not come near them. They do show that a misleading answer from your chatbot is your problem under the Act.

The Canadian tribunal decision on 14 February 2024 in the Air Canada case is not UK law, but it shows how an argument that a chatbot is a separate entity gets treated. The tribunal said the chatbot was part of the airline's website and that the airline was responsible for all information on its website, whether it came from a static page or a chatbot. It awarded damages of about 650 Canadian dollars, plus interest and fees. For a UK merchant, the sensible conclusion is to treat anything an automated tool says on your site as your statement, and to review the answers it gives on delivery, returns and refunds as you would any page of policy.

A chatbot installed from an app store is still your chatbot on your site, and the transcript it leaves is evidence of what was said to the customer. Check how long the tool keeps transcripts and whether you can export them before a complaint arrives, not after.

Three habits follow. Keep policy answers tied to a source page that you maintain, with a date on it. Give the chatbot a way to hand over to a person on delivery and payment questions, where the studies above show customers least want a bot. Read a sample of transcripts every month and fix the answers that are wrong.

## Where BYOM sits

Kina is the AI operator in BYOM. It reads your Shopify catalogue and stock, and the helpdesk you have connected, and drafts the change. Every Action names its approver, and you can approve a change you asked for yourself when your role allows it.

Related: [See Kina](https://byom.co/kina), [See approvals](https://byom.co/approvals), [What an AI operator is](https://byom.co/blog/what-an-ai-operator-actually-is), [What BYOM is](https://byom.co/blog/what-byom-is).

## Sources

- [IBM, AI agents vs AI assistants](https://www.ibm.com/think/topics/ai-agents-vs-ai-assistants)
- [The Conversation, Chatbots are on the rise, but customers still trust human agents more, 2025](https://theconversation.com/chatbots-are-on-the-rise-but-customers-still-trust-human-agents-more-259980)
- [Torkin Manes, BC Tribunal confirms companies remain liable for AI chatbot information, 2024](https://www.torkin.com/insights/publication/bc-tribunal-confirms-companies-remain-liable-for-ai-chatbot-created-information)
- [Shopify Help Center, Shopify Flow](https://help.shopify.com/en/manual/shopify-flow)
- [Shopify Help Center, Shopify Flow triggers reference](https://help.shopify.com/en/manual/shopify-flow/reference/triggers)
- [Zapier, Pricing, read 2 October 2026](https://zapier.com/pricing)
- [Make Help Center, How credits are counted, read 2 October 2026](https://help.make.com/credits)
- [Model Context Protocol, Specification 2025 06 18, Tools](https://modelcontextprotocol.io/specification/2025-06-18/server/tools)
- [legislation.gov.uk, Digital Markets, Competition and Consumers Act 2024, Part 4 Chapter 1](https://www.legislation.gov.uk/ukpga/2024/13/part/4/chapter/1)
- [Pinsent Masons, DMCC Act overhauls UK consumer law enforcement](https://www.pinsentmasons.com/out-law/analysis/dmcc-act-overhauls-uk-consumer-law-enforcement)