Which small-business AI agents take actions, ask for approval and keep a log? Six agents compared

Of six AI agents a small business is likely to meet in 2026, only one documents an approval step before it acts. Shopify's help page says Sidekick "presents changes for your review before applying them."¹ Fin and HubSpot Breeze document actions and bill per outcome. Meta's Business Agent "stops answering that chat" once it hands off.⁶ Google's GBP agent is a US pilot with no official documentation.⁸ None of the pages we read documents a per-action receipt or a follow-up schedule.

This comparison reads each vendor's own words, dated, against six tests. Where a vendor page could not be reached, the row says so and names the secondary source used instead.

How we got this: on 1 October 2026 we read the help, pricing and documentation pages listed in the sources for each agent. A cell says "not documented" when the page we read is silent, which is not the same as "does not exist". Vendor features change monthly; this table carries a check date for that reason.

Which AI agents can take actions for a small business?#

An AI agent "takes actions" when it changes something outside the chat window: an order, a product, a calendar slot, a record in another system. Answering a question does not count, however good the answer. On that definition, four of the six agents below document some form of action, and two of them act mostly for someone other than the owner.

The six agents also sit in different places. Shopify Sidekick works for the merchant inside Shopify admin. Fin and HubSpot Breeze answer customers on the business's own chat or help desk. Meta's agent answers WhatsApp, Instagram and Messenger. Google's agent answers messages on a Business Profile. Yelp Assistant acts for the consumer and reaches the business as a request.

That placement matters before any feature comparison. An agent that edits your product catalogue carries a different risk from one that tells a customer your refund policy, and a different risk again from one that books a customer into someone else's calendar.

What counts as acting, approving and logging? The six tests#

The six tests below are the columns of the comparison. Each one is phrased as a question an owner can put to a vendor and check against a written page, not a sales call.

1. Takes actions. Does the agent change state outside the conversation, and does the vendor list which actions?
2. Asks approval. Must a person approve before the change is applied, and can the owner choose which actions need approval?
3. Logs per action. Is there a record of each action showing what was asked, what was done and what it was based on, or only counts in a report?
4. Follows up. Does the agent write again on a schedule when the customer goes quiet, or does it wait to be spoken to?
5. Hands off with a reply-by time. When the agent stops, is the customer told who will answer and by when, and does the human receive the context?
6. Billing unit. What does the vendor charge for, and does a handoff or a non-answer count as a billable outcome?

The distinction between tests 2 and 3 should be kept strict. Approval is a control that operates before the action. A log is evidence that operates after it. An owner needs both, because approval on every action does not scale and a log without approval only tells you what went wrong.

The comparison table: six agents, six tests, checked 1 October 2026#

The table records what each vendor's page says, in the vendor's words where possible. "Not documented" means the page we read is silent on the point.

Actions, approval, logs, follow-up, handoff and billing for six AI agents

Agent	Takes actions	Asks approval	Logs per action	Follows up	Handoff	Billing unit	Verified from
Shopify Sidekick (works for the merchant)	Yes: "analyzing data, managing orders, or editing products"	Yes: "presents changes for your review before applying them"	Not documented	Not applicable (owner-facing)	Not applicable	Listed within Shopify Magic; not priced separately on the page read	Shopify Help Center, primary12
Fin (formerly Intercom)	Yes: "Takes action to update external systems"; data connectors "retrieve or update data"	Customer verification only: a one-time password before sensitive tasks	Report counts: Triggered, Escalated, Handed off, Pending; filter by procedure outcome	Not documented	Configured "Procedure handoff" is billable	$0.99 per outcome, 50/month minimum; qualification $9.99	fin.ai and Intercom Help, primary34
HubSpot Breeze Customer Agent	Yes: "performs an action (like a password reset)"	Not documented in source read	Not documented in source read	Not documented	Resolved if "no handoff to a human within 72 hours"	$0.50 per resolved conversation from 14 Apr 2026	Resolve247, secondary; HubSpot page not reached5
Meta Business Agent	Reported: books appointments, qualifies leads, routes to humans	Not documented	Reported: "daily briefings of overnight chats"	Not documented	"Stops answering that chat"; "loses earlier conversation context" on hand-back	Reported as "included in WhatsApp Business Premium subscription tiers"	360dialog docs, primary; TechCrunch, secondary67
Google GBP AI agent (US pilot)	Not documented beyond answering messages	Not documented	Two tabs: "Handled by your AI agent" and "Needs your attention"	Not documented	"Customer handoff" toggle notifies the owner	No published price; no official Google documentation	Partoo, secondary8
Yelp Assistant (acts for the consumer)	Yes, for the customer: schedules via Vagaro, Zocdoc, Calendly, RepairPal; obtains estimates in 450+ categories	Business side: AI "guidance" for owner-written replies	Not documented	Not documented	Not applicable	Reported: ChatGPT quote requests billed as standard Yelp leads	Yelp, primary; Tygart Media, secondary91011

Read across the rows and two columns stay almost empty: per-action logs and follow-up. Read down the billing column and the units differ in a way that matters more than the prices: outcomes, resolutions, subscription tiers and leads are four different things.

Does Shopify Sidekick make changes without approval?#

Shopify Sidekick does not, according to Shopify. The help page states that Sidekick "presents changes for your review before applying them," and that for longer tasks it "continues working in the background, even if you close the chat or navigate away, and notifies you when it's ready for review."¹ The documented actions include "analyzing data, managing orders, or editing products," and working "with third-party apps to answer questions and complete tasks."

Two limits apply to that answer. First, the page we read does not describe an undo function or a history of applied changes, so an owner who approves a bad edit should not assume a one-click reversal exists. Second, Sidekick is a merchant's assistant. The page contains no statement about Sidekick talking to customers, so its approval model tells you nothing about how a customer-facing agent on your store behaves.

The review-before-apply model is the one owners say they want. In Bluehost's 2026 survey of 350 small businesses, 16% had deployed AI agents and 6% highly trusted AI with their brand voice; Bluehost reports that "owners prioritize approval mechanisms before AI directly engages customers."¹³

What happens after Meta's, Google's and Yelp's agents hand off?#

The platform agents answer on channels the platform owns, and their handoff behaviour is where most of the owner's work begins. Meta's is the best documented. The 360dialog documentation states: "Once the agent hands off to a human, it stops answering that chat," and "On hand-back, the agent loses earlier conversation context."⁶ The same page notes the handoff message is fixed and cannot adapt to the customer's language.

Google's pilot is less documented than Meta's. Partoo reports a "Customer handoff" toggle that notifies the owner when the AI passes a conversation on, a "Needs your attention" tab for the passed conversations, and no official Google documentation, no tone controls, unknown data sources and no stated way to correct answers.⁸ Neither vendor documents a promise to the customer about when a person will reply.

The handoff is the step most likely to fail. Pied Piper's 2025 service-telephone study measured a 56% failure rate when dealership AI tried to transfer callers to a human.¹⁴ That is a phone figure from car dealers, not a written-channel one, but it shows where the risk sits.

Yelp is a different case. Yelp Assistant acts for the consumer: it can schedule through Vagaro, Zocdoc, Calendly and RepairPal and "obtain estimates from professionals across 450+ service categories."⁹ For the business, the action arrives as a request in the Yelp inbox, and Tygart Media reports that quote requests started in ChatGPT are billed as standard Yelp leads "with no source tag."¹¹ The owner's side gets "real-time AI-powered guidance" to write better replies, which the owner still sends.¹⁰

Fin vs HubSpot Breeze: what counts as done, and what you pay for#

Fin and HubSpot Breeze both bill per outcome, and the definition of the outcome is the term an owner should read before the price. Fin's pricing page lists five outcomes. A "Resolution" occurs when "No further help is requested after Fin's last answer." A "Procedure handoff" occurs when "Fin completes a Procedure you've configured to hand off to a human." Each costs $0.99, with "50 outcomes per month minimum"; a qualification costs $9.99. The page also states: "You're not charged when a conversation is simply passed to your team without an outcome."³

HubSpot's definition, as reported by Resolve247, is time-based. A conversation is resolved when "The agent shares a content source (like a knowledge base article) or performs an action (like a password reset), and there's no handoff to a human within 72 hours of the agent's final message," or when a lead is marked qualified, partially qualified or not qualified. The price is "$0.50 per resolved conversation," effective 14 April 2026. Resolve247 also reports that a reply after 72 hours reopens the conversation and "the 72-hour clock resets."⁵ HubSpot's own pricing page could not be reached, so these terms should be confirmed in your HubSpot contract.

On approvals, Fin documents a customer-side control: "Fin can send a one-time password to the customer's email and require them to enter it before performing the task."⁴ That verifies the customer. It is not an owner approval tier, and the source we read for HubSpot does not describe one either.

Why the per-action log matters when an agent gets it wrong#

A log is the evidence that remains when an agent's statement is disputed. In Moffatt v. Air Canada (2024 BCCRT 149), the British Columbia Civil Resolution Tribunal rejected the airline's argument that its chatbot was a separate legal entity and ordered it to pay CAD $812.02.¹² That is one Canadian small-claims decision, and it binds no court elsewhere. This is not legal advice; check with counsel for your jurisdiction.

The practical point does not depend on jurisdiction. If a customer says your agent promised a refund, you need to see what was asked, what the agent answered, which policy or document it relied on, and when. Counts in a report ("Handled by your AI agent", "Escalated") tell you volume. They do not let you reconstruct one conversation's reasoning. That per-action record is what we call an action receipt; our 12 questions for an AI vendor include it.

Questions to put to any vendor before you switch an agent on#

• Which actions can the agent take, listed by name, and which can I restrict?
• Which actions require my approval before they apply, and can I set that per topic?
• For a single conversation, can I see the question, the answer, the source used and the time, and can I correct it?
• Does the agent follow up when the customer goes quiet, and how many times?
• When it hands off, what is the customer told about when a person will reply, and what context do I receive?
• Is a handoff, an abandoned conversation or a non-answer billable under your definition of an outcome?

What we could not verify#

Four points remain open as of 1 October 2026. HubSpot's own Breeze pricing page was not reachable, so its row relies on Resolve247. Meta's pricing is reported by TechCrunch, not a Meta pricing page. Google has published no documentation for the GBP agent. Shopify's page does not say whether Sidekick keeps a change history. We will update the table when primary pages answer these.

Where Valet fits#

Valet is an AI assistant for a small business's written inquiries: email, website forms, marketplace lead emails and requests on its free Valet page. Against the six tests it works this way. It answers only from the owner's own facts and offers, holds and books real calendar times, touching only events it created. Each topic is set to Autopilot, Ask me first or Always you, and money, complaints and contracts go to the owner with one-tap replies while the customer is told when the owner will reply. Every reply gets a Valet receipt: what was asked, the rule used, the sources used, where it was sent from. The Valet follow-up cap is two follow-ups by default. Pricing is a small base plus a fee only for bookings or quotes Valet wins. To see where your own inquiries stall today, start with the free lead check: forward 10 to 20 recent inquiries and see which went unanswered and which stopped after one reply.