guide
Only One AI Support Vendor Publishes a Setting for How Far Its Agent Can Go
Chatfuel ships four autonomy levels and starts every bot on the one where the AI writes nothing.
Chatfuel is a small social-commerce chatbot with no about page, no published founding date and a documentation site written for marketing agencies. It publishes the clearest statement of AI autonomy we have found anywhere in this category.
| Level | Who creates the booking | Can the AI reschedule? |
|---|---|---|
| Intent Collection | A human — the AI only collects what the lead wants | No |
| Booking with Human Approval | A human, after the AI has collected everything | No |
| Booking with Human Review | The AI — a human reviews it afterwards | Yes, then notifies a human |
| Full | The AI, end to end | Yes |
And then the sentence that matters more than the table:
“New bots start at the safest level, Intent Collection, and you move them up as the client's trust (and your setup) earns it.”— academy.chatfuel.com/booking-and-calendar, read 21 August 2026
The default is the level where the AI writes nothing.
Why this is the question, not the capability
Every platform in this directory now advertises an agent that takes action. We have spent months establishing what that phrase actually covers, and the answer is usually less than it sounds.
But capability turns out to be the second question. The first is who decides, and on what default. An agent that can issue a refund and does so unsupervised is a different product from an agent that can issue a refund and queues it for approval, and they are sold with the same sentence.
Almost nobody publishes their answer. Chatfuel publishes a slider, names who creates the record at each setting, gives an honest cons list for each, and states which one you start on.
What Chatfuel gets right, in detail
Three things, and each is separately rare.
It names the actor rather than the capability. Not "the AI can book appointments" but, at each level, whether a human or the AI creates the record. That distinction is the whole thing and most documentation elides it.
It states the default and justifies raising it. Trust is earned by the setup, not granted by the purchase.
It publishes the downside of each level, including the observation that at Human Review a mistake sits in the calendar until somebody looks at it. That is a real operational cost written by the vendor whose product incurs it.
There is also a warning about the interaction that breaks the model: every level below Full ends with the conversation going to a human, so if handoff is switched off, the AI collects everything and has nobody to pass it to. Check that handoff and the team inbox are set up and staffed before you dial autonomy down.
And the honest limit
The write lands in Chatfuel's own calendar. Not Google Calendar, not your scheduler. The autonomy model is excellent and the surface it governs is small, and the documentation page that would cover reaching a CRM reads, in full, coming soon.
Now compare a vendor that cannot say whether its agent acts at all
ChatBot by Text is a twenty-four-year-old public company with audited accounts. Two of its help pages answer the central capability question differently, in the same month.
One page, Custom Skills vs Instructions, opens its list of what a skill can do with "Call external services via API (webhooks support)".
Another page, Custom Skills, says "Below is a list of all the actions your AI agent can perform" and then enumerates twenty of them across customers, chats and tickets. We fetched that page and counted occurrences of the word webhook in all 175KB of it. Zero.
So one page says the agent can call your API, and the page that claims to list every action it can perform does not contain the mechanism. Text's marketing sides with the permissive reading, showing skill templates that call an order-status webhook. The external writes that are unambiguously documented sit in Workflows, which fire on an event rather than being chosen by the agent.
Autonomy settings are a luxury when the underlying question is unresolved.
Where the rest of the directory sits
Most vendors publish something adjacent to autonomy without publishing autonomy itself. The adjacent thing is usually a confidence threshold, which controls when the agent gives up rather than what it is permitted to do.
| Vendor | What is published | Is it an autonomy control? |
|---|---|---|
| Chatfuel | Four booking autonomy levels with a stated default | Yes |
| Zendesk | An adjustable confidence threshold | No — controls answering, not acting |
| Intercom | Escalates when confidence thresholds are not met; procedures can trigger approvals | Partly |
| Gorgias | Actions gated by construction — only runnable inside a skill or guidance | Partly, by design rather than setting |
| NICE | "Configurable approval workflows" | Named, not documented |
| Sierra | Supervisory layers, topic and keyword filters | No |
| Salesforce | Trust Layer toxicity detection, dynamic grounding | No |
| Meta Business Agent | Handoff triggers you cannot configure | No — and explicitly so |
Gorgias's approach deserves credit as the structural version of the same idea: its agent can only run an action when that action is included in a skill or guidance, and its own advice is to keep those minimal to start. That is a default-deny posture achieved through architecture rather than a settings page, and it produces a similar outcome.
Meta sits at the other end. Its agent starts a handoff automatically on signals like low confidence or a customer asking for a human, and Meta states that you do not configure the triggers. Combined with the agent acting as primary responder and your application receiving standby webhooks, that is the least buyer-controlled arrangement here — though it is also, unusually, honest about being so.
The pattern underneath: whose record does it write to
Three of the four platforms we added this month can write something. What separates them is where it lands.
- Manychat writes to its own contact record. Lead capture populates system or custom user fields, and AI Goals end in skip, send a message, trigger an automation, or assign to a person. The HTTP call is a flow block you author, so the AI reaches outside only by triggering an automation that contains one.
- Chatfuel writes to its own calendar, at the two higher autonomy levels only.
- ChatBot by Text writes freely to Text's own records — customers, chats, tickets — with twenty named actions. Whether it can write anywhere else is the question its two help pages answer differently.
- Meta Business Agent writes to your systems, through connectors you author and register with a schema the agent populates itself. Meta ships zero pre-built connectors, so every system it touches is one you built.
Writing to the vendor's own record is not nothing, and for a lot of use cases it is enough. But it is a materially different product from writing to your order system, and the marketing language for both is "takes action".
What to ask, and the order to ask it in
- In your product today, what can the agent change without a human approving it first, and what is the default when we switch it on?
- Where is the setting that changes that, and can we see it in a demo rather than hear about it?
- When the agent creates a record, does it land in your system or ours? Ask for the field, not the feature name.
- If a human has to approve, where does that approval queue live and what happens to it overnight?
- What happens when the agent gets it wrong and the customer has already acted on the answer? No vendor in this category publishes a runbook for this, so the answer will be improvised, and how it is improvised tells you a lot.
The first question is the one to insist on, because the honest answer to it is a setting rather than a paragraph. If a vendor can show you the control, they have thought about this. If they describe their safety philosophy instead, they have not shipped the control.
Quotes above were matched against raw page source on 21 August 2026, including the occurrence counts used to establish absences. Where a vendor's own pages disagree, both are quoted and neither is treated as settled.
Frequently Asked
Can I control how much an AI support agent is allowed to do on its own?
At most vendors, not through a published setting. Chatfuel is the exception, publishing four booking autonomy levels that name who creates the record at each: Intent Collection where a human creates everything, Booking with Human Approval, Booking with Human Review where the AI creates it and a human checks afterwards, and Full. New bots start at the safest level. Most competitors publish a confidence threshold instead, which controls when the agent stops answering rather than what it may do.
What is the default autonomy setting when I turn on an AI agent?
Almost nobody publishes one. Chatfuel states that new bots start at Intent Collection, the level where the AI writes nothing. Gorgias reaches a similar place structurally rather than by setting: its agent can only run an action included in a skill or guidance, and its advice is to keep those minimal to start. Meta goes the other way, making its agent the primary responder with handoff triggers you cannot configure.
Does the AI write to my systems or to the vendor's?
Ask about the field, not the feature. Manychat writes to its own contact record and reaches outside only through an HTTP block you author inside an automation. Chatfuel writes bookings into Chatfuel's own calendar. ChatBot by Text writes to Text's records through twenty named actions. Meta Business Agent writes to your systems, through connectors you build and register yourself, since Meta ships none pre-built.
Why do two of ChatBot by Text's help pages disagree about webhooks?
We do not know, and we report both. The Custom Skills vs Instructions page lists calling external services via API with webhook support as the first thing a skill can do. The Custom Skills page says it lists all the actions the agent can perform, enumerates twenty of them, and contains the word webhook zero times in 175KB of page source. Until it is confirmed in-product, treat the external-write capability as unresolved and test it during a trial.
What should a good answer about agent autonomy sound like?
A setting, shown on screen, with a stated default. If a vendor answers with a safety philosophy, filters, supervisory layers or a trust framework, they have described how the agent decides rather than what it is permitted to do. Those are different things, and only one of them is something you can change on a Tuesday when a refund goes wrong.
Tools Mentioned
Full reviews, pricing tiers and where each one breaks.
Chatfuel
Social-commerce chatbot whose AI books, reschedules and cancels appointments end to end — into its own calendar. Publishes two mutually incompatible price lists on two of its own properties.
ChatBot by Text
Per-seat chatbot from the LiveChat family, mid-rebrand to Text. Reads live Shopify order data; changing a record needs a webhook you host. Auto-refills resolutions at $0.99 when you hit the cap.
ManyChat
Priced per active contact rather than per seat, with a real free tier at 25 contacts. The AI answers and captures; external writes run through an External Request block you author, not through the AI.
WhatsApp Business Platform
Meta's own channel and, since July 2026, its own agent. The only platform here billing AI per token rather than per resolution, and the only one where the vendor is also the channel.
You Can Also Look Into
Half These AI Agents Cannot Touch Anything Until You Build It
Fourteen platforms, read from their own documentation, split by who has to build the thing the agent acts through. It decides your implementation cost.
Which AI Support Tools Can Actually Issue a Refund in Shopify
Four of the six e-commerce platforms in this directory can change something in your store, and two cannot touch an order at all. Here is which, and the question that separates a refund that ships from a refund your engineer builds.
What Is Agentic AI Customer Service? Definition and 3 Tests (2026)
Every vendor now calls their product agentic. Most of them mean a chatbot with a better model behind it. Here is the distinction that actually holds.
Whether Your Customer Can Reach a Human Is a Setting
The consumer press covers the tricks for escaping a bot. Nobody covers the other side, where the escape hatch is a configuration decision somebody at your company made.
SOURCES
WRITTEN BY AR · UPDATED 2026-08-21
I read the fine print. Vendor pricing pages, billing definitions, terms, funding filings and acquisition notices — then I do the arithmetic nobody publishes: what a platform actually costs at your volume, what its headline metric is really counting, and who owns it now. I do not run benchmarks, and no page here pretends otherwise.