Guided actions
The support agent that works your site for the customer.
Most support bots end at a link. Elennai answers from your own content, then opens the right page on your site, fills in what it already knows, and stops at the step that needs a person. You decide which flows exist and where it has to ask first.
What actually happens when it acts.
One question, five steps, and a stop before anything that matters.
The visitor asks in their own words
"I want to send back the jacket I ordered last week." No menu, no keyword, no form to find first.
The agent answers, then offers to do it
It replies from your own returns policy and attaches the page it read, then offers the flow that matches: start a return. The visitor presses it, or keeps reading.
It picks the flow you saved
Only flows in your catalogue can run. The agent matches the request to one of them; if nothing matches well enough, it stays an answer.
It navigates and fills
It opens the returns page and completes the steps on that page: the order number it already has from the conversation, the reason the visitor gave. A visible cursor moves to each field so the visitor can follow every move.
It stops at the step that matters
The confirm button is a sensitive step, so the run stops there and hands control back. The customer presses it themselves, and the conversation carries on underneath.
Recorded, scanned, or written. Never guessed.
Three ways into your catalogue, and one rule that covers all three.
Record it yourself
Open the recorder, do the task once the way a customer would, and each click and field becomes a step. It is the fastest route for anything behind a login, because you are already signed in while you record.
Scan your site
Point Elennai at your site and it crawls the pages, finds the forms and the flows they belong to, groups them, and proposes them with names. You read the list and accept the ones that are real.
Write it by hand
Edit any step, reorder them, mark one sensitive, or build a flow from scratch when you know exactly what it should do. Everything a scan proposes is editable before you save it.
Nothing enters your catalogue without your save. A scan proposes; a recording drafts; you decide what becomes a flow the agent may run.
The model never writes a selector.
This is the line that makes the rest of the page safe to promise. Every step is derived from the real HTML on your page: the element the recorder captured or the scanner found, with several independent ways to locate it again. The language model is used for what language models are good at - reading the page, grouping the steps into a flow and naming it in your words - and it is never asked to produce a target.
- It cannot invent a button. A control that does not exist in your HTML cannot become a step, because steps are built from what was found, not from what was imagined.
- It cannot drift. The saved flow is data, not a prompt. Running it twice does the same thing twice, in the same order.
- It stops rather than guesses. If the page has changed enough that no locator matches with confidence, the run ends and says so. Clicking the wrong thing is worse than not finishing.
Answer. Guide. Act. You decide how far it goes.
The same three rungs and the same four switches the home page promises, stated here in full.
Answer
It replies from your own content and attaches the sources it used, so the answer can be checked. When it has nothing, it says so instead of inventing something.
Guide
It takes the visitor to the right place and shows them what to do there. Nothing is clicked on their behalf at this rung.
Act
It navigates to the right page and completes the steps on that page: filling the fields it already knows and stopping at the step that needs a decision.
Four switches, all yours.
- Off by defaultGuided actions do nothing until you turn them on, one saved flow at a time. A new agent can only answer.
- GatedThe agent can only run flows that are in your catalogue because you recorded them, accepted them from a scan, or wrote them by hand. It has no general permission to use your site.
- SensitiveMark a step sensitive and the run stops there and asks before it goes any further. Anything that costs money or changes an account belongs in this group.
- Hand it to a humanAny flow can end with a person instead, and the conversation arrives in your inbox with the full transcript and what the agent had already done.
It can chain flows you have already approved into a longer one. It cannot invent a new flow on its own.
What it will not do, in writing.
You will meet these in the first week, so you may as well read them now.
A full page reload ends a run
The agent works inside the page the visitor is on. If a step causes your site to reload the document, the run stops there and the agent says what is left to do. Most checkout and account flows in a modern app never reload, which is why this is a limit rather than a wall.
Login-gated pages need the recorder
A scan sees your site the way an anonymous visitor does, so it cannot propose flows that live behind a sign-in. Record those yourself while you are logged in and they work the same as any other flow.
Single-page apps continue across views
If your site is a single-page app, register your router once and the agent can carry a run from one view to the next without a reload. That is the one case where a flow spans more than one screen.
It cannot invent a flow
It can chain flows you have already approved into a longer one. It cannot decide on its own that a new sequence of clicks would be a good idea, and there is no mode that lets it.
The price, and what counts as a run.
Included on Pro and Scale. €99 a month on Starter and Growth.
Every plan that has guided actions includes 250 of them a month. If a month runs long, add a block of 250 more for €49 a month without changing plan. A guided action is counted when the agent starts a run for a visitor, not when it merely mentions that one exists.
The five questions everyone asks.
What exactly is a guided action?
A flow on your own site that the agent can run for a visitor. It navigates to the right page and completes the steps on that page: opening the form, filling in what it already knows from the conversation, and stopping on the step that needs a decision. A visible cursor moves across the page while it works, and the visitor can take over at any moment.
Can it complete a whole journey across several pages?
It navigates to the right page and completes the steps on that page. If your site is a single-page app and you register your router, a run continues across views. A full page reload ends the run, and the agent tells the visitor what is left rather than pretending it finished.
What happens when my site changes?
Each step carries several ways to find the element it needs, so a moved button or a renamed class usually changes nothing. When the page has changed enough that the agent cannot be certain, the run stops instead of clicking something else, and a re-scan proposes the corrected step for you to accept.
Can it act on a page that requires a login?
Yes, if you recorded the flow while signed in. The in-browser recorder captures the steps from the real page, so account and dashboard flows work. A scan cannot reach those pages, because it visits your site as an anonymous visitor.
What does the visitor actually see?
A cursor that moves to each field and control, the field filling in, and a short line of text saying what is happening. Nothing is hidden and nothing happens in a background tab. If the visitor has reduced motion turned on, the cursor jumps between steps instead of gliding, and the run still works the same way.
Record the task you are tired of explaining.
Sign up in under a minute, no card. Connect your content, record one flow, and let the agent finish that job for the next customer who asks.