# Edit Your AI with Assembly (/docs/guides/campaign-toolkit/assembly)
Voice AI campaigns and your Receptionist have an **Assembly** tab (it used to
be called Employees). It's where you talk to careCycle about your AI employee: how it handles calls, what it says, and what you want it to
say instead. You don't need to touch a prompt or know anything about how AI
works. You just chat, test, and approve.
This guide covers how the tab works and, more importantly, how to actually get
the results you're looking for.
How your AI "thinks" about a call [#how-your-ai-thinks-about-a-call]
Before you change anything, it helps to know what's under the hood.
Behind every AI employee is basically an **exemplary script**. That's the happy
path: how the call should go when everything goes right. On top of that, it
knows how to handle the stuff people say most often on these calls: "I'm busy,"
"Is this a robot?", "I already have an agent," "How'd you get my number?", and
so on.
Then there are **escalations**. These are the moments where the AI needs to
*do* something: transfer the call to your team, book an appointment, or log an
inquiry for someone to follow up on. The script decides what the AI *says*; the
escalations decide what it *does* about it. More on those
[below](#escalations).
And on every call, careCycle hands it what your records say about the member:
their policies, appointments, open requests, and recent calls. See [What Your
AI Knows on Every Call](/docs/guides/concepts/what-your-ai-knows).
Step 1: Onboard yourself to the campaign [#step-1-onboard-yourself-to-the-campaign]
Open your campaign and click the **Assembly** tab. You'll land in a chat with
careCycle.
The best first move is to just ask questions. Treat it like getting onboarded
by someone who already knows the campaign inside and out:
* "Walk me through the happy path on this call."
* "What does it say if someone's busy?"
* "How does it handle someone who says they already have an agent?"
* "What happens if nobody's available to take a transfer?"
Starting prompts [#starting-prompts]
Not sure what to ask? A new chat offers a few starting prompts. Click one to
send it:
* **Show the happy path**
* **Review recent calls**
* **Improve the opening**
* **Explain how this employee handles a situation**
* **Make responses shorter**
After each answer, careCycle suggests a few follow-ups to keep exploring.
Reviewing calls [#reviewing-calls]
Once your campaign has been running, careCycle can pull up real calls,
transcripts, and results to explain what's working and what isn't. It looks
at the last 7 days unless you ask for something else, like "yesterday," "last
month," "the last 48 hours," or "since we launched."
careCycle explains the AI's instructions in plain English instead of showing
you the raw prompt. Greetings and voicemails are the exception: you can see and
edit those word for word.
Step 2: Tell it what you want [#step-2-tell-it-what-you-want]
Once you get the gist of how the AI works, now's your chance to tweak things.
If you're starting from scratch [#if-youre-starting-from-scratch]
Start with the context, not the words. Tell careCycle:
* **Who** you're calling and **when** (where are they in their journey with
you?).
* **Why** you're calling them.
* What the **call to action** is, if there is one (transfer, appointment,
just a check-in).
* How you want it to **sound**.
If you already have a script in mind [#if-you-already-have-a-script-in-mind]
The best results come from a mix: give the general context (who, why, when),
then paste in a script that paints a picture of what you're going for.
Here's the thing about scripts though. Think about the last time you called
your bank or your cell provider. You can usually tell when the person is
reading off a script versus actually talking to you. Most of the good ones
still have a script, they just don't sound like it. Same with cold-call sales:
you get a script, but the best calls are the ones that wander off it.
AI is different. If you give it a script, it'll say that script almost word
for word. It's a robot, after all. So a script that sounds a little stiff when
a human reads it sounds *way* stiffer coming out of an AI.
The good news is we've spent a ton of time on exactly this problem. careCycle
is trained on our best practices for voice AI scripts. It won't ignore what
you ask for or rewrite your whole script, but it will push back on lines that
work between two humans and fall flat through text-to-speech.
A few tips for writing lines that sound human [#a-few-tips-for-writing-lines-that-sound-human]
* **Write it how you'd say it.** Ums, uhs, and repeated phrases are your
friends: "I wanted to uh, I just wanted to see if you had a couple minutes."
* **End most turns with a question.** On an outbound call, you called them, so
you're driving. A question gives them a reason to respond. *Most* turns,
though, not every one. A question after every single line gets annoying and
sounds robotic.
* **Think in phases.** A call has steps: the opener, the reason for the call,
the ask, handling a "no," and the close. Big changes to how those phases flow
are what the main chat is great at.
Approving changes [#approving-changes]
When careCycle has a change ready, you'll see a **Proposed Employee Change**
card. It tells you what it's changing, why, and which calls it applies to (for
example, daytime calls, after-hours calls, or appointment callbacks), with
the **Current** and **Proposed** versions side by side. If something has to
happen first, like setting up an escalation, the card says so under **Before
this can be applied**.
Not quite right? Just keep chatting. Tell it what to change and it'll come back
with a new proposal.
When you're happy, click **Apply Change**.
**Apply Change** goes live right away. If your campaign is active, the very
next call uses the new version. If you want to try it out first, click **Test
Proposed Changes** to run a simulation against the proposal before it touches
anything.
Step 3: Test it in a simulation [#step-3-test-it-in-a-simulation]
Once you've gone back and forth a few times, you'll want to make sure the AI is
actually following the script and flow you have in mind. That's what the
simulator is for.
Click **Simulation** in the top right. You can pick a **Call Scenario** (like
**Existing Customer**, **New Customer**, or **After Hours (Existing
Customer)**) and a **Customer** to play, or leave both alone to use a sample
customer. Then click **Start Simulation**.
You play the caller and type back and forth with the AI. It uses the same
instructions and escalations your real calls use.
The simulation is text, not voice, and it never actually transfers anyone,
books anything, or logs an inquiry. When the AI tries to, you'll see a **Test
Action** badge showing what would have happened. A transfer or a hang-up ends
the simulation. If you picked a **Customer**, their details (like their name,
agent, and plan) fill in the script, but the AI doesn't get their policies,
appointments, open requests, or recent calls. Use a [test
call](/docs/guides/getting-set-up/test-your-campaign) to hear it use those. See
[What Your AI Knows on Every Call](/docs/guides/concepts/what-your-ai-knows).
Step 4: Coach it line by line [#step-4-coach-it-line-by-line]
Think of the simulator as your space for testing, verifying, and then making
surgical edits to the exact phrasing. Say you really like the script, but you
want to add an "uh" in the middle of a long turn to break it up. This is where
you'd do that.
Under every AI response you'll see three buttons: **Good**, **Bad**, and
**Improve**. **Good** saves right away. Pick **Bad** or **Improve** and you
get:
* **Coaching Note**: free-text guidance. What's wrong, or what you want.
* **Suggested Response**: open it to write the exact line you wish it had
said. Turn on **Use Exact Wording** if it should say your line word for
word.
Here's how I use them:
* **Small wording tweaks** (same message, just swapping some words or sounds):
I just rewrite the line in **Suggested Response**.
* **The direction is off and I don't have a perfect line in mind**: I write my
thoughts in **Coaching Note** and let the AI figure out the fix.
* **I know exactly what I want, and it applies beyond this one line** (like
"change this line to X, and fix any other lines like it"): I use both.
Mark the lines you love as **Good** too, so careCycle knows to keep them.
Submit your coaching [#submit-your-coaching]
Keep going until the end of the conversation, or stop whenever you've seen
enough. Then click **Submit Coaching**. That ends the test and sends all your
notes to the main chat, where careCycle turns them into a proposed change.
Review the proposal, push back if you need to, and click **Apply Change** when
it looks right.
Retest your fixes [#retest-your-fixes]
Click **Retest These Fixes** to replay the same conversation (same caller
lines, same customer) against the fixed version. careCycle shows you the
before and after for each line you coached, so you can see whether the fix
landed without retyping the whole call.
It takes a few reps (and that's normal) [#it-takes-a-few-reps-and-thats-normal]
AI isn't perfectly predictable. You can coach a line, apply the fix, re-run the
sim, and still get something you don't love. That's ok! Give the feedback
again and let it take another swing. Good voice AI scripts are built through a
lot of back and forth. The more reps and edge cases you throw at it, the better
it'll sound on the phone.
Campaign briefs [#campaign-briefs]
You can also build a whole new Voice AI campaign by chatting. Click **Create
Campaign**, choose **Voice AI**, then **Start with Assembly**.
1. **Describe the call.** Tell careCycle what the campaign should accomplish,
or click **Attach a Script** to start from one you already have. Starting
ideas like **Help New Callers** and **Check In After Enrollment** fill in
the box for you.
2. **Answer a few questions.** careCycle asks what it needs to know, like the
insurance type and when the calls should happen.
3. **Review the brief.** As you go, it writes a **Campaign Brief**: the goal,
who you're calling, how the conversation goes, the escalations, the voice
and tone, and what success means. **What Will Be Created** shows exactly
what you'll get, and **Still to Resolve** lists anything left open.
4. **Create it.** Scroll through the brief and click **Create Campaign**.
careCycle builds it while you watch, then opens it in Assembly with the
Escalations panel ready to set up. It's created as a draft, so nothing
calls anyone until you activate it.
Resuming drafts and conversations [#resuming-drafts-and-conversations]
You don't have to finish in one sitting.
* **Briefs:** click **Save and Come Back Later**. Your unfinished briefs wait
under **Saved Campaign Briefs** on the Campaigns page. Click **Continue** to
pick up where you left off.
* **Assembly chats:** open **History** in the top right of the Assembly tab
and pick any chat to keep going, or click **New Chat** for a fresh one.
Start from a template [#start-from-a-template]
Seriously, use our templates. They're tried and tested scripts that have run
more calls than any one agency could, and we keep updating them with what we
learn. The best play is to pick the template closest to the kind of call you
want to make, then tweak its script to match. That way you're not starting
from zero and finding every edge case on your own.
Escalations [#escalations]
In the **Assembly** tab, click **Escalations** in the top right to set what the
AI does when a situation comes up on a call. Each row is a scenario, with one
setting for **Business Hours** and one for **Outside Operating Hours**:
* **Transfer Call**: send them to one of your [routing
groups](/docs/guides/campaign-toolkit/routing).
* **Book Appointment**: book them on one of your
[calendars](/docs/guides/campaign-toolkit/calendars).
* **Log Inquiry**: write it up for your team to follow up on. See [Handle
Inquiries](/docs/guides/answer-every-call/handle-inquiries).
After hours never offers a transfer, since nobody's there to pick up. Use
**Manage** in any menu to create or edit routing groups and calendars without
leaving the page. Click **Add Escalation** for a new row, and **Save Changes**
at the bottom of the panel when you're done.
Saving escalations marks any proposals you haven't applied yet as outdated.
Apply or finish up those first, then save your escalations.
See [Escalations](/docs/guides/campaign-toolkit/assistant-behaviors) for the
full rundown.
Bulk-assign a scenario across campaigns [#bulk-assign-a-scenario-across-campaigns]
To set what the AI does for a scenario in many campaigns at once, go to
**Settings > Calling > Scenarios**. Under **Default Escalations**, click a
scenario, pick the action for business hours and after hours, and choose
which campaigns to update (or **Select All**). Click **Save and Update \[number]
Campaigns**. New campaigns start with these defaults too, and any campaign
can still override them in its own Escalations panel.
Version history [#version-history]
Every change you apply creates a new version. Open **History**, then
**Versions** to see who changed what and when. If a change made things worse,
click **Restore Version** on the previous one. Restoring creates a new version,
so nothing gets lost.
Good to know [#good-to-know]
* **Your chats and simulations are private to you.** Teammates won't see them,
but they'll see the versions you apply.
* **You need edit access** to the campaign to apply changes, restore versions,
or save escalations. Anyone who can view the campaign can chat and run
simulations.
* **One change can cover more than one call type.** Check **Applies to** on
the proposal so you know if it's touching daytime, after-hours, and callback
calls all at once.
Where Transfer Call sends people.
The calendars Book Appointment uses.
Hear it live before you turn it on.