Mike HeapMike is an experienced Product Manager who focuses on all the “non-development” areas of My AskAI, from finance and customer success to product design, copywriting, testing and more.
Table of Contents
- What do you need before you start?
- What are the four ways to add AI to your existing helpdesk, and which should you pick?
- How to add a third-party AI agent to your helpdesk, step by step
- Step 1: Check your admin access and what the app will ask for
- Step 2: Install the agent from your helpdesk's marketplace
- Step 3: Point it at your knowledge
- Step 4: Scope which tickets it sees, and pause the AI you already have
- Step 5: Run it in internal-note mode first
- Step 6: Set the handoff before you let it reply
- Step 7: Turn on direct replies for one ticket type
- How to test it before you let it near customers
- What breaks, and how do you tell?
- What should you do next?
- FAQs
Created time
Sep 7, 2026 02:59 PM
Title length (<60)
Author
Last optimised
Ecomm?
Tags
Image
Publish date
Aug 15, 2026
Video
Slug
how-to-add-ai-to-existing-helpdesk
Featured
Type
Article
Ready to Publish
Ready to Publish
By the end of this you'll have a third-party AI agent answering real tickets inside the helpdesk you already run, live in about an afternoon, with your macros, tags and routing untouched.
You add AI to an existing helpdesk by installing a third-party agent from your helpdesk's own app marketplace, pointing it at your help center and your historic tickets, then scoping it to one class of ticket. Nothing migrates: the macros, tags and routing you already built keep working, and the agent installs as an approved app inside the desk you already pay for. Most teams are live in an afternoon with no developer, and the real trade-off is that you now run two vendors (you pay your helpdesk for the desk and the agent for the tickets).
You own the inbox. You built the macros, the tags and the routing rules that keep it moving, and you know exactly which of them will break if someone touches them.
You've also already looked at whatever AI your helpdesk sells you. I suspect you want a second opinion before you commit to it.
The job is wiring a third-party AI agent into the helpdesk you already run, so it answers tickets. Running one particular desk and want the exact click path? Zendesk, Intercom, Freshdesk, Gorgias and HubSpot each have their own walkthrough.
The whole thing is an afternoon's work, and the only hard gate is admin rights in your own desk. You won't need an engineer (you will need an admin, and there is no way round that one).
What do you need before you start?
TL;DR: Admin access in your helpdesk, a published help center (or your last 5,000 tickets if you don't have one), and about an afternoon. No developer for the route most teams take.
- Permissions: admin, or the role your helpdesk lets you hand the job to. Every desk calls it something different, and buries it somewhere you'd have to hunt for. Zendesk calls it "Manage Zendesk Support apps" and switches it off by default on every custom role. Intercom calls it "Can install, configure and delete apps", under Apps and integrations. Gorgias gates installing an app at "Admin roles". Freshdesk hands it to Administrators, and HubSpot asks for Super Admin or App Marketplace access. Our walkthrough for your own desk carries the exact click path.
- Plan tier: on Zendesk, admins and billing admins keep full access to Support app management and nobody can take it off them. The delegable version of that right, the custom-role permission above, is Enterprise only and switched off by default on every custom role, so you only need Enterprise if you want to hand the job to someone who isn't an admin. Gorgias documents app installs on all plans. What your helpdesk charges for its own AI is a different question, and we get to it next.
- Data prep: a published help center. No help center yet? Train on Historic Tickets drafts starter knowledge from your last 5,000 historic tickets instead.
- What you can skip: rewriting the knowledge base first. It will eat a week, and you don't need it to start. The gaps surface from real tickets afterwards, which is quicker and a lot less depressing.
If your helpdesk's own AI is already answering customers, plan to pause it. Two agents replying to one ticket is the single most common thing I see go wrong here, and Step 4 is where you deal with it.
What are the four ways to add AI to your existing helpdesk, and which should you pick?
TL;DR: Four routes: the AI your helpdesk already sells, a third-party agent drafting internal notes, the same agent replying directly, or moving helpdesk entirely. Most teams start with the internal notes and stay.
Route | Best for | Setup time | Needs a developer? | Trade-off |
A. Your helpdesk's own AI add-on | Teams already on the tier that includes it | Hours | No | You're billed for the AI's usage, so the bill climbs as the agent does more |
B. A third-party agent in internal-note mode | Proving it before a customer sees anything | About an hour | No | It drafts for an agent to send, so handle time drops while every ticket still reaches your team |
C. A third-party agent replying directly | Most teams, once B has run for a week | About an afternoon | No | You work inside the app's model of routing |
D. Move to a helpdesk with AI built in | Teams whose helpdesk is the actual problem | Weeks to months | No | You rebuild the macros, tags and routing you already have |
Route A is the AI your helpdesk already sells you: Zendesk AI, Intercom Fin, Freshdesk Freddy, Gorgias AI Agent and HubSpot Breeze. On four of those five the agent is included in the plan you already hold, which is not the same as it running (Gorgias puts an "Add AI Agent" control right on its pricing page). On three of them, Zendesk, Intercom and Gorgias, you're then billed only for the conversations it resolves; Freshdesk charges per session whether it resolves anything or not.
Zendesk includes AI agents in every Suite and Support plan and bills per automated resolution against a per-plan allowance. Intercom includes Fin on all three plans at $0.99 per outcome, and will sell it on top of a helpdesk you already run (no seats needed).
Gorgias includes its AI Agent on every plan with an allowance, then charges $1.50 per automated interaction past it. Freshdesk gives you 500 Freddy AI Agent sessions once, to try it, then sells more at $49 per 100.
HubSpot is the one that asks you to move plan. Its product catalog puts the Breeze customer agent on the Professional or Enterprise edition, running on HubSpot Credits.
Route A is a reasonable start if you're already sitting on the tier that includes it. The catch shows up later, once the agent is handling enough for the bill to be noticeable (Zendesk and Gorgias both include an allowance first, so it stays invisible early on).
Route B is the boring-but-effective option: a third-party agent running in internal-note mode. It drafts a reply on every ticket and writes it as a private note.
Your agents send it, edit it or bin it, and the customer never sees a thing. It's the cheapest way I know to find out whether an agent can answer your tickets.
Route C is the same agent, replying to customers directly, scoped to the ticket types you choose. This is where most teams land after a week on Route B.
Route D is moving to a helpdesk that has AI built into it. If your desk is the actual problem, that's a real answer and I'm not going to talk you out of it.
Our roundup of the best AI help desk software lays that field out properly. Go there if you're picking a new platform.
If you're not sure, start with Route B and give it a week. You get the write path into the ticket from day one, and nothing reaches a customer until you say so.
Teams treat "switch to an all-in-one platform" and "add AI to the stack I already run" as one decision. Collapsing them turns a one-afternoon job into a quarter-long project.
How to add a third-party AI agent to your helpdesk, step by step
TL;DR: Seven steps. Check your access, install from your helpdesk's marketplace, connect your knowledge, scope what it sees, run it as internal notes, set the handoff, then let it reply on one ticket type.
Pick the agent before Step 1. It needs a real write path into the ticket, tagging, a handover it can trigger, and an internal-note mode (that last one is what lets you run it for a week without a customer seeing anything).
Step 1: Check your admin access and what the app will ask for
Start in your helpdesk's admin area. Any marketplace agent asks for the same two things: read and write on tickets or conversations, and an identity to post as.
Check what your role actually allows before you install anything. Where that setting lives depends on the desk: Freshdesk keeps it at "Admin → Team → Roles", Intercom at "Settings → Workspace → Teammates". Our walkthrough for your desk covers the rest.
If you don't have the permission, ask someone who does. HubSpot documents a request-an-install flow for exactly this: you pick a user who does have it, and they get an email naming the app and what it's asking for. Most desks have some version of that conversation, even if it happens over Slack.
Either way, you're done with this step when you know whether you can install an app yourself or need to ask someone who can.
Step 2: Install the agent from your helpdesk's marketplace
Still in your helpdesk, this time in its app marketplace. Install the approved app from the marketplace itself.
On Zendesk, installed apps live at "Admin Center → Apps and integrations → Apps → Zendesk Support apps"; on Gorgias, the app store sits at "Settings → Apps → App Store". Every desk puts its marketplace somewhere different.
Marketplace apps go through the marketplace and helpdesk approval process, so they're secure and your helpdesk knows the app is there. They're more stable too, and less likely to be taken down.
We have an approved app in each helpdesk we support, and you can see the AI Support Agent by My AskAI listing in the Gorgias app directory if you want to know what one looks like from the outside. Setup runs 10 to 15 minutes, with no code and no developer.
Helpdesk | The step-by-step guide |
Zendesk | |
Intercom | |
Freshdesk | |
Freshchat | Covered by the Freshdesk walkthrough |
Gorgias | |
HubSpot |
You're done when the app is installed and connected, and you can see it in your helpdesk.
Step 3: Point it at your knowledge
This step happens outside your helpdesk, in the AI tool's own console. Connect three things, in this order: your help center, then your macros and saved replies as exact answer pairs, then your historic tickets.
The help center is already written, the macros are the answers your team already agreed on, and the ticket history is where the long tail lives (all the questions nobody thought to document).
In My AskAI that's the Knowledge view. It reads the Zendesk, Intercom and Freshdesk help centers, including the articles sitting behind a login.
It also reads Google Drive, Notion, Confluence, SharePoint, OneDrive, Dropbox, Salesforce and Shopify. Macros land as Custom Answers, which the agent gives back word for word.
Train on Historic Tickets drafts knowledge from your last 5,000 historic tickets by default, and we'll go higher on request if you have the volume.
If you want it answering "where is my order?", the User Data API connects your own backend so the agent can read the live record and answer from it. Our Train on Historic Tickets feature page covers the setup for teams with no help center worth connecting yet.
You're done when the agent can quote your own help center back at you in a test question.
Step 4: Scope which tickets it sees, and pause the AI you already have
Back in your helpdesk. Give the agent a narrow slice first: one channel, or one class of repeat question.
Scope it with one control. Three controls means three places to look when it misfires (and it will misfire in week one).
Then check nothing else is answering. If your helpdesk's own AI is live, pause it.
The native agent is usually switched on in more than one place, and turning it off where you configured it doesn't turn it off everywhere. One operator posted this on the Intercom community:
"We are not using Fin anywhere in our product and are not currently interested in implementing it. It's never been integrated into a workflow and I have never manually turned it on." Intercom community thread
Intercom's answer named a "Simple deploy" surface where it was still switched on (the operator had never turned it on there deliberately). On Intercom specifically, the full check also means "Settings → Channels → Messenger → Conversations → With Fin". That control is separate from the deploy surface.
Your desk will have its own version of that split, so check every surface it exposes before you go further.
In My AskAI, our AI Tagging doubles as the scoping control. It reads what the customer wrote and classifies the ticket against your existing helpdesk tags.
Each tag carries its own reply setting: the agent can reply to everything under that tag, give no reply, or be blocked from replying. It runs natively inside Zendesk, Intercom, Freshdesk, Freshchat and Gorgias, and it's billed at $0.05 per attribute per ticket.
Edel Optics uses it exactly this way, blocking our agent from replying on tag categories like faulty items.
You're done when only the tickets you chose reach the agent, and nothing else is answering them.
Step 5: Run it in internal-note mode first
Still in the AI tool, even though what it produces turns up in your helpdesk. Turn on notes mode: the drafts land as internal notes and the customer sees nothing.
Run it side by side with whatever answers today for a week, and read the drafts. Plenty of teams never move past this, and I count it a good place to stop: an agent that drafts and a human who approves is a setup in its own right.
In My AskAI, Internal Notes mode is the default on install, so this is where you land anyway. Your agents also get the same drafts and lookups through the AI Copilot Chrome Extension, which works inside any of our integrated helpdesks.
Our explainer on what an AI copilot is covers the difference between the two if you want the background.
YouGarden ran our agent in Freshdesk notes mode for a month before flipping it to direct replies. Edel Optics started the same way on Zendesk.
By the end of the week every new ticket should carry a draft note, and no customer should have received anything.
Step 6: Set the handoff before you let it reply
Same console as Step 5. Decide when the agent stops, and write down both triggers: the customer asks for a person, and the agent isn't confident.
Then look at what the human picks up, because I see this half skipped more than anything else in the setup. They should get a summary of the conversation so far, and the agent shouldn't reply again until someone hands the ticket back to it.
In My AskAI those rules are Guidance, written in plain English, and the type you want here is Handover & Escalation. Keep each entry under 75 words.
When the handoff fires, the conversation is summarized onto the ticket as an internal note and a person takes over inside the same helpdesk. Edel Optics set up an "I don't know" handover so unanswerable tickets go straight to a human. Our Guidance feature page has the rule syntax.
You're done when you've asked for a human in a test conversation and a summarized note lands on the ticket.
Step 7: Turn on direct replies for one ticket type
Back to the AI tool, with your Step 4 scoping still in place in the helpdesk. Pick the narrowest, highest-volume repeat you have: the one your team could answer in their sleep.
Let the agent reply on that alone, watch it for a week, then widen one class at a time. Going live on everything at once is how CSAT drops and the project stops.
In My AskAI you flip from notes to direct replies, and our per-tag reply control from Step 4 keeps the scope tight while you widen. Self-Learning then drafts new knowledge by comparing the agent's reply to what your team actually sent after a handover, so the resolution rate climbs without engineering work.
You're done when real customers are getting real answers on exactly one ticket type, and you can name it.
How to test it before you let it near customers
TL;DR: Four checks, each with a pass mark. Don't turn on direct replies until the handoff one passes.
- Ask it something only your help center knows. Pass: it answers from your own content, and you can see which source it used. In My AskAI you ask Echo why the agent gave that answer and where it got it from.
- Ask it something nobody has documented. Pass: it says it doesn't know and hands over, rather than inventing an answer. This is the check I care about most.
- Ask for a person. Pass: it stops, hands over, and the summary is on the ticket before your agent opens it.
- Read a week of note-mode drafts against what your team actually sent. Pass: you'd have sent the draft, or edited one line, on the large majority of them.
The go-live gate is check 3 passing and check 4 coming back clean, and nothing should touch a real customer until both do.
Once it's live, the numbers do the watching for you. My AskAI's Insights scores 100% of conversations for AI CSAT.
A drop shows up in days, so you can act on it inside the same week. Our posts on AI resolution rate benchmarks and on containment versus deflection versus resolution cover what those numbers should look like.
What breaks, and how do you tell?
TL;DR: The most common failure is two agents answering one ticket, yours and the one your helpdesk was already running.
When someone tells me the agent answered wrongly, the information it read is almost always the cause. Out of date, ambiguous, wrong or conflicting: an article that says one thing and a macro that says another.
We made the same point on LinkedIn under the line most "AI hallucinations" aren't hallucinations, they're your own docs.
Symptom | What's actually happening | How to tell, and what to do |
Two replies land on one ticket, and one is from an agent you thought was off | Your helpdesk's own AI is still live on a surface nobody checked | Check every surface the desk exposes: chat and email deploy separately, and a quick-deploy path is separate again from a workflow. Then re-test on a real ticket |
The agent's own reply sets your automations off again | The AI's write is a ticket update, so any rule that fires on "ticket is updated" fires on the agent's reply | Stamp the agent's writes with a tag and exclude that tag in the rule's conditions, so the rule fires on customer updates only |
The agent never picks the ticket up at all | The routing rule that was supposed to hand it over isn't matching | Re-read the rule's conditions against a real ticket, then check rule order. An earlier rule can close, tag or reassign the ticket before yours ever runs |
Replies stop overnight and nobody changed anything | The connection's access has lapsed, or another admin removed the app | Work out which of the two it is before you touch anything, then check the app is still installed |
It answers confidently and the answer is wrong | The knowledge it read is out of date, ambiguous or contradicted somewhere else | Trace the answer back to the source it used, then fix the source. That fixes it for every future ticket |
The vendors document two of these themselves. Zendesk describes what happens when its own automations and an outside tool write to the same ticket at once: "Errors are likely when a ticket gets updated by business rules and the API at the same time. This creates a situation called a race condition."
Every ticket create or update re-runs your business rules in that same cycle, so a rule and an outside write can land on the same ticket at once.
Intercom publishes its own article on assignment workflows not firing. The third row happens often enough that the vendor wrote a page about it (rule order is the usual culprit).
On the fourth row the two causes look identical from the outside and need opposite fixes. Zendesk's own documentation draws the line between them. Either your helpdesk can no longer tell who is calling, which means the connection itself has dropped, or it knows exactly who is calling and is refusing them, which means someone changed a permission.
Read the error your helpdesk gives you before you start reconnecting anything.
What should you do next?
TL;DR: Widen one ticket type at a time, and let the knowledge maintain itself.
Widen the scope weekly, one class of ticket at a time. Read the Self-Learning drafts as they arrive and approve the ones that are right.
Only then go looking at the actions that need live data: order lookups, refunds, address changes (the User Data API from Step 3 is what feeds them).
If you'd rather see what the integration does before you install it, our helpdesk integrations page is where to start.
Read it even if you're happy where you are. The trained agent, the custom answers, the guidance and the tasks all come with you if you ever change helpdesk.
You can test the whole thing first, too, on our 30-day free trial: every feature unlocked, no card.
And when you want the menu-by-menu version in your own desk, the Zendesk, Intercom, Freshdesk, Gorgias and HubSpot walkthroughs each carry it end to end.
FAQs
Do I have to migrate off my helpdesk to use an AI agent?
No. A third-party AI agent installs into the helpdesk you already run as an approved marketplace app, and your macros, tags and routing rules stay exactly as they are. Edel Optics layered our agent onto Zendesk after its own AI underwhelmed them, and YouGarden did the same on Freshdesk after a tender that included Freddy AI and Intercom Fin, and neither of them moved helpdesk.
Should I use my helpdesk's built-in AI or a third-party agent?
If you're already on the tier that includes it and your volumes are modest, your helpdesk's own AI is a reasonable place to start. The native add-ons are usually charged per successful outcome, the way Zendesk, Intercom and Gorgias all price theirs. Ours charges per ticket.
Will this break my existing macros, tags and routing?
No. Installing a marketplace app doesn't edit your rules, and scoping is additive: you add a condition, a tag or a view, and what's already there keeps working.
There is one interaction to plan for. A rule that fires on "ticket is updated" will fire on the agent's own reply, so tag the agent's writes and exclude that tag in the rule's conditions.
Can I roll this back if it doesn't work?
Yes, and at three levels. I reach for the softest first: switching the agent back to internal-note mode, which stops anything reaching a customer while the agent stays running and trained.
Next is disconnecting the integration. Last is uninstalling the app from your helpdesk's app settings, and on some desks you'll also need to remove the connection from the app's own side.
Can I set up AI customer support without any coding?
Yes. The marketplace install is no-code, and connecting your help center, writing guidance and setting the handoff are all done in a normal admin interface. Our guide to setting up an AI support agent with no code walks the whole thing through if that's the part you're worried about.
What does it cost to add AI to my existing helpdesk?
It depends on which billing model you land on. Your helpdesk's own AI is usually charged per successful outcome, while a third-party agent like My AskAI charges per ticket or per credit whether or not it resolves, so the bill doesn't climb as the agent improves.
Usage-based extras such as AI Tagging are charged separately on top. Our pricing page has the plans: for a team under 5,000 tickets a month, Pro at $199 a month covers it, with 1,000 credits included, $0.12 a credit after that, and 5 team seats.
