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?
- No-code, low-code or custom build: which route do you need?
- How to set up an AI support agent with no code, step by step
- Step 1: Install the agent into your helpdesk
- Step 2: Point it at the knowledge you already have
- Step 3: Backfill from historic tickets when your help center is thin
- Step 4: Write the rules in plain English
- Step 5: Scope which tickets it sees, and pause the native AI if one is already answering
- Step 6: Run it silently before it says a word to a customer
- Step 7: Find the line, the tickets your own systems have to answer
- How to test it before you let it near customers
- How do I get AI to grade the drafts for me?
- What breaks, and how do you tell?
- What should you do next?
- FAQs
Created time
Aug 25, 2026 05:04 PM
Title length (<60)
Author
Last optimised
Ecomm?
Tags
Image
Publish date
Aug 6, 2026
Video
Slug
how-to-set-up-an-ai-support-agent-with-no-code
Featured
Type
Article
Ready to Publish
Ready to Publish
By the end of this you'll have an AI agent answering real tickets inside the helpdesk you already run, live in less than a day, with no developer and nothing rebuilt.
The no-code route is four moves: an admin approves the agent as an app in the helpdesk you already work in, then you point it at the documentation you have. You write the tone and escalation rules as ordinary sentences, then scope which tickets the agent is allowed to touch. If you're switching from Fin or Zendesk AI, we can have you live inside a day, and only one job needs a developer: any answer that lives in your own order or billing system.
You own the queue. The repetitive tail is eating your week. Password resets, WISMO, the same five policy questions typed a hundred different ways.
Every time you look at AI, someone says engineering has to scope it first. You don't have engineering. I've been on plenty of those calls.
You'll finish with an agent answering one ticket type, plus a test you can run before it goes near a customer. You'll also have a written list of what still needs a developer.
Nothing you've already built gets torn out. Your macros, tags and routing rules stay where they are, and we can reuse your existing macros as Custom Answers. The agent adds its own tags and its own triggers alongside yours.
What do you need before you start?
TL;DR: Admin rights in your helpdesk, whatever knowledge you already have, and less than a day. No developer and no engineering ticket.
- Permissions: admin rights, and a seat. For our Zendesk app you must be an Admin in your Zendesk account to connect, and the agent uses one of your existing Zendesk ticket agent seats or a new one you create for it. Intercom asks you to approve a separate authorizations screen and wants a team-member seat for the email sender, because the AI sender can't be the bot account. Each knowledge source you connect carries its own authorisation screen too. Allowing access isn't the end of it: you get a Connecting to Zendesk screen, land back in My AskAI, then choose Internal Note or Direct replies and whether the agent answers only the first message in a ticket or all of them.
- Plan: installing is an admin-approved OAuth marketplace app authorisation. On our side the Zendesk Tickets integration runs on all paid My AskAI plans, and signup needs a business email address.
- Data prep: published help center articles, a public website, or your ticket history. Any one is enough to start. If your documentation is thin, Step 3 backfills it from your ticket history.
- What you can skip: rewriting the help center, cleaning up your tag taxonomy, and picking a final vendor. We recommend all three eventually. None of them has to happen before you test. Run this on what you have today.
No-code, low-code or custom build: which route do you need?
TL;DR: Three routes exist, and only the custom build needs engineering time. Most support teams want the middle one, a third-party agent that installs into the helpdesk they already run.
Vendors use these words loosely. By no-code I mean you configure everything in a user interface. You write your rules in plain English and never touch a build step.
Low-code means a user interface, plus a small snippet or a webhook. Somebody technical writes it for you. Custom build means a codebase you own, host and maintain.
Most tools sold as no-code are standalone builders (a separate product with its own chat window, sitting outside your helpdesk). Here's the catch: an agent answering in its own widget never touches the queue your team works in, so the macros, tags and routing rules you already run apply to none of those conversations. We build the other kind.
Route | Best for | Setup time | Needs a developer? | Trade-off |
Your helpdesk's own AI | Teams already on a tier that includes it | Hours | No | Priced and gated by the same vendor, so you inherit their model, their limits and whatever they decide to charge |
A third-party agent that installs into your helpdesk (no code) | Teams keeping their helpdesk, macros and routing | Less than a day | No | You run two vendors instead of one, and you own the knowledge quality |
Build it yourself on an agent builder or framework (custom) | Teams with engineering capacity and an unusual workflow | Weeks | Yes | You own the maintenance from launch day, forever |
Route 1 is the AI your helpdesk already sells you: Zendesk's own AI agents, Intercom Fin, and the equivalent inside Freshdesk or Gorgias. You stay on one invoice and one security review instead of two of each.
Zendesk says its AI agent answers from your help center content, and stops where an article is missing or unpublished. Third-party agents work the same way: ours answers from whatever knowledge you connect, and stays quiet where it doesn't cover the question.
Intercom documents Fin handing off on a language mismatch, low confidence, or a customer asking for a person. Route 2 agents hand off on the same triggers.
Zendesk sold that add-on as Advanced AI; it's now branded Copilot.
Route 2 is an approved app in your helpdesk's marketplace that answers tickets inside the helpdesk you already run (we build one of these). Your agents keep the same inbox, and your macros, tags and routing rules keep working.
Route 3 is a codebase. A production support agent keeps costing engineering time long after the demo works. That's why I tell teams without spare engineering capacity that building your own almost never makes sense.
If you're not sure, start with Route 2. You find out whether AI works on your tickets without touching anything you've already built, and you can reverse it.
How to set up an AI support agent with no code, step by step
TL;DR: Seven steps. Install, load your knowledge, backfill from historic tickets, write the rules in plain English, scope what it sees, run it silently, then find the one thing that needs an API.
Step 1: Install the agent into your helpdesk
Start in your helpdesk. Third-party agents install one of two ways: as an approved app from the helpdesk's own marketplace, or as an API integration somebody wires up on your side.
We'd favor the app. It has been through the marketplace and the helpdesk's own approval process, which means it's secure. The helpdesk is aware it's there, it's more stable, and it's less likely to be taken down.
An admin finds it in the helpdesk's own app marketplace and approves the OAuth scopes once; you migrate nothing. The exact menu path differs by helpdesk. Follow our per-helpdesk guide for Zendesk, Gorgias, Intercom or Freshdesk if you want the clicks.
Ours is an approved app in Zendesk (Support Inbox), Zendesk Messaging, Intercom, Freshchat, Freshdesk, Gorgias and HubSpot. Setup is typically 10 to 15 minutes. We can move you off Fin or Zendesk AI in less than a day.
You're finished here when the app shows as installed and connected on the helpdesk side.
Step 2: Point it at the knowledge you already have
Now you're in the AI tool's own console. Every agent needs something to answer from. On our onboarding calls, this is the step I always walk through in detail.
That means URLs and file pickers: help center address, public website, and the PDFs and SOPs sitting in a shared drive.
In My AskAI you connect a help center URL, sync a website, and upload PDFs and Word files. You can also connect Google Drive, Notion, OneDrive, Dropbox, SharePoint, Confluence, Salesforce and Shopify.
Connected sources resync every 24 hours; websites refresh every two weeks. Where the wording has to be exact (regulated answers, or an imported macro library), Custom Answers return a fixed reply verbatim.
Our guide to training AI on your knowledge base covers what makes a source retrieve well. The Knowledge feature page lists every source we support.
You're finished when each source is listed as synced with a page count next to it.
Step 3: Backfill from historic tickets when your help center is thin
Still in the AI tool. "Our documentation is a mess" is the objection I hear most. Teams who say it assume they need a three-month writing project before they can start.
Luckily, an agent that can read your past tickets can draft the knowledge nobody ever wrote down. That's all configuration.
Our own docs on historic-ticket training say it in one line: "Don't already have help docs/a help center? No problem, we can help you generate one automatically from your historic tickets."
Connect the helpdesk and My AskAI trains on 5,000+ historic tickets. It then generates around 20 grouped help articles, usually visible within 6 hours of connecting. They land at
Improve > Self-learning, unpublished, for you to review before anything reaches a customer.
If you skipped it at onboarding, start it yourself from
Knowledge by toggling on Train on historic tickets, up to a week after signup. After that, ask us in chat and we'll trigger it.Barn Owl came to us with roughly 70,000 legacy tickets. We train on the recent ones, which are the tickets worth learning from.
Our guide to optimizing a knowledge base for AI agents covers what to fix in the drafts.
You're finished when draft articles are waiting for review in the self-learning queue.
Step 4: Write the rules in plain English
Still in the AI tool. On any competent agent, you configure how it behaves. That means tone, terminology, when to stop and fetch a human, and what it must never say.
If a vendor makes you write that in a scripting language, you're looking at a low-code product wearing a no-code label.
We call ours Guidance rules. They come in three types: Communication & Style, Context & Clarification, and Handover & Escalation. Keep each one under 75 words.
A real escalation rule reads like a sentence somebody typed in a hurry. One of ours, straight from our Guidance docs: "If a user asks about setting up multiple brands, multiple accounts, or requires multiple AI agents, immediately escalate the conversation to a human support agent."
The panel shows how often each rule fires, so you can see which ones earn their place. The Guidance feature page has the full set of examples, if you want a starting library.
You're finished when you have at least one rule of each of the three types saved.
Step 5: Scope which tickets it sees, and pause the native AI if one is already answering
Back in your helpdesk. Decide what the agent is allowed to touch before you switch it on, and start narrow: one channel, or one ticket type.
If your helpdesk's own AI is already replying, pause it here. Two agents answering the same customer is the most common self-inflicted failure we see in a rollout.
With My AskAI you scope by channel and by brand. Our standard advice is one agent per brand, so one brand's articles never answer another brand's tickets.
Where your helpdesk has tagging (Zendesk, Intercom, Freshdesk and Freshchat), you can layer per-tag reply control on top, as an optional metered add-on. Each tag gets a setting for whether the agent replies to tickets under it.
Tagging reads what the customer wrote and assigns a reason-for-contact category or sentiment. I'd use it to route anything angry to a person.
You're finished when you can name the scope in one sentence, and any native agent that was replying is paused.
Step 6: Run it silently before it says a word to a customer
Back in your helpdesk, watching real tickets. The safe first move is a mode where the agent drafts on every live ticket and the draft lands as an internal note.
Customers see nothing. Your team reads the answers the AI would have sent and grades them beside whatever is answering today.
Internal Notes mode is our default. You check the quality before you switch anything on.
You can stop here for good. One of our customers, a Swedish DTC apparel brand on Gorgias, handles roughly 2,400 tickets a month this way, with direct replies held back. Sizing and comfort questions need a human's judgment on brand voice.
You're finished when internal-note drafts are appearing on live tickets.
Step 7: Find the line, the tickets your own systems have to answer
Back in your helpdesk, reading those drafts. Everything we've done to this point was configuration. The rest needs an API.
After a few days you'll see it yourself: the tickets that stall halfway, where the answer sits in a system only you own. That's order status, plan changes, refunds and account edits.
An agent can look something up or change something only if it can call an API that exposes it. Somebody has to expose it. That is the only place you need a developer.
Custom Tools and APIs are the one part of our standard setup that takes developer time.
Shopify is a pre-built connector, so product, order and customer lookups (including Shopify Markets) need nobody technical.
Everything else runs through User Data, where a CRM, billing system or order database connects via an API endpoint, or through Tasks and Tools for actions.
You write the Task in natural language; a developer writes the endpoint. Alex, our CTO, works with customers on the complex ones, on Scale and Enterprise.
Meanwhile, hand that ticket class to a human on purpose. The AI takes the repetitive bulk and escalates the rest with a conversation summary attached, so your team picks up cleanly.
You're finished when you have a written list of the ticket types that need an API, plus a routing rule sending them to a person.
How to test it before you let it near customers
TL;DR: Grade a week of internal-note drafts against what your agents sent. If the agent is right on your top five ticket types and hands over cleanly on the rest, you're ready to flip it on.
- Batch-test your top five ticket types. Run the questions you know arrive weekly, and check which knowledge source produced each answer. Pass: the right source, every time.
- Grade a week of internal notes against the human reply that followed. I write the pass mark down before I start reading. Set it afterwards and it becomes whatever you got.
- Force a handover. Ask for a person mid-conversation. Pass: the summary is good enough that the customer never repeats themselves, and the AI stays quiet until it's handed back.
- Test one thing it should refuse. Ask something outside your knowledge. Pass: it escalates instead of improvising.
- Test out of hours. The 24/7 outcome is why I hear from most of the teams that sign up. Pass: an overnight ticket reads as well as a midday one, because the same knowledge and rules apply at 3am.
When an answer looks wrong, ask our Echo assistant why the agent gave it and which source it used. That audit view lives in your team's dashboard.
Go-live gate: one ticket type with direct replies on, everything else still in notes mode.
How do I get AI to grade the drafts for me?
Reading a week of internal notes by hand is the slow part. I'd hand the first pass to a model, then read the ones it flags myself. Keep the pass mark you wrote down in check 2, and paste your own pairs in.
You are grading a customer support AI's draft replies before they are switched on for customers.
I will paste pairs of [AI internal-note draft] and [the human reply that was actually sent]. My top five ticket types are [list them]. My written pass mark is [paste it].
For each pair, return one row: ticket type, whether the draft would have resolved the ticket on its own (yes / no / partly), what it got wrong, and which of these four causes it was.
1. Stale knowledge - the draft matches the published article, and the article is out of date.
2. Missing knowledge - nothing published covers the question.
3. Needs data held in another system - order status, plan changes, refunds, account edits.
4. Handover failure - it should have escalated and did not, or escalated when it should not have.
Then give me a count per cause, the three articles worth fixing first, and the ticket types to keep routed to a person until an API endpoint exists.
Where the text alone does not tell you, write "unverified, check the ticket" rather than guessing.What breaks, and how do you tell?
TL;DR: The most common failure is the AI being right about something your help center says and your team stopped doing eighteen months ago.
Symptom | What's actually happening | Fix |
Nothing happens at all after install | The traffic is arriving on a channel the agent was never scoped to, or replies are paused | Our Zendesk Tickets docs set the default: "By default your Zendesk Ticket AI agent will respond to Email tickets that are “New” or “Open” (identified by a webhook trigger)." Web forms and other channels need their own trigger. Check the Pause AI agent toggle while you're there |
Confident answers that are wrong | An article changed or came down at the source, and the old copy is still sitting in the agent's knowledge | Our Connections docs are explicit: "If you “Unpublish” or move an article back into “Draft” in Intercom, Zendesk or Freshdesk, you will also need to delete it from your My AskAI account." Fix the article, republish, then confirm in Improve > Inspect & Logs, which shows the knowledge sources used to answer |
Answers arrive in the wrong language | A Custom Answer was written in the customer's language rather than in English, or somebody wrote the language rule as a Guidance rule | Our Custom Answers docs ask for English, because they are "automatically translated into the user’s language when answering questions." And our Guidance docs list changing language under what guidance should not be used for; the language feature is the right lever |
Handover fires on everything, or never fires | The handover setting and your escalation rules are pulling in different directions | On Zendesk Tickets you pick either Cannot answer (hand over whenever it can't answer) or Escalation request (hand over only when the customer asks). Choose one, then narrow each escalation rule to a named scenario. Inspect & Logs filters to Escalated to human, so you can read the ones firing most |
Two replies land on the same ticket | Your helpdesk's own acknowledgment is still switched on underneath direct replies | Our Zendesk Tickets docs say it plainly: "If you’re using your AI agent to directly reply to users in Zendesk Tickets, we recommend you disable automatic email replies (for ticket creation acknowledgement), otherwise users will receive multiple emails." Check it again after any helpdesk settings change |
Confident wrong answers are the failure I'd bet on. Most so-called AI hallucinations in support come straight out of your own docs.
No tuning fixes an article that says 45 days when your policy has said 50 for a year. The quieter version catches people out too, where replies are switched on and nothing comes out because a person got there first. That one is deliberate, and our docs say so. "To stop the AI agent replying in Zendesk tickets, just reply to the ticket yourself and the AI agent will automatically stop replying."
What should you do next?
TL;DR: Once one ticket type is answered well and hands over cleanly, the next job is knowledge coverage. Handovers tell you where it's thin.
Widen the scope one ticket type at a time. After each one, read the handovers first.
Every handed-over ticket is a question your knowledge couldn't answer. Our Self-Learning drafts a new article by comparing what the AI said to what your agent sent. You approve it or you rewrite it.
Honeygain answers roughly 600 tickets a month from that auto-drafted knowledge alone.
YesLMS reached 76% AI resolution on about 300 tickets a month, with 88% AI CSAT.
If your remaining blockers are all API-shaped, you are back at Step 7. Write the list, then keep those ticket types routing to a person until the endpoints are scoped.
Where the help center is the blocker, Train on Historic Tickets is the shortcut I'd take. It drafts your starter knowledge from the tickets you've already answered, so you skip the write-up you would otherwise do by hand.
FAQs
How to make an AI chatbot without code?
Pick a third-party agent that installs into the helpdesk you already run, approve the app as an admin, and point it at your help center or website. From there it's plain-English rules for tone and escalation, plus a scope setting for which tickets it sees. No build step, and on our side setup is typically 10 to 15 minutes.
What is a no-code AI?
No-code means every part of the setup happens in a user interface. The build step is what separates the three approaches.
Approach | What you do | Who does it |
No-code | Connect sources, write rules as sentences, set scope | You, in the interface |
Low-code | The same, plus a small snippet or webhook | You, plus somebody technical |
Custom build | A codebase you own, host and maintain | Your engineers |
Can I set up AI customer support without any coding?
For the standard setup, yes: connecting a helpdesk, loading knowledge, writing tone and escalation rules, and scoping which tickets the agent answers are all configuration. The exception is anything that reads or changes data in a system only you own, like order status against a bespoke backend; that needs an API endpoint, and exposing one is a developer's job. Meanwhile I'd route those ticket types to a person.
How do I set up AI customer support that runs 24/7?
24/7 here means the clock is the only difference, so there are no hours to configure. I'd run the out-of-hours check for a week, read those drafts beside the daytime ones, and widen the scope when they match.
Can I roll this back if it doesn't work?
Yes. We make notes mode the default for this reason: while the agent drafts as internal notes, nothing is visible to a customer, so rolling back is a toggle. The install is an app authorization, so removing the app removes the access, and the app never modified your macros, tags, views or routing rules.
Does this cost anything on my current helpdesk plan?
Installing a marketplace app is an authorization, so on most helpdesks it costs nothing on their side; a few gate app installs behind a tier, so check yours. The agent itself is priced separately by whoever you pick. Ours is roughly $0.10 per ticket the AI handles, and the 30-day free trial unlocks every feature with unlimited tickets and no card, so you can prove it before you pay.
