Duckie AI: Complete Guide to Features, Pricing & Limitations (2026)
Duckie AI's support agents investigate a ticket and close it. Customer pages publish 90%, 92% and 95% under three labels. No price. What to ask on the call.
Summarize with AI
Duckie is the YC-backed vendor behind an AI agent that investigates a support ticket, takes the action it needs and closes it on its own. It suits a technical B2B support team in fintech, dev tools or SaaS, whose tickets need investigating, account actions taken and engineering context pulled in before a reply goes out. The catch is commercial: Duckie publishes no price and no trial terms, so the only route to a figure is a sales call that asks for your ticket band first.
You have probably just come off a demo, or a founder has forwarded you the Y Combinator profile and asked what you think.
You go looking for the two things any support lead needs before an internal conversation: what does it cost, and what happens when the AI gets something wrong on a ticket that mattered? Neither answer is on the site.
I'm Mike, co-founder of My AskAI. We sell an AI support agent too, so I have a stake in this.
By the end you'll know what Duckie does, what it does not publish, what to press on during the call, and whether it fits your stack.
Duckie at a glance
| Fact | Detail |
|---|---|
| What it is | Agentic support platform for technical B2B teams |
| How you're billed | Per-ticket (Grid customer page); rate unpublished |
| Best for | Investigation-heavy tickets in fintech, dev tools and SaaS |
| Not a fit for | Teams needing a published price |
| Biggest limitation | No public pricing and no public trial terms |
| Our verdict | 7/10 |
What is Duckie AI?
The homepage tells you who Duckie thinks it is selling to. The line is: "Support teams become AI builders, shipping agents that resolve tickets, investigate issues, and automate any workflow behind support."

I read that as three things:
- Agents that decide for themselves which knowledge and tools to use on a ticket.
- Visual workflows and runbooks for the paths you want followed exactly.
- A layer of safety rules (Duckie calls them guardrails) that get checked before the agent does anything.
Duckie's documentation sets out the loop plainly. A message arrives, safety rules run, the agent searches knowledge and calls tools, a reply goes out, and the conversation gets classified. Every message creates a run I always ask to open in a demo.
It started as an internal tool. The about page says one of the founders was tired of answering the same questions on call, so she built a bot that could search the documentation and answer from it. That origin still shows, I think, because the product is aimed at tickets with a stack trace or a broken integration in them.
The company is YC-backed and based in San Francisco.
Duckie draws its own line between a chatbot and an agent on the same page:
"We don't build chatbots that explain how to do things. We build AI that actually does them, so you don't have to keep closing tickets manually." — Duckie's about page
That is a fair description of what it does. Duckie's agent takes the action a ticket needs and closes it.
How easy is it to set up Duckie AI?
Duckie's quickstart promises you can be up and running in ten minutes. You connect a support channel under Settings → Connections with OAuth, add knowledge under Train → Knowledge, create an agent, and deploy it.
I rate Launchpad highest in the setup flow. It reads your past tickets and drafts the starting configuration for you, and the docs give you the dials: "The time window can be 1-12 months, and the ticket limit can be 100-5,000 tickets."
The guidelines, knowledge items and runbooks it generates stay in draft until a human approves them.
Duckie is direct about who does the setting up. The about page says your support lead should be able to configure the agents in plain English, with no engineering help and no long timeline.
Buyers get stuck on the commercial path. The homepage CTA is a demo request with a work-email field, and the call page asks for your monthly ticket band before it books anything. A signup form exists, but Duckie publishes no trial terms alongside it (worth flagging to whoever signs the contract).
Once you are in, the safe first deployment is Testing Mode. The deployment modes page documents two safeguards: Internal Notes Only, where the agent's replies land as internal notes for your team, and No Write Actions, where anything marked as a write action gets skipped. Duckie's own recommended order is playground, then testing, then live.
What channels does Duckie AI work in?
The integrations page groups everything into four buckets: support platforms, messaging channels, knowledge sources and issue tracking. Duckie's framing is that each integration can be a knowledge source, a channel for agent actions, or both. We think about our own connectors the same way.

On the helpdesk side you get Zendesk, Intercom, HubSpot, Freshdesk, Plain and Pylon. On the messaging side, Slack, Discord, Gmail and a website widget you embed on your own pages.
The supported integrations table in the docs adds two the marketing grid leaves out. Microsoft Teams is a full agent channel, though it does not sync knowledge, and it has its own docs page (worth a second look if your team already lives there). Web pages and URLs count as a knowledge source.
Both lists describe the same product: one built for typed, technical support. Gorgias is on neither one, which rules Duckie out for a lot of ecommerce teams, and neither carries a voice channel.
The internal use case gets less room on Duckie's site than I think it deserves. Slack works as both a knowledge source and a channel, so the same agent can answer your own engineers and CSMs. Automox runs an internal Slackbot alongside its customer-facing deployment.
What are the limitations of Duckie AI?
Nothing about cost is published: Duckie has no pricing page on its site, and no page in its documentation names a price, a tier or a pricing model; the in-app billing docs describe how to upgrade a plan but never name one or its price. That, in my view, is the limitation most likely to slow you down.
The one path to a figure is a demo call, and it asks for your ticket band first. That makes an early budget conversation with your finance lead almost impossible. We publish our own rate for exactly that reason.
The commercial terms are not public either. Support coverage outside your own hours, what the support agreement promises, and how firm the roadmap is are all left to the call.
Write those three down before you dial in.
Every published outcome carries a different label (worth checking which one before you compare it to anyone else's). On the customers page you'll find a 90% ticket automation rate, a 92% ticket deflection rate and a 95% resolution rate, sitting beside each other.

Those are three different measurements. Duckie's own performance metrics docs are explicit about it: "Configure resolution rules before using the page to judge agent quality."
Resolution rules are where you decide what counts as resolved, deflected, escalated or unresolved. Each customer sets that bar for themselves. We wrote a full breakdown of containment, deflection and resolution if you want the definitions side by side.
It is built for technical support. The published customer list is fintech, IT automation, logistics and dev tools.
There is no Gorgias integration and no published language list. If your day is refunds, order status and returns at volume, this is the wrong product.
What knowledge sources can I train Duckie AI on?
Duckie starts where your answers live: in tickets your team has already handled. Launchpad reads them from a connected ticketing platform and drafts guidelines, knowledge items and runbooks for review.
The documented knowledge sources cover Notion pages and databases, Confluence spaces, Google Drive files, Slack channel history, web pages, and help center articles from Zendesk and Intercom. The marketing grid adds Sentry, Fireflies, Guru and Skilljar, and engineering context comes in through Jira, Linear and GitHub (a lot of surface area to keep in sync).
You can also write knowledge directly inside Duckie for the answers that exist nowhere else.
Control sits at the agent level (worth checking before anything else). Duckie's security overview puts it like this: "For agents, enable only the tools, knowledge, guardrails, workflows, runbooks, and callable agents needed for that agent's job."
The same knowledge serves both the customer-facing replies and the internal-note drafts. Testing Mode is a deployment setting, so there is no separate, safer corpus for the drafts a human reviews (the safer default until you trust the agent more).
What features does Duckie AI have?
Duckie gives you three architectures and expects you to pick per job. Start with one agent, and add the stricter forms when a process has to be followed exactly (compliance-flavored tickets, mostly).
Autonomous agents work it out at run time. Workflows are a visual sequence with conditions; runbooks are a step list the agent follows. I read runbooks as the safest starting point for anything compliance-flavored.

Bigger setups delegate to sub-agents, so one agent hands billing questions to a billing specialist. Tools are what let any of them act: Duckie's own tools, your connected apps, custom tools you define, and MCP servers, which are a standard way to plug an outside system into an agent.
Actions are the reason a team buys this at all. Duckie's homepage walks through a plan change in Stripe, a cancellation with a retention offer and a refund policy check. I'd start with the Stripe example if you want to see this live.
Rules sit on top of those: refunds over $200 go for human approval, and customer details get stripped before any tool call.
Whether the AI takes those actions itself is your call. Refunds, disputes and account changes can run on their own, or get flagged and handed to your team. Where a mistake is expensive to undo, start with the handoff.
For a support lead, I start with guardrails: "Guardrails are safety constraints that protect your customers and brand by defining what your agent cannot do and when it should escalate to humans."
Escalation triggers include an angry customer, a legal question, a VIP account, or someone asking for a person. Alongside them, tagging and resolution rules classify each conversation as it closes.
Testing Mode is how Duckie handles draft-for-approval. With Internal Notes Only switched on, the agent writes what it would have said as an internal note, and your team decides whether to send it.
How do I improve Duckie AI responses?
The testing layer is the strongest part of the product. There are three ways into it. Replay testing pulls real conversations from Slack, Zendesk, HubSpot or Intercom and runs them again (handy for catching a regression before a prompt change ships).
The playground under Test → Playground is the interactive one. Test → Batch Test runs a set at once and scores each one from 1 to 5, with optional per-metric scores and notes.
When something goes wrong on a live ticket, run history is where you go. Duckie's docs describe it as "a complete log of every agent execution", built for debugging, auditing and understanding what the agent did. Each run holds the message that triggered it, every step, the knowledge retrieved, the tools called and the final reply.
Alerts cover safety-rule violations, tool failures and escalations, and land in email or Slack (the Slack route is the one a support team will see). Duckie's own suggested rhythm is a daily health check, a weekly look at the runs behind any metric change, and a monthly trend review. A support engineer at Automox describes the assistant side of that work:
"Being able to use the assistant to create things like runbooks, guidelines, and guardrails, and troubleshoot specific runs has made things quite a bit easier." — Jonathan White, Staff Support Engineer, Automox, on Duckie's homepage
Duckie's own documentation treats an autonomous agent as something you keep tuning after launch.
I've sat with a team that ran an autonomous agent for months. Their two problems were a version change that moved accuracy in ways nobody expected, and upkeep the vendor never took back.
Ask who owns tuning after launch, and what a version upgrade does to your safety rules. Ask how you see what was resolved, and who picks up the phone when it drifts.
What resolution rate can I expect from Duckie AI?
Duckie's customers page carries three figures. I take the word attached to each one as seriously as the number.
Grid reports a 90% ticket automation rate. Automox reports a 92% ticket deflection rate. Vanquish Trader reports a 95% ticket resolution rate (three different metrics, worth noting before you line them up against each other).
Automation, deflection and resolution count three different things. Ask which of the three any number on a slide is measuring.
Duckie's performance metrics docs define deflection rate as deflected runs divided by total tickets, and resolution rate as resolved runs divided by total tickets. Which runs count as which comes from the resolution rules each customer sets up.

Nearly every vendor in this space does something similar, us included, which is why we spell out how we count. It means a Duckie number and a competitor's number are rarely measuring the same thing.
Our own AI resolution rate benchmark study pooled 195 rated deployments across 38 vendors. The median came out at 70%, with the middle half sitting between 56% and 80%.
Take our own figure with a grain of salt too. It is an aggregate, the labels differ between vendors exactly as they do here, and it only covers vendors who publish a number at all.
A number in the mid-nineties is worth a follow-up question. Ask what drove it, because the fastest way to a perfect deflection rate is to take away the path to a human. The customer satisfaction score is where that shows up first.
What AI model does Duckie AI use?
Duckie's homepage promises every frontier model, and names OpenAI, Claude, Gemini, Mistral, Llama and DeepSeek as switchable options.
The security page is more specific about what runs by default: "By default, for closed-source models, we use OpenAI and Anthropic." It states zero-day retention at OpenAI, says Anthropic does not use customer data for training, and notes that the open-source models it runs share no data with third-party vendors.
Teams who want a tighter boundary can choose an AWS Bedrock-only setup with zero retention and no training.
I have made this argument before. Model choice is the vendor's job, it changes every few months, and a production support agent is a dozen tuned calls behind a single reply.

Three things decide answer quality: whether the right knowledge gets found before the model writes anything, whether the safety rules hold when a customer pushes, and whether you can open a conversation afterwards and see what happened.
Duckie documents all three, including prompt injection defenses. The worry there is a customer message or a synced page carrying instructions meant to override how your agent behaves.
Duckie does not publish which model handles which step, or any accuracy figure per model. Almost nobody in the category publishes that.
What languages does Duckie AI work in?
There is no language count on the homepage, the about page or the integrations page, and nothing in the documentation names a supported set.
We would expect the agent to handle a French ticket, because modern models handle plenty of languages by default. Duckie publishes no language commitment you could hold it to, and nothing on how the search side behaves when your knowledge is English and the ticket is French.
Make it a test. Replay testing can pull real conversations from Zendesk, Intercom, HubSpot or Slack, so ask for a replay of ten real tickets in each language you support, and read the replies with someone who speaks it.
How secure is Duckie AI?
Duckie states SOC 2 Type II certification on its about page and in the homepage footer. Ask for the current report during the security review, which is what our own reviewers ask us for.
The security page is more detailed than most at this stage. It commits to:
- 99.9% uptime, and no rights reserved to use or disclose customer data.
- Direct data access limited to two named people, multi-factor authentication for employee access, and deprovisioning within 24 hours.
- Customer data kept off the public internet, with failover across multiple US regions.
The documentation adds the operational layer your reviewer will ask about: workspace and custom roles, API keys with scopes, per-agent tool access, account-safe actions, safety rules, and the run history that gives you an audit trail.
Duckie's self-hosting docs cover Docker and AWS setups, for teams who need to control where data sits or reach systems that are not exposed publicly. Automox has run it that way since the first version.
Four things are not addressed on any public page: HIPAA, a GDPR compliance statement, data residency for the hosted product, and how long hosted data is retained. Put those on the list for the call.

Who is using Duckie AI?
Grid is a fintech handling more than 70,000 tickets a month, and reports a 90% ticket automation rate. Its product manager describes what changed for his team:
"Prior to Duckie, managing the bizops team or the customer support team was about managing humans. Now, it's becoming more about someone who can manage an AI system." — Matthew Kim, Product Manager, Grid
Automox is IT automation, reports a 92% ticket deflection rate, and has run Duckie self-hosted since the first version.
Abbiamo moves more than three million deliveries a month through its logistics API. It went from three support employees to zero, and reports a 66% cut in support cost while ticket volume grew 30%. Its founder puts the outcome in headcount terms:
"Duckie costs the same as one full-time employee I had. And I got rid of three. But more than the savings, I don't need to deal with hiring people, training people, watching the knowledge walk out the door when they leave. That's the real value." — Aurélio de Pádua Gandra, Founder & CEO, Abbiamo
The logo wall adds Vellum, Snowplow, Mintlify and Vapi: developer tools and data infrastructure, where the tickets are technical.

How much does Duckie AI cost?
What Duckie publishes about pricing
Nothing. There is no pricing page on Duckie's site, and no plans or pricing page anywhere in the documentation.
The demo is the only way in. Duckie's call page says: "Book a 20-minute call with us to explore if Duckie could be a good fit for your customer support team."
It asks for your monthly ticket volume band before it will show you times, which suggests volume drives the quote, though Duckie does not state the model. Whether you are charged per ticket, per resolution, per seat, or on a platform fee is not published anywhere.
You cannot put a range in front of your finance lead, and you cannot rule Duckie in or out before spending an hour on it (a bad trade for what should be a yes/no decision).
What to establish on the call
- The unit of charge. Per ticket, per resolution, per seat, a platform fee, or some combination. This is the single number the rest of the conversation hangs off.
- What counts as a resolution. Resolution rules are configured per customer, so if you are billed per resolution, ask who signs off on the definition and whether it can change later.
- Overage and minimum term. What happens in a spike month, and how long you are committed for.
- The self-hosting premium. Three deployment models are documented, and running Duckie in your own infrastructure is unlikely to cost the same as the hosted product.
- Who owns tuning after launch. Ask whether ongoing tuning is included, chargeable, or yours to do.
- Which model tier you are on. An AWS-only setup is a different cost base to the default.
What the same volume costs on a published-price agent
| Assumption | Value |
|---|---|
| Tickets per month | 2,000 (chat) |
| AI resolution rate | ~70% (field median, context only) |
| Duckie price per ticket | Not published, quoted on a call |
| Human agents | Unchanged either way |
For that volume, My AskAI's Pro plan is the lowest one that covers it. That is $199 a month including 1,000 credits, then 1,000 more at $0.12 a credit, so $319 a month. Usage add-ons such as Tasks, Tools, Tagging and Live Translation are billed on top.

You pay per ticket whether the AI resolves it or not, so the bill does not climb as the AI gets better.
How do I work out what Duckie has to quote?
Duckie's number arrives on the call, so it pays to have your own break-even worked out before then (and the six questions above ready to read out). Paste this into ChatGPT or Claude with your figures in the brackets, then take the output into the call.
You are helping me price an AI support agent for my support team.
My numbers:
- Monthly ticket volume: [e.g. 2,000]
- Channel mix: [e.g. 80% chat, 20% email]
- Helpdesk: [e.g. Zendesk]
- Fully-loaded cost of a human-handled ticket: [e.g. $4.50]
My published-price benchmark: $199 a month including 1,000 credits, then
$0.12 per credit after that, charged per ticket whether the AI resolves it
or not.
Do this:
1. Calculate my monthly bill on the published-price benchmark at my volume.
2. Calculate the highest price-per-resolution a vendor could charge before
it costs more than that benchmark, at resolution rates of 50%, 70% and 90%.
3. Put the three break-even prices in a table with the assumptions beside them.
4. Add a one-line follow-up question for each of these six items, so I can
read them out on the call: unit of charge, what counts as a resolution,
overage, minimum term, self-hosting premium, and who owns tuning after
launch.
Where you do not have a figure, write "unverified, ask the vendor" instead
of guessing. Do not estimate any vendor price that is not published.Does Duckie AI have a free trial?
There is no trial offer on the homepage, the about page or the customer stories. A "Create an account" form asks for an email and a password, with no terms next to it, and the quickstart points you at either that form or a demo booking.

Duckie's evaluation path is a pilot you agree during a sales process. Testing Mode is a good one once you are inside, because Internal Notes Only lets the agent draft on real tickets with nothing reaching a customer.
Either way, if your buying process needs you to try software before you talk to anyone, Duckie will cost you a call first.
Is Duckie AI worth it?
This rating rests on Duckie's own published surfaces: the homepage, the about page, the customers page, the security page and the product documentation. Where Duckie publishes nothing (price, trial terms, supported languages), the gap is named as a gap rather than filled with an estimate.
Duckie earns its seven out of ten on the parts most agents in this category skip: tickets that need investigation, real actions with permissions attached, a testing layer you can run before anything reaches a customer, run-level history you can audit, and a self-hosting option a security team will take seriously.
Every one of those is documented in Duckie's own docs.
The three points it drops are all commercial. You cannot price it, you cannot try it, and the terms your procurement team needs all live on the call.
If those are dealbreakers, we've ranked the other autonomous agents worth looking at in our roundup of Duckie AI alternatives.
Whatever you end up testing, take the same checklist into every demo.
Can you see, per conversation, exactly what the AI used and why? Can you define what "resolved" means, and check it?
Who owns tuning six months in? And what does support look like on the day it starts drifting?
What are the pros and cons of Duckie AI?
Pros
- Built for tickets that need investigating. Autonomous agents with sub-agents, custom tools and MCP servers, plus Sentry, GitHub, Jira and Linear as context, which is why Grid can have it walk a member's full ledger before replying.
- An operating loop you can inspect. Testing Mode drafts as internal notes or blocks write actions, replay and batch testing catch changes before they go live, and run history logs every step of every execution.
- Security a reviewer can read. SOC 2 Type II, documented access control and prompt-injection defenses, account-safe actions, and self-hosting in your own infrastructure.
Cons
- No published pricing and no published trial terms. There is no pricing page and no plans page in the documentation, so a budget number needs a 20-minute call.
- Commercial terms only surface on the call. Support coverage, what the support agreement promises and how firm the roadmap is are not on any public page.
- Outcome numbers carry a different label on each customer page. Automation, deflection and resolution sit side by side at 90%, 92% and 95% on the customers page, and resolution rules are configured per customer.
- Built for technical B2B. No Gorgias, no voice channel, no published language list, and a customer set that is fintech, IT and dev tools.
Our rating
FAQs
What is Duckie AI?
Duckie is a YC-backed platform that lets support teams build AI agents which resolve tickets, investigate issues and run workflows behind support. It works inside your existing helpdesk and channels. Duckie describes the difference from a chatbot on its about page: "We don't build chatbots that explain how to do things. We build AI that actually does them, so you don't have to keep closing tickets manually."
Does Duckie AI support self-hosting?
Yes. Duckie's self-hosting docs cover Docker and AWS deployment options for teams that need to keep data inside their own infrastructure or reach systems that are not exposed to the public internet. Automox has run its deployment self-hosted since the first version, which is the strongest evidence the option holds up in production.
How much does Duckie AI cost?
Duckie does not publish pricing. There is no pricing page on its site and no plans page in its documentation, so you get a figure by booking a 20-minute demo call, which asks for your monthly ticket band first. The four things I ask on a call like that: what the unit of charge is, what counts as a resolution, what overage costs, and what the minimum term is.
Does Duckie integrate with Zendesk and HubSpot natively?
Yes, both, and each one can serve as a knowledge source and as a channel the agent replies in. Duckie's supported integrations table sets out what each ticketing platform does:
| Helpdesk | Knowledge source | Agent channel |
|---|---|---|
| Zendesk | Yes | Yes |
| Intercom | Yes | Yes |
| HubSpot | Yes | Yes |
| Plain | No | Yes |
| Pylon | No | Yes |
Freshdesk is on the marketing grid too. Gorgias is on neither list.
Who is Duckie AI for?
Technical B2B support teams, on the evidence of who Duckie publishes. Its named customers are fintech (Grid, Vanquish Trader), IT automation (Automox), logistics (Abbiamo) and healthcare (Medtelligent), and its logo wall adds developer-tool and data-infrastructure companies like Vellum, Snowplow, Mintlify and Vapi. One of its customers put the shift well:
> "Anyone can hop on Claude and try to build a chatbot. But to build one that's actually a full system like this? I don't think that's realistic as a one-man job at home." — Trevor Feldmann, Co-Founder, Vanquish Trader, on Duckie's homepage
Is Duckie good for non-technical support teams?
Probably not the best fit. Every customer Duckie publishes is a technical B2B company, and Gorgias, the helpdesk most ecommerce teams run, is not on the integrations list. The product is built around investigating a ticket before answering it, which is worth paying for when tickets are complex and less so when they are order-status questions at volume. If that is your day, we'd start you somewhere else.
What's the difference between AI chatbots and AI agents for customer support?
A chatbot answers a question from your content. An agent takes the action that closes the ticket: it looks up the order, issues the refund, resets the account, then writes the reply. The practical test I use is whether the tool can call your systems and change something in them. Duckie's own framing is that it builds AI that does the thing.
What can Duckie AI be trained on?
Past tickets first, through Launchpad, and the ticket window is set there: "The time window can be 1-12 months, and the ticket limit can be 100-5,000 tickets." On top of that, the documented knowledge sources are Notion, Confluence, Google Drive, Slack history, web pages, and Zendesk or Intercom help center articles, with Sentry, Fireflies, Guru and Skilljar on the marketing grid.
Does Duckie AI have a free trial?
Duckie publishes no free-trial or free-plan terms. A self-serve signup form exists, with no trial terms stated on it, alongside a demo-call path. Once you are inside, Testing Mode is the closest thing to a safe pilot, because the agent drafts as internal notes for your team. Our own trial runs 30 days with every feature unlocked, no ticket limit and no card.
What AI model does Duckie AI use?
Several. Duckie's homepage promises every frontier model and names OpenAI, Claude, Gemini, Mistral, Llama and DeepSeek. Its security page states the default: "By default, for closed-source models, we use OpenAI and Anthropic." An AWS Bedrock-only setup is available for teams who need one, and Duckie documents its prompt-injection defenses separately.
What channels and messaging apps does Duckie AI reply in?
Slack, Discord, Microsoft Teams, Gmail and an embeddable website widget, alongside the ticketing platforms. Slack is the only one of those that also syncs knowledge, so channel history feeds the same agent that answers in it. Microsoft Teams appears in the docs table but not on the marketing grid (the docs are the fuller list of the two), and there is no voice channel on either.
Is Duckie AI SOC 2 compliant?
Duckie states SOC 2 Type II certification on its about page and in its homepage footer, and its security page promises audit reports to customers. Ask for the current report during your security review. HIPAA, a GDPR compliance statement, data residency and retention periods are not addressed on any public page, so put those on the same list.
Is Duckie AI backed by Y Combinator?
Yes. Duckie is a YC-backed company based in San Francisco, and its Y Combinator profile carries the founding details.