AI onboarding, for a SaaS product, means using AI to get a new customer from signup to their first real result. Today it usually means one of three things: AI that writes your tours and checklists, AI that answers a new user's questions inside the app, or AI that handles the admin around a human-led implementation. An AI onboarding assistant should go further: do the setup with the customer, in their own account, and keep that customer set up after they leave. That is what customer onboarding automation should mean: the setup gets done, not described.
All three of today's approaches are real improvements, and all three leave the setup itself with the user. This post is about customer and user onboarding, not HR software for new hires, which goes by the same name. Below: what the tools do, where they stop, and what it takes to treat onboarding as a job the AI owns.
What AI onboarding tools do today
If you are evaluating AI for onboarding, the market sorts into three jobs. The first is generating the guidance: tours, tooltips, checklists, and demos, written or assembled by AI. The second is answering the new user in the app from your help content. The third is running the project around a high-touch implementation.
Here are some of the names you will meet, described from their own pages. It is a map, not a ranking.
- Pendo: an AI guide builder that creates in-app guides from a prompt.
- Userflow: Smartflow builds a flow while you click through your own app, and an Adoption Agent answers users' questions from your knowledge base and website and can perform tasks for them through Action Flows.
- Chameleon: agents that find where new users stop making progress and build the fix, with personalization of who sees it marked as coming soon.
- Userpilot: an agent, Lia, that analyzes product data, recommends a change, and builds the in-app message or tour.
- Appcues: Appcues AI answers your team's questions about how flows perform, follows your writing rules, and suggests experiences to build.
- Arcade: AI-written copy and voiceover for interactive demos.
- Valuecase: for high-touch onboarding, an agent that drafts and updates the plan in each customer's space, chases customers, drafts emails, and updates the CRM.
I want to be fair to this work, because it is good. A tour built from real drop-off data beats one a product manager guessed at two years ago. A checklist that adapts to the user's role beats one list for everyone. An assistant that answers at 11pm beats a help article nobody finds. If your onboarding is a static tour today, an AI onboarding tool for SaaS from that list will make it better.
Look at who the AI is working for, though. In the first and third jobs it works for your team: the product manager gets a flow in minutes, and the CSM gets the follow-up drafted (we look at how a CS team's week changes when the AI owns the work instead in AI agents for customer success teams). In the second it works for the user, mostly by talking, with some assistants able to trigger a flow someone scripted in advance. Most writing about AI user onboarding ends with a list like this one. What the list leaves out is who does the setup.
The setup is still the user's job
Picture the trial from the other side. Someone spent twenty minutes on your site, decided your product might fix a real problem, and signed up on a Tuesday between meetings. They land on a dashboard with a zero in every tile. No data, no integrations, no teammates.
A checklist in the corner says 0 of 6, and step one is "Connect your data source." They have three systems that could be the data source. The checklist is well written, and an AI may have tailored it to their role. It is still six pieces of work assigned to a person who has not yet seen the product do anything for them. They finish one step, hit a question, and the next meeting starts.
The tab closes.
Nobody in this story did anything wrong. Tours and checklists were the sensible design when software could only show and tell. The numbers say how far showing and telling goes. In Userpilot's benchmark data, the average onboarding checklist completion rate across 188 B2B SaaS companies is 19.2%, and the median is 10.1%. Average activation, across 62 companies, is 37.5%.
Source: Userpilot SaaS product metrics benchmarks. The median company sees about one new user in ten finish its checklist.
AI-powered onboarding of the generating kind can improve those numbers at the margin, because it improves the instructions. It does not change who does the work. The first session is short, and the user spends it on setup they never wanted to do. A customer should not open a new product and find homework. They should find progress.
What an AI onboarding assistant should actually do
It should do the setup. With the user, in their first session, in their real account.
This is newly practical for a plain reason: products now expose their own actions as tools. If your product has an MCP server, or an API you can put one in front of, an agent can call the same operations your interface calls. The assistant no longer has to point at buttons. It can do what the buttons do.
Take a hypothetical shipping app. The new user types one sentence: "We ship from Auckland, mostly by post, and our orders come from Shopify." A tour would highlight the Settings menu. An assistant that does the work answers with a log:
acting as matthew@hartwell.io
create_pickup_address · Auckland · done
add_carrier · postal service · done
create_shipping_rule · parcels under 3kg go by post · done
connect_store · Shopify · waiting on your approval
Three things in that log matter. The first is identity. Every call runs as that user, with their permissions, in their account, so the assistant can do nothing the person could not do by hand. I made the longer case for that in what to expose to your customers' AI agents, and it rules out the shared admin key.
The second is the honest gap. One step sits behind a login the assistant should not hold, so it waits. A good assistant does everything it can, then hands back the short list only the human can finish. One approval replaces six chores.
The third is what the user saw. Their own account filled in while they watched, and they ended the session with the product working on their problem. Some people would rather click through and learn the interface, and the assistant should walk them through when they ask. That should be the user's choice.
Then there is what happens after the tab closes, because setup rarely finishes in one sitting. Today the follow-up is a nurture email: "You're 40% set up. Finish your checklist." The assistant should send something like this instead:
Your Shopify store is connected and 14 orders came in overnight. Three are over 3kg and no rule covers them. Want me to send heavier parcels by courier? Just reply.
The customer replies from their phone: "Yes, and cap it at 25kg." The assistant makes the change as that customer. When they next sign in, the rule exists and a summary is in the thread. Onboarding stops being a sequence the user completes and becomes work the user delegates.
What we hear on onboarding calls
We have a lot of calls with software companies about exactly this, and three things come up that the onboarding guides rarely mention.
The hardest steps are in someone else's system. An IT asset management company told us its fastest route to the aha moment runs through integrations, and each one means a guide: go generate this key, register this app with Microsoft Azure or Intune. No assistant in your product should hold the customer's Azure admin rights. What we do today is narrower and works: the AI knows the exact screen in the other system, sends the customer straight there, and picks up when they come back with the key. That is the "waiting on your approval" line in the log above, made as short as it can be.
The vendor, not just the user, often wants the customer to learn. A hospitality software company we demoed to raised it first. Its product had grown so many settings and automations that new users did not know where to start and missed the features that would help them most. They liked an agent that could check a setting through their MCP server and answer correctly, but they wanted it to show customers where things live rather than quietly configure everything. We think that is right for many products. The choice between doing a step and teaching it should be made per step, by you, based on whether the customer needs to understand that setting to get value later.
The trial is empty in a specific way. The same IT asset management team pointed out that a prospect describes their setup on a demo call, then opens a trial full of generic data that reflects none of it. The context exists, in the CRM notes from the call. It just never reaches the product. An assistant that reads those notes and builds the account to match is doing onboarding work no checklist can.
Onboarding is a job, not a session
That email exchange is where most onboarding tools stop counting, and where the real work starts. A customer is not set up once. New teammates join, a second store gets connected, a rule stops covering this month's orders. Treated as a task, onboarding ends when the checklist does. Treated as a job, it has a state the AI keeps true for each customer: is this account set up for the way it works today? When that slips, the AI acts in the product, and loops in a person when the call is theirs. It is the same move we argued for with customer health scores: the number matters only if something is responsible for keeping it true.
That shift also changes who can be paid for it. A checklist is a feature. Getting every new customer set up, and keeping them set up, is work a customer would happily hand over. I made the longer case for the job as the unit software companies will sell in outcome as a service.
How to automate customer onboarding with AI
Getting there is less about choosing a model than about a few decisions you make about your own product.
Start with one activation milestone and write down every step between signup and it. I covered how to choose that milestone in SaaS onboarding best practices. For the account-level side, getting a new customer connected, synced and actually live, see SaaS implementation as a job. Then sort the steps into two piles: the ones that need a decision from the customer, and the ones that are just labor. The second pile is usually the bigger one. The assistant should ask for the decisions and do the labor.
Next, make the labor callable. Each step needs a tool the agent can use, and this is the real price of admission. If a step has no tool behind it, the assistant cannot do that step dependably, and you are back to narrating. Begin with additive, reversible actions like creating a rule or importing a list. Keep destructive ones behind a confirmation.
Be honest about where this does not apply. If a new user gets value from your product in two minutes with no configuration, you do not need any of it. If a decision carries real liability, a person should still make it. And agents make mistakes, so every action should leave a record the customer and your team can read.
Finally, change what you measure. Tours completed and questions answered describe the assistant's activity. The number that matters is accounts that reached the milestone. That is the test I would apply to AI agents for SaaS onboarding and activation: did the account end up set up?
Where Aimdoc sits
This is what we build. Aimdoc delivers the work your customers need, on top of your product. It connects securely to your product through your MCP or API, and to your systems, like your CRM, Slack and email. Every call runs as the customer, with their permissions, in their account. Setup is one of the clearest cases: a new customer asks in your app or replies to an email, and the setup gets done instead of explained.
Where your MCP has no tool for a step, a task definition teaches the browser path. That is a fallback, not magic, and the AI is only as capable as what your product exposes.
Keeping every new customer set up is also the kind of job you can offer as a service and sell to your own customers. What you charge is yours to set, through your own Stripe account; Aimdoc's prepaid credits only cover the AI's work.
Starshipit, a shipping platform in New Zealand and Australia, uses Aimdoc to help new users with setup and support, and their story is the fullest public example. In the first five weeks of that rollout, 23% of new users engaged it unprompted, and the median first conversation came 15 minutes after signup. If your trial opens on an empty dashboard, you can start on your own product without talking to us.