The fastest way to overpay for a website AI agent is to start by browsing features. On a recent call the founder had been comparing vendors for two weeks and was paralyzed between three, all overkill for a site that fielded maybe forty chats a day. The right question is which one will do the three jobs an agent actually has: read intent, pull from a real source, and hand off to a human when it's stuck. Everything else is decoration. I look at data plumbing first. Can the agent ingest the help docs, product pages, and Notion notes the business already lives in, or does it want a fresh schema? The teams that ship in a week accept files, websites, and Notion out of the box; the ones that ship in a quarter picked a tool that needed retype everything. Actions come next - Zendesk, Stripe, a calendar - because every custom handoff is another week and another invoice. Then guardrails before clever features: identity verification, escalation routing, and the rule that lets a human type "I'll take this" and have the agent step aside. Outbound campaigns and AI Replies with Human Takeover are worth adding, once the plumbing holds. A vendor that skips the guardrails to show off the toys will cost you a refund dispute later. Picking an agent without that conversation is how a quiet line item turns into a budget meeting.
Picking an agent that fits the business
I lost a Monday to a vendor demo that did everything except answer the question I had. The founder had been comparing tools for two weeks and was now paralyzed between three, all of them overkill for a site that fielded maybe forty chats a day. We spent the first hour untangling what the agent actually had to do, which is the same three jobs every agent has: read intent, pull from a real source, and hand off to a human when it gets stuck. Once that list was on the whiteboard, the vendor shortlist got a lot shorter.
Training the agent on what the business actually knows
The fastest path to a useful agent is to skip the chatbot builder's "type your FAQ here" field and feed it the messy pile the business already has - help docs, the Notion page the founder wrote in 2022, the product catalog, the return policy nobody has reread since launch. On the e-commerce shop I work with, that meant dropping in the catalog, the shipping PDF, and a two-year-old FAQ Notion the owner had half-forgotten, and watching the agent start answering "do you have this in a medium?" correctly the same afternoon. The boring content wins.
What I feed it first
I start with the three documents a new hire would grab on day one: the shipping policy, the return window, and the product catalog. Then the long-tail content - sizing notes, the founder's old blog posts, the FAQ Notion that everyone pretends doesn't exist. If a new employee needs it, the agent needs it.
What I delete before training
Conflicting policies, duplicated help articles, and any page that still says "coming soon" six months after launch. The agent treats all of it as gospel, and I would rather not explain why the bot quoted two return windows in the same chat. I have watched that exact moment land on a Tuesday.
How I know the training stuck
I run the same six test questions the client used during the sales demo, at the end of the week. If the agent handles the shipping, return, and stock-availability trio without flinching and politely declines the refund question it isn't allowed to touch, the training took. If it invents a return window, I haven't fed it enough - or I've fed it too much conflicting copy. Both happen, and both are fixable by the next morning.
The first seven days
I watch the escalation queue, not the deflection rate. Deflection is a vanity number that goes up when the agent guesses, and guessing is what refunds are made of. The honest signals are how often the agent hands off cleanly, how often it cites the right policy page, and how many chats the human team had to reopen because the bot said something it couldn't back up. If those three numbers hold steady for a week, the launch held.
When to stop adding features
The temptation after a clean first week is to bolt on the outbound campaign, the proactive offer, the survey follow-up. I wait. Adding features during launch week is how a working agent becomes the thing that quoted two return windows in the same chat. One feature at a time, a week apart, and only after the escalation queue has stayed quiet. Boring discipline, but it is the difference between an agent that survives the month and one that becomes a Tuesday story.