Strip away the buzzwords and AI customer support is a plain thing: software that answers customer questions from your own content, handles a few real tasks like order lookup and booking, and hands everything else to a person. That is the whole product. It is not a robot employee, and any vendor selling it as one is setting you up to be disappointed by a very confident FAQ.
I set these systems up for customers most weeks, and this guide is the version I give people on a call. What the technology does well right now. Whether your team needs it at all (maybe not). How to roll it out in two weeks. What it costs under the three common pricing models, and where it goes wrong when it goes wrong.
What it does well today (and what it still can't)
The strong cases first.
An AI agent trained on your content answers instantly, at any hour, in whatever language the visitor writes. It does not get tired of a question, and it does not get slower on the fortieth repeat. If most of your inbox is "what is your return window", "do you ship to Ireland", and "how do I change my plan", a grounded bot clears that layer from day one. The language part is easy to underrate: a visitor writes in Portuguese, the bot answers in Portuguese, and you did not hire anyone.
Lookups are the second win. Connected to your store or your calendar, the bot checks an order status, reschedules a booking, or fetches an account detail. This is the point where AI support stops being a smarter FAQ page and starts closing conversations end to end. "Your order shipped Tuesday, here is the tracking link" ends a conversation. A policy quote does not.
The third win is quieter: triage. Even when a human needs to take over, the bot has already asked what the problem is, collected the order number, and captured an email address. Your first human reply starts from context instead of "could you tell me a bit more?"
Those three cover the heavy share of real support load, though rarely the interesting share.
Now the other half, which vendors mention less.
The model cannot make judgment calls. It should not decide whether a specific customer deserves a refund outside policy, and it cannot repair the relationship with someone who is genuinely angry. Those conversations need a person, and getting them to a person fast is part of the product doing its job.
It also cannot answer anything that is not in your content. A grounded bot is a mirror of your documentation. Thin docs produce a thin bot. And an ungrounded bot, one that fills the gaps by guessing, is your brand improvising in public. That failure mode is big enough that we wrote a separate post on why chatbots make things up, and I would read it before signing anything.
Do you actually need it?
Here is the part of the call where I sometimes talk people out of buying.
If you get fewer than, say, a hundred conversations a month, volume is probably not your problem. Scattered docs, unclear policies, or a slow inbox usually are, and those are cheaper to fix directly. Write the five answers you repeat most often into a public FAQ. Set up a shared inbox and a same-day reply habit. That may genuinely be the whole fix, and it costs you an afternoon. A bot layered on top of thin content just automates thin answers, faster.
The case for AI customer support strengthens as specific pressures appear. Volume is the obvious one: repeat questions crowding out work only you can do. After-hours traffic is the second: if customers buy while you sleep, every morning opens with a backlog and the people who asked at 11pm have been waiting the longest. Language is the third, since a bot answers in the visitor's language without a hiring plan. And repetitiveness is the quiet qualifier: if fifteen question types cover most of your inbox, that inbox is largely automatable. If every conversation is a bespoke negotiation, it is not, and no amount of AI changes that.
There is an awkward middle, of course. At a few hundred conversations a month a fast inbox still copes, and plenty of teams run that way, at the cost of somebody answering chat at dinner. At that size the deciding factor is usually somebody's evenings and what they are worth. I have set up bots for teams of two whose only goal was to stop checking chat after 7pm, and that counted as success.
A practical test before any trial: tag one week of conversations by hand. If more than half are questions your website already answers, or should answer, the math works. If not, fix the docs first and come back when the volume does. For how to read those numbers without fooling yourself, see customer service metrics that matter.
The anatomy of a good setup
Every setup I do has the same five parts. The tools vary. The parts do not.
Knowledge
What the bot is allowed to know: your website, help docs, policy pages, uploaded files, and the Q&A pairs you add for anything not written down anywhere. The mistake to avoid is blind indexing, where a crawler swallows your whole site (navigation menus, legal boilerplate, a 2019 blog post about a discontinued product) and you find out from a customer. Pick a tool that shows you exactly what it indexed and lets you exclude pages before anything goes live, then refreshes on a schedule so answers do not drift stale. We wrote up the whole process in how to train an AI chatbot on your website.
One more thing to feed it beyond the site: your real email replies. The way you actually answer "can I change my shipping address" is usually better than the way the policy page says it, and pasting those answers in as Q&A pairs is the fastest quality win in the whole setup.
Behavior rules
Scope and tone. Which topics the bot must refuse (payment disputes, legal threats, anything medical), how it should sound, and how strictly it sticks to your content. If you are weighing this whole approach against a scripted flow builder, the rule-based comparison covers when each one wins. Good rules are dull and specific. "Do not discuss pending disputes; offer to connect a person" beats a paragraph asking the bot to be helpful and friendly.
The welcome experience
The first screen decides whether anyone types. It should say it is a bot, say what it can help with, and offer three or four quick questions that map to your actual top questions, plus a visible way to reach a person. One screen, no theatrics, no fake typing indicator from "Emma". A gym's version: "I'm the front-desk bot. I can help with class times, memberships, and freezes, or pass you to the team." Two lines, scope set, exit visible.
Handoff
The seam between the bot and you. When does a human take over (the visitor asks, the bot cannot answer, the topic is sensitive), where does the conversation land (shared inbox, Slack), and what does the visitor see at 2am when nobody is online? For small teams the honest answer is asynchronous: collect contact details, promise a reply, keep the transcript so nobody repeats themselves. The full decision logic is in live chat vs chatbot.
Actions and integrations
The difference between answering about things and doing things. Order lookup against your store, bookings against your calendar, a lead saved into your CRM, and for anything sensitive, an approval step so a human signs off before the action runs. In practice these matter more than any benchmark of answer quality, because resolved conversations come from doing.
The two-week rollout
You can index a website in minutes. Trust takes two weeks. Here is the plan I run with customers, day by day where it matters.
Before day 1, pick the number that will decide the verdict: resolution rate, first-reply time, hours saved per week, or after-hours conversations handled. One number, written down, so that week-3 you is not grading on vibes.
Week 1: index, break it, soft-launch
Days 1 and 2: point the bot at your site and your key documents, then review what got indexed and cut the junk. Expired promotions, superseded policies, the press release from 2021. Add Q&A pairs for the things you type into email every day but never put on the site. Every team has a dozen of these and nobody notices until the bot cannot answer them.
Day 3: try to break it. Ask thirty real questions pulled from your inbox, then get hostile. Ask things your content cannot answer. Request discounts. Demand promises. Ask about competitors. You are testing the refusals as much as the answers here. A bot that says "I don't know, want me to connect you to the team?" passed. A bot that improvises failed, and you want to learn that on day 3, in private.
Days 4 and 5: fix what the breaking found. Add the missing pages, tighten the rules, correct the wording. Then soft-launch on low-stakes pages: help center and docs first, homepage and pricing later, once it has earned the promotion.
Week 2: read everything, then expand
Days 6 through 10: read every transcript, every day. Twenty minutes with coffee. You are looking for three things: wrong answers to correct the same day, questions where the bot had no source (each one is a to-do), and whether people who wanted a human found the door. This ritual is worth more than every launch decision you made in week 1, and the teams who skip it are the ones who write to me confused a month later.
In parallel: wire handoff to where your team actually lives, turn on lead capture if it fits your business, and expand the bot to more pages as the transcripts justify it. By day 14 you have real numbers against whatever you decided success meant, and you can keep, tune, or kill it without a single vendor-comparison meeting.
What AI customer support costs
Three pricing models dominate the market, and they behave very differently in a busy month.
Per resolution. You pay for each conversation the AI resolves. Intercom's Fin charges $0.99 per resolution, and Zendesk and others price AI the same way. It sounds fair, pay for outcomes, until you run the math against growth. Say you get 1,000 conversations a month and the bot resolves 60 percent of them: about $594 that month. Now a product launch triples your traffic, and the bill lands near $1,780 in the exact month everything else got expensive too. Illustrative numbers, but the shape holds at any scale: the meter runs fastest precisely when the tool works best, and the vendor defines what counts as resolved. I have opinions about that arrangement, and so does Siddharth. The per-resolution pricing trap is the long version.
Per seat. Pricing carried over from helpdesk software: a fixed price per human agent, say $50 to $100 each. It keeps costs flat as conversations grow, which is something. But it prices the wrong axis. The point of AI support is that conversation volume stops mapping to headcount, so as the bot answers a larger share you keep paying per human for software whose job is to need fewer humans on the routine layer.
Flat plans with conversation pools. A fixed monthly price covering a set number of conversations. Predictable in busy months, no incentive for anyone to inflate resolution counts, easy to budget a year out. This is what we chose for Hey Support (the free plan covers 50 conversations a month, paid plans pool from 1,000 to 20,000), so discount my bias accordingly, then do the busy-month math for each model yourself before you sign anything. Our numbers are on the pricing page.
Whichever model you are quoted, run it against three months: a quiet one, a normal one, and your busiest on record. The quiet month tells you the floor. The busy month tells you the truth.
One cost no pricing page lists: your attention. Whatever the model, budget the two weeks above plus a standing twenty minutes a week. A neglected bot decays quietly, and the cheapest plan in the world does not fix that.
The risks, and how to manage them
Wrong answers are the risk everyone asks about, and the fix is architectural, not aspirational. Grounding, so the bot answers only from indexed content. Citations, so every answer shows its source. Refusals, so below a confidence bar it says it does not know and offers a person. If a product lacks those three, no prompt begging it to be truthful will save it.
Tone drift is subtler. The bot sounds slightly off-brand: too chirpy, too formal, apologizing three times per answer. You fix this with explicit written instructions, not by hoping. Give it your voice as rules ("short answers, no exclamation marks, link the page you cited") and check the transcripts weekly.
Privacy deserves plain questions before you sign. Where is your content stored? Is it used to train models? Can you delete it, and does deletion actually delete? Your content is your business. A vendor should answer all three without hedging; at Hey Support the answers are your workspace, never, and yes.
Transparency is the easy one to get right. Say it is a bot. In some places disclosure is required by law, and everywhere it is the difference between a customer feeling helped and feeling tricked. No fake human names, no fake typing delays, no stock photo of an agent who does not exist.
And the risk nobody names in a sales call: over-automation. Once the bot clears most of the queue, hiding the human option entirely starts to look efficient. Resist it. Deflection without an escape hatch teaches customers to leave quietly, and the metric that celebrates it will look great right up until the reviews arrive.
Ten questions to ask any vendor
Take these into every demo, including ours. Plain questions get plain answers, or they get squirming, and both are information.
- What does the bot do when the answer is not in my content?
- Can I see exactly what was indexed, and exclude pages before going live?
- Does every answer cite its source?
- Can it invent a URL, or does it only share links that exist in my content?
- How does a customer reach a human, and how many clicks does that take?
- Is my content used to train models, and can I delete it completely?
- What would a month at three times my normal volume cost me?
- If pricing is per resolution, who defines resolved, and can I audit it?
- Which actions can it take (order lookup, booking, lead capture), and which require human approval?
- What do I see after launch: transcripts, knowledge gaps, feedback on answers?
The answers to 7 and 8 will teach you more than the demo will. A vendor who squirms at "who defines resolved" has told you how the invoice works.
Two honest weeks
You do not need a procurement quarter, a scoring matrix, or a committee. You need your real content, your real questions, and two weeks of paying attention. Index the site, try to break the bot, soft-launch, read every transcript, then look at the number you picked on day one and decide. Two honest weeks beats a quarter of procurement theater, and most tools have a free tier that covers the whole experiment.
The verdict at day 14 is usually one of three: it works, so keep it and stop thinking about it; it half-works, so tune the knowledge for another two weeks; or your volume was never the problem, in which case you learned that for free.
If you want to run those two weeks on Hey Support, start on the free plan. Fifty conversations a month is enough to find out.
