Real-world HubSpot Customer Agent deployment lessons learned

HubSpot's Customer Agent takes an hour to switch on and a month to trust. Four deployments in, here is what we learned about knowledge, cost, reliability and ownership, and the ten decisions to make before you click "Start Free Access".

Here's a recording of my webinar and you can download the slides, sample regression test questions and go-live questionnaire.

In the last few months we have built four HubSpot Customer Agents. Two are live, two are in development and ready to go. One was for a very large company with a lot of traffic and a complex web estate; the others were smaller, and one was our own. This post is the condensed version of that experience: the hard-won bits, so your deployment goes right from the start.

The autopilot problem

The fundamental challenge with Customer Agent, as with most AI, is that the first demo is mind-blowingly good. In an hour you can point it at your website, especially if HubSpot hosts it, let it spider your knowledge base, and watch it answer reasonable questions reasonably well. It looks like a lot of AI. It is extraordinarily plausible.

It is not enough. Plausible is not the same as right, affordable, doing the job you expected, or doing it reliably.

I spent years training as a pilot: private licence, commercial licence, night rating, instrument rating, exams, books, hours in the cockpit. Then you reach the point where you fly an aircraft with an autopilot, press a button, and the plane flies itself. On a good modern aircraft the pilot's job is roughly the last 200 feet of the landing and the first 200 feet after take-off. Most of the time it works beautifully. And sometimes it does something strange, and you find yourself asking: why is it turning right?

That moment, "what on earth is it doing?", is the moment you want to avoid with Customer Agent. It is representing you. It is on your website. People know they are talking to an agent, but they still expect it to be a valid representative of the company.

My wife recently bought a service for the house over the phone from an AI agent. The experience was excellent: it walked her through the options, quoted £150, took her card details, and then said: "So we'll be charging your card £450." That one moment cost the company the deal and our trust, and we had used them before. Their AI did a bait-and-switch on price. You do not want your Customer Agent doing anything like that.

So this post is about four jobs: deciding what Customer Agent knows, making sure it does what it is meant to, making sure it earns its keep (because you will start paying for it), and working out who owns every piece.

Say which chatbot you mean

Customer Agent is one of several things inside HubSpot that sound like a chatbot. Get the terminology straight before you plan anything.

• Chatflows are version one. They manage the thing that pops up, the consent text, the buttons. They are deterministic: the same input gives the same output every time. They come with your subscription and cost nothing extra.
• Customer Agent is the AI. It is probabilistic: ask the same question twice and you may get two different answers. If you have set it up well they will be very similar, but it is the same as asking ChatGPT twice. It consumes credits (more on that below).
• Reply recommendations use the same capability to draft answers for human support reps in the help desk. Useful, but a different tool.

The two work together. Even when a Customer Agent is live, it talks through a chatflow window. You can let the agent be the only thing the visitor sees, or you can run a traditional chatflow with buttons (book a meeting, get support, and so on) and make "talk to our AI assistant" one of the options. At that point the agent plugs in like a person taking over a live chat. It is a kind of Turing test.

That second pattern matters because once the agent is live it is billable, and you want to be careful about who you make that billable activity available to. I covered the wider shift to credits-based pricing in HubSpot Unbound 2026: 134 announcements and one urgent question; this post is the practical sequel.

Knowledge is the product

Customer Agent can learn from several places: your website, HubSpot knowledge bases, past support ticket answers, uploaded documents (PDFs, Markdown files), short answers, and documents you mark as context the agent can use but must not cite or link, such as a price list.

The temptation is to point it at the site and say "spider everything". We did exactly that for one client: everything except the foreign-language pages. Four and a half thousand pages. It slowed the agent down and diluted the knowledge. More pages did not mean better answers.

Take Articulate as a smaller example. We have about 450 blog posts, 40 or 50 pages, 120-odd webinars and 20 or 30 gated pieces. That is a lot of stuff, and a lot of it is not useful to someone asking a question in a chat window. Two hundred past events give a visitor nothing.

So the real work was curation, done outside Customer Agent with our own tools:

• A curated crawl with path exclusions (events, press, locale prefixes, gated content, investor and legal pages), which took 4,500 pages down to about 950.
• A glossary of terms scraped from the site and ingested as one Markdown file, so the agent has definitions to draw on without a crawl diluting them.
• A pass through what was left to remove anything out of date. Contradictory content is worse than missing content: the agent may cite either version.

The exclusion rules got so complicated that I built a small simulator: given this set of inclusions and exclusions, which pages will HubSpot actually ingest? Expect to allocate real time here, before you get anywhere near the trial.

Crawling fails quietly

Sometimes the agent is not ingesting what you think it is. On one large site a web application firewall was silently blocking HubSpot's crawler. No error; just thin answers. Fixing it meant several round trips with IT security and web operations, and it was blocking the launch. If you run a scan and the content is not there, that is the first place to look. (The same bot-protection problem hides sites from AI search engines; see Is your website invisible to AI search?.)

Other things to know: HubSpot will not ingest pages marked 'noindex' even when you want it to. Content on other platforms needs separate permissions and often a separate team. HubSpot's interface for all this is, charitably, adequate.

Plan maintenance windows

HubSpot re-spiders every seven days. If you change what it spiders, especially external sites, you trash the existing knowledge and start again. On a site with thousands of pages that is hours, during which the agent is answering from the first few hundred pages it has seen. The answers are plausible and wrong. Treat scope changes as planned changes, and use short answers as a hot-fix layer for new content in between.

Guidelines are suggestions. Actions are commitments.

This is the biggest technical lesson from all four builds.

Customer Agent is not deterministic, and internal stakeholders need to understand that. As the implementer I have spent a fair amount of time saying, "That's AI. I didn't write that answer." The levers you have are limited: the knowledge base; a handful of guideline boxes (how the bot should talk, how it hands off to humans, safety rules, and a fourth box I use for privacy protections such as "if someone starts giving you card details, tell them to stop"); and actions. The guidelines are your root prompts, they are capped at 2,500 characters, and small changes can produce large downstream effects.

In every build the valuable behaviour was the same: when a visitor shows buying intent (asks about pricing, a demo, or to speak to someone), offer a meeting link. Every time. That is lead generation; it is the thing the client is paying for.

Putting that rule in the guideline boxes did not work reliably. Handoff instructions are consulted late, often after the agent has already fallen back to "sorry, I can't help with that". HubSpot's prospecting action looks exciting on paper but is prescriptive: it wants to ask qualification questions first, and on one client's CRM it raised GDPR and permissions questions we were not comfortable with.

What worked was a custom action used as a tripwire. Custom actions are designed for jobs like "enter your order number and I'll look it up". I wrote one whose trigger is buying intent, with no inputs, that fetches a static file containing the meeting link and serves it. Because actions sit at the top of the agent's interpretation order, with a higher priority than guidelines and handoff rules, it fires reliably. It is a hack, and I say so. But it stopped buyers being sent to the default "I can't help" answer.

A note to HubSpot: ungate custom action development, allow a little programmatic information gathering inside actions, and ship a simple "offer a meeting" action that does not insist on qualifying people first. Until then, this is a reasonable workaround.

Filter the noise before it costs you

Customer Agent generates value in two ways: support deflection (answering a question before it reaches a human) and lead generation (helping people buy). Everything else, nicer navigation, helping people find content, may be good experience, but those two are the ones with money attached. Costs avoided or revenue added are what excite business leaders.

So before building anything we analysed a year of contact-form submissions for one client. A very large proportion were job seekers, gig seekers, vendors and quasi-support tickets. If you have ever read a contact-us inbox, you know the pattern. You do not want to pay Customer Agent to tell people how to apply for a job.

In every deployment we therefore put a deterministic chatflow in front of the agent: a short menu (work for us, book a sales meeting, get support, and "talk to our AI assistant"). Only that last button summons the Customer Agent. It is filtering out a lot, though not all, of the noise, and it costs nothing.

You pay per resolution, not per message

HubSpot's billing model takes a moment to get your head round. Customer Agent charges for conversations it resolved: anything that did not end up in the lap of a human being. A handoff to a person is free. In effect it charges you for the human conversations you avoided, which is odd but makes a kind of sense. It also means a visitor who takes a link and wanders off is a billable resolution, and so is a booked meeting.

The mechanics, as of October 2026 (check the HubSpot knowledge base; this changes):

• Fourteen days free, then credits. Most portals have a monthly credit allowance, and at roughly 50 credits per resolved conversation that covers around a hundred conversations a month if nothing else is using them. For many HubSpot customers that is plenty.
• A 72-hour window. A visitor can come back, so a resolution is confirmed up to three days after the conversation. Credit burn lags conversation volume. Build that into your monitoring.
• Spend limits do not take the agent offline. You can cap credits, but hitting the cap does not switch the chatflow off. Visitors keep expecting an AI and get nothing. Monitor the cap and either raise it or take the chatflow down.
• The clock starts when you click "Start Free Access", not when you deploy. Do not click until you are ready to tune hard. Plan the fourteen days as a sprint: weekly review of transcripts and the questions people are not asking, deflect more, fix routing, widen the throttle step by step.

It is a huge mistake, and I have watched it happen, to switch it on and say "right, we've got a chatbot, done".

It is software. Treat it like software.

Everyone is into vibe coding and writing things in plain English, but a Customer Agent is software. You are programming it with prompts and knowledge instead of code, and it represents your business the way your website does. So the disciplines that go with software apply.

Test it with a bank of questions. It is not deterministic, so one pass proves nothing. Build a regression bank covering your services, buying-intent phrasings, the pre-screen, your red lines and the handoff. For live trials ours now run to 150 or 200 questions and growing. Re-run the whole bank every time you change the prompts and every time you change the knowledge. You will think version two still behaves like version one. It does not; small changes have large effects. Test mode is free, so there is no excuse. The sample bank is in the PDF at the top of this post.

Test on the live site without going public. Add ?test_chat=true to any URL and a chatflow set to that targeting rule appears for you and nobody else. Share the link internally and stakeholders can try the agent on real pages with real knowledge, at no cost.

Version control. Every iteration of every prompt lives in a GitHub repo. HubSpot keeps no history for these fields; without a repo, a bad edit on a Friday afternoon is unrecoverable.

A bug tracker and a release cadence. We track feedback and bad answers, and we release on a schedule ("new version Saturday morning") rather than tweaking live. It is very easy to make a mess of your week by doing this casually.

It touches more owners than you think

Think about the stakeholders you would involve when launching a new website. Who needs to be asked, who informed, where does the knowledge come from, is there legal input, security input, technical input? A Customer Agent needs all of them: the website team, support and docs, legal (transcript retention, AI disclosure), security (that firewall), sales, and whoever administers HubSpot. One client has employees in Germany, which means works councils and consultation. It is more complicated than it first appears, and planning it early is critical.

The sales handoff deserves its own mention. If the agent is going to offer meeting links, the people on the other end need HubSpot calendars connected and configured properly, with sensible defaults. We did a lot of that set-up work. It is a whole other webinar.

And be ruthless about scope. One language first (you cannot QA what you cannot read), one mission, one handoff. Live chat is a programme (seats, hours, routing, the no-one-available experience), not a toggle. Defer it. It is not a switch; it is a process.

Chatbots are not the only option

A penultimate thought. I have seen some very good AI interfaces that are not "type a question, get an answer" chatbots. Ubiquiti's support assistant for its networking kit is a nice example. We are building something similar, codenamed Marlinspike, where a slightly constrained conversation drives search across a curated library of our content. It is still supported by a lot of AI in the background, but the shape is different, and you could imagine it working well for support. The same knowledge problem sits underneath both: curate the library first and whatever you put on top gets better. There are other options; a chatbot is one of them.

The demo takes an hour. A working AI chatbot takes the next month.

Here is the honest sizing. The demo version of Customer Agent takes an hour. A plausible go-live candidate for a small business is a day's work. If it is mission critical, enterprise scale, or there is more than one stakeholder, you are looking at a month of calendar time and days of work to get it right, and then you need to monitor it like a newly launched website.

Below are the ten decisions to make before you click "Start Free Access". Tick the ones you have actually made, not the ones you plan to make.

If you want a second pair of eyes on your deployment, email me at matthew@articulatemarketing.com or book a Customer Agent readiness call. I am happy to talk Customer Agent, chatflows and HubSpot automation all day.