AI Support Agent

An assistant that answers your support tickets and live chat from your own documentation, cites the passage it used, and refuses anything it cannot ground. It ships switched off and is promoted one step at a time.

4 min readUpdated 6 August 2026ai, support, tickets, knowledge-base

The addon reads your support conversations and answers them from your knowledge base and the product documentation. Every answer carries the passage it came from, so a customer can check it and your team can audit it afterwards.

What makes it safe to switch on is what it refuses to do.

Every answer is assembled the same way: the customer's question retrieves passages from your own documentation, a grounding gate decides whether there is enough evidence to answer at all, and only then does the model write. With too little evidence it hands over to a person instead of guessing.

The shape above is the whole product. The model is never asked what it knows about your platform — it is handed passages from your documentation and told to answer from those. That is why the quality of your docs is the quality of your support, and why a page that is out of date becomes a confident wrong answer rather than a missing one.

What it refuses

Most of the engineering in this addon is in the answers it declines to give. An assistant that answers everything is an assistant that invents your refund policy.

  • It will not answer below a confidence floor. If retrieval cannot find a good enough source, it hands the conversation to a person and records the question as a gap for you to close.
  • It will not state your own numbers from somebody else's documentation. Fees, limits, review times, supported countries and refund policy can only be answered from articles you wrote. It will never lift a figure out of the general product docs and say it in your voice. That gate applies in every language, not only English.
  • It will not name the vendor. Your install is white-labelled and the assistant's answers are part of your brand. It writes a page's name and its in-platform path, never an external link.
  • It will not act on an account unless you let it. Account tools are read-only by default, and the actions it may take are an allowlist you fill in yourself — empty at install.

The ladder

It ships off, and it is promoted one rung at a time. Nothing here happens because you installed the addon.

Mode What happens
COPILOT The assistant drafts. A human reads it and presses Send.
AUTO_TICKET It answers tickets on its own. Live chat still goes to a person.
AUTO_ALL It answers tickets and live chat unsupervised.

COPILOT is the shipped default, and the console will not let you skip ahead casually: promotion is measured against how much your agents actually changed its drafts on your own tickets. See Autonomy.

What it costs you to run

You bring the AI service — either your own key or the managed connection — and you set a daily and a monthly ceiling in dollars. When a ceiling is reached the assistant stops answering and escalates instead; it does not degrade quietly or overspend. The defaults are $5 a day and $100 a month.

The bill falls the longer it runs. The nightly harvest turns the replies your agents wrote by hand into draft articles for review, so the questions you answer twice become questions that answer themselves. And once one of your own articles answers a question outright, the assistant can serve it verbatim with no model call at all — that last one ships off and has no control on any screen in this build, which is covered in Cost and budgets.

The order to set it up in

Install

Connect an AI service, create the persona, and get the first answer out of it.

The console

Nine screens, and the order an operator actually uses them in.

Knowledge

What it is allowed to answer from — and why a fresh install answers nothing.

Autonomy

Copilot, then tickets, then everything. What each rung requires.

What your customers see

The customer-facing side lives with the rest of the platform's help, because that is where a customer reads it:

Requirements

Bicrypto core. No other addon is required. If you have the FAQ & Knowledge Base addon installed, your published FAQ entries are indexed alongside everything else.

The addon also ships a documentation pack per product, and two rules decide how much of that reaches a customer. Both are worth knowing before you conclude the assistant is ignoring an addon you paid for.

Only products registered as extensions are indexed. On the shipped default, aiSupportScopeToInstalledProducts matches each pack's slug against your enabled rows on Admin → System → Extensions, and a pack that does not match is not merely skipped — anything a previous run indexed from it is withdrawn. Seven products you can buy have no extension row at all: Solana, Tron, Monero and TON, and the Binance, KuCoin and XT provider connectors — blockchains are rows in ecosystem_blockchain and exchange providers are a setting, so neither ever produces an extension row to match against. Their packs ship with the addon and are excluded on every install, including yours. Switch that setting off on Settings → Retrieval to index everything shipped instead — see Knowledge pipeline.

And only customer-facing pages can be quoted to a customer. Audience is a hard filter at retrieval rather than a ranking nudge, so a page written for you is unreachable from a customer's question — by design, and it is what stops your fee settings being read back to somebody's customer. Most addons ship a customer help section; the ones that are purely operator tools, such as AI Market Maker, the Binary AI Engine and MailWizard, ship none, so their pack contributes nothing to a customer answer even when it is indexed.

Neither rule is a silent failure. A question those rules leave uncovered escalates and lands on Gaps like any other.