Raise your agent

A support AI you can't shape is a gamble: it sounds like someone else's brand, improvises your policies, and you find out from a customer. Today Iris ships the other kind — an agent you raise.

A character you choose

Pick a tone — friendly, neutral, professional, matter-of-fact, even lightly humorous — and how long answers should run. Add words it should prefer and words it must never use. Then write guidance the way you'd brief a new teammate: "If the plan isn't clear, ask which plan before answering." "Never discuss legal specifics." "Hand any payment dispute to a human." Plain language in, behavior out.

Exact answers, exactly

Some sentences you don't want paraphrased — refund policies, guarantees, anything legal approved word by word. Pin them: when a customer asks that question (however they phrase it), your stored answer is sent verbatim. Not a summary. Not a remix. Your words.

Knowledge beyond your help center

Feed it snippets — internal truths the AI may use but customers never see listed. Point it at a page on your website and it reads it. Upload a PDF or DOCX and it learns it. All of it stays inside your workspace; none of it is ever cited publicly.

Rehearse before you ship

Here's the part we're proudest of. Before a change goes live, rehearse it: Iris replays your last real customer questions under the new configuration, and an AI examiner grades every answer — was it grounded? did it follow your guidance? did it hand over what it should? You get a score and the failures, question by question. Simulated personas — the frustrated refund seeker, the confused first-timer — are one more click.

Every change is a version

Nothing applies silently. Every save is a numbered version recording who changed what, and any previous version is one click from being live again. The ledger shows replies, closes and satisfaction per version — so you can see whether v12 actually beat v11, in outcomes, not vibes.

Or just tell it

All of this has forms. None of it needs them. Open the Trainer and talk: "Be more professional, and never handle refunds yourself." Your agent proposes the change as a diff — what's being removed, what's being added — and waits. Apply it, rehearse it first, or throw it away. It's the way you'd train a person, pointed at software.

Name it, feed it, come back tomorrow

Give your agent a name and a face — Maya with a photo beats "AI Assistant" every time, and both live in the same versioned character file as its tone. Feed it files right in the chat: drop a price sheet and it learns every row, share a voice note and it's transcribed into knowledge, paste a screenshot and the trainer offers to remember the facts inside it — with your approval, like everything else. And every training conversation is a session: reopen one and continue exactly where you left off; start fresh and it remembers nothing. Your training history reads like a log, because it is one.

The part that compounds

Here's what separates raising an agent from configuring one: it learns from the work your team already does. Every time a teammate answers a customer differently than the AI would have, that's a correction — and the Trainer notices. Once a week it hands you a short review: "Your team overrode me six times, and I stayed silent on these five questions. Here are the snippets and rules that would have fixed them." One tap each, every one a diff you approve. Pin the answers it must always get right into a golden test set, and every change re-takes that exam automatically — so fixing one thing can never silently break another. Five minutes a week, and the agent you raised keeps getting sharper on its own.

Support AI you can shape, test, and roll back — and that improves while you sleep. That's the difference between switching an agent on and raising one.