By the Onsites AI team · Last updated · 5-minute read
An AI copilot inside a help desk is not a chatbot on the website; it lives in the reply box, next to the human, and its whole job is to shorten the distance between "conversation open" and "useful answer sent." In Onsites AI that means four actions — draft, suggest, translate, summarize — priced at prepaid credits that only burn when you actually use them. This guide walks through what each action does in the flow of a real conversation, what it costs, where copilots save the most time (spoiler: not where you expect), and the two habits that separate teams whose AI output reads like a colleague from teams whose customers can smell the robot.
Draft (3 credits) is the flagship: with a thread open, Copilot reads the conversation plus the workspace context — the CRM record, the handbook, history — and produces a first reply you edit rather than a blank box you battle. The realistic saving is the hardest 60% of any answer: finding the right facts and opening the right structure, after which the human adjusts specifics and hits send. Suggested reply (6 credits) works the front line: for a new inbound message it proposes the next response inline, which turns triage time into answering time on high-volume, repetitive mail. Translate (3 credits) turns the multilingual thread from a chore into a click, keeping business register rather than literal translation — the difference between "we will dispatch tomorrow" and grammatically-correct-but-strange. Summarize (6 credits) compresses a 40-message thread into the two lines a teammate needs at a handoff — the feature that quietly saves the most organizational time, because it converts long histories into portable context for every person who touches the case next.
Credits are prepaid, never expire, and price AI at roughly a cent each: a $10 pack carries 1,100 credits, $20 carries 2,400, $50 carries 6,500. At those rates a year of aggressive copilot use costs less than one per-seat SaaS subscription month: teams of two to five typically burn 500–2,000 credits a month depending on volume and appetite — that is $5–25 of runway per person covered, with a spend limit enabled by default so enthusiasm cannot surprise the card. The worked example worth internalizing: 200 conversations a month, drafts on 40% of them (240 credits), translations on a quarter (150), summaries on escalations (120) — roughly 500 credits, one $10 pack, every few months. The monthly cost walkthrough prices busier teams, but the shape holds: copilot spend scales with use, not with headcount — six agents who barely click the button cost the same as two who work it hard.
The copilot's ROI is uneven, and knowing the skew directs the habit. Biggest wins: first replies on routine tickets (draft turns ninety seconds of composition into fifteen of editing), multilingual support (the difference between a hesitant reply and a confident one is more than the credit cost), and thread summaries at handoff — the quiet compounder that makes knowledge transfer free. Modest wins: polishing prose in short replies, where the human version was already fine. Net negative zone: forcing AI onto novel, sensitive or escalated cases where the human context is the whole task — the button exists to be ignored there. The teams that report the best numbers do one thing differently: they treat drafts as proposals with a reading step, and they let the obvious ones through untouched while reading the ones whose thread has money, emotion or contract in it.
Copilots respond to instruction quality like very fast interns: the more of your intent in the prompt, the less editing after. Three instructions carry most of the value. Say what the answer must contain — "refund the difference, mention 48-hour posting, no apology theater" produces a sendable draft; bare "help me reply" produces a generic one you will rewrite. Say who it is for — "for a long-term wholesale buyer" changes temperature and formality more than any tone preset. Say what it must not do — "do not quote policy, do not offer a replacement" saves a whole editing pass on money threads. These micro-prompts cost nothing (the draft is 3 credits whether directed or not), take ten seconds to type, and are the actual skill behind "AI drafts well": the model was always capable — the team finally started steering it.
Habit one: read-before-send is a role, not a mood. The flow is AI proposes, human disposes — every outbound message has a named sender, and the desk records it (audit trails matter for exactly this). What you must not do is auto-send anything to a customer; the moment a draft goes out unread, the copilot has become the support team, with your brand as its tone of voice. Habit two: feed it truth or watch it confabulate. A copilot is only as grounded as the workspace behind it — a lean handbook, current CRM fields, connected history. Given no facts it will still write you an answer, confidently, with plausible details that are entirely invented; given facts, it cites them (Harness answers carry visible sources for this reason). Teams that feed their AI real policy text get answers consistent with their own prices and terms; teams that skip that step get grammatically perfect fiction.
Depth comes in three sensible rungs. Rung one, free and permanent: skip AI entirely — the desk is complete without it, and nothing in the free tier pushes you toward it. Rung two, the copilot layer: one $10 pack on the shelf, drafts and translations on the repetitive half of the queue, the reading habit above — this is where 90% of teams should live, and the watermark disappears with this first payment too (any first payment removes "Powered by Onsites AI" permanently). Rung three is the assistive machinery beyond the copilot: Harness for asking questions of your workspace with sources, and agent tools for actions — priced higher per run and only worth it once the daily rhythm is understood. Start at the rung your volume justifies; the credits never expire, so the ordering mistake is the cheap kind.
What does an AI copilot actually do inside a help desk?
Four things in the reply box: drafts a first reply from the thread and workspace context (3 credits), proposes a suggested reply to new inbound messages (6), translates your answer into the customer's language with correct register (3), and summarizes long threads for handoffs (6). A human reads and sends every message.
What does the copilot cost per month for a small team?
Credits are prepaid and never expire at roughly $0.01 each — packs of $10/1,100, $20/2,400, $50/6,500. A typical 2-5 person team using copilot daily burns a few hundred to a couple thousand credits a month, so roughly $10-30 of runway.
Do we need AI to use the free plan?
No — AI ships off and the workspace is complete without it. Copilot runs only when you buy a credit pack, and any first payment also removes the "Powered by Onsites AI" watermark permanently.
What should the copilot never be allowed to do?
Auto-send anything. Every draft needs a named human sender — the audit trail exists for this. And never trust it on novel, sensitive or escalated cases where human context is the task.
Why do copilot answers sometimes make things up?
Because the model writes plausible text from whatever grounding it has. Give it a lean handbook, connected CRM context and thread history and it cites your real facts; give it nothing and it invents plausible details. That is why handbook feeding and human review are the two habits that matter.