Freshdesk AI Playbook: Freddy AI Agent, Copilot and Insights

Freddy AI in Freshdesk is three pieces: two separate add-on bills and an Enterprise extra, and the trick is switching them on in the right order. Freshdesk's pricing page lists the Freddy AI Agent with the first 500 sessions included once per account on Growth, Pro and Enterprise, then $49 per pack of 100 sessions added to the current billing cycle.^[1] Freddy AI Copilot, the per-agent assistant, costs $29 per agent per month billed annually or $35 per agent per month billed monthly, on Pro and Enterprise only. Freddy AI Insights comes only with Enterprise. Our verdict: start the Email AI Agent in shadow mode, license Copilot for your busiest agents, and treat Insights as an extra. Never upgrade for it alone.

What Freddy AI does on each Freshdesk plan

Freshworks splits Freddy into a customer-facing AI Agent, an agent-facing Copilot and a reporting layer called Insights, each behind a different gate. The table shows what a standalone Freshdesk account can turn on. The Chat AI Agent, which answers web chat, WhatsApp, Instagram and SMS, exists only in Freshdesk Omni, because standalone Freshdesk has no live chat for new accounts. Our Freshdesk Omni guide covers that route if chat is your main channel.

The plan itself is billed per agent, and every agent on the account sits on the same plan, so these gates apply to the whole team; the Freshdesk pricing breakdown has the seat math. The two AI bills run on different meters. Copilot never uses sessions. Only customer-facing AI conversations draw on the session balance, while Copilot is a flat license per chosen agent.

Step 1: Feed the Email AI Agent before you switch it on

The Email AI Agent answers new tickets from your mailboxes, support portal and web forms, never chat. An admin builds it in AI Agent Studio, and it needs at least one mailbox and one knowledge source before it can deploy. Sources can be solution articles, uploaded PDF, Word or text files, crawled public URLs and hand-written Q&A pairs. Retrieval works by meaning rather than keyword, and knowledge edits are usually indexed within one to two minutes, so fixing a weak article pays off almost at once.

Two filters decide what the agent can see. It uses only published, public solution articles, and only from categories whose visibility is All Users or Bots, so an internal folder stays invisible to it. Studio limits are generous for a small team: 50 AI agents per account, knowledge files up to 35 MB with 200 per agent, and public URLs capped at 10 per agent and 25 per account, crawled up to 3,000 pages as text only.

Before going live, run Evaluate. Freshdesk's help center recommends a test of at least 20 queries, and Freddy can also generate sample queries from your knowledge sources. Read the misses, not the hits. Every wrong or empty answer points at a missing or ambiguous article, and fixing that article fixes the agent for every customer who asks the same thing.

Step 2: Run shadow mode, then set the guardrails

Shadow mode is the safest first deployment, and it is available on Growth, Pro and Enterprise.^[2] The agent writes its answer as a private note and never emails the customer, and the ticket stays open. You can limit shadow mode to particular customer groups or ticket sources. It handles only tickets that arrive after you enable it, so plan a window of real traffic instead of expecting it to clear your backlog. Live mode is not retroactive either.

When the drafts look right, tune four settings before letting it send. The confidence threshold is High or Medium, set once per agent with no per-topic override. A daily ticket limit and a maximum number of turns, each with a minimum of 10, cap how far a bad answer can spread. Handover phrases, such as a customer asking for a person, route the ticket to your team, and a list of topics the AI must not answer acts as a hard block, whereas custom instructions are only guidance.

Know what customers will see. Every AI email carries a fixed legal footer saying an AI wrote it, and you cannot remove it. A Helpful click closes the ticket, while Not helpful hands it to a person with a private-note summary of the question and the handover reason. While the AI holds a ticket, your agents' reply box is disabled. On Pro and Enterprise, automations can fire on the handed-over, resolved and timed-out events to tag or escalate AI tickets.

Step 3: Budget the session meter, not the seats

An email session is a 72-hour window from the customer's first email, however many messages follow; on Omni, a chat session covers 24 hours.^[3] The 500 complimentary sessions arrive once per account for new customers and last until used, not every month. After that, packs of 100 join the current billing cycle at once, and unused purchased sessions do not carry over to the next cycle. At 100% usage, deployed AI agents stop taking new queries until you buy more. Alerts go out at 70, 80, 90 and 100%, and auto-recharge adds a pack when fewer than 50 sessions remain, for online payment only.

Size the packs from your volume. By our arithmetic, a $49 pack of 100 sessions, added to your current monthly, quarterly or annual billing cycle, works out to $0.49 a session. On monthly billing, a team whose AI handles 300 email conversations a month uses up the free 500 during month two, buys one pack that month, then needs three packs, $147, every month after. Because packs come in hundreds and expire with the cycle, 310 sessions a month means four packs. That gap adds up.

The rival models show what a session buys. Help Scout charges $0.75 per AI resolution and Intercom charges $0.99 per Fin outcome, so both bill per resolved or completed conversation, while a session is spent each time the AI agent takes up a conversation, whether or not it resolves it. Comparing paid sessions with paid resolutions, after each product's included allowance, and by our arithmetic, a $0.49 session costs less per solved ticket than a $0.75 resolution when more than about 65% of sessions end resolved, and less than a $0.99 outcome above about half. Zendesk includes five automated resolutions per agent a month on Support Team and Suite Team, then charges $1.50 committed or $2.00 pay-as-you-go.

Step 4: License Copilot for the agents who carry volume

Copilot is bought for chosen agents on Pro and Enterprise; an agent consumes a license by holding the Freddy AI Copilot User role. It bundles a writing assistant, ticket summaries, a reply suggester, an article generator and suggester, similar-ticket and canned-response suggesters, sentiment analysis, auto triage and Live Translate. Some pieces need data first: the article suggester wants 25 or more published articles, similar tickets about 5,000 tickets, and auto triage 2,000 or more historical tickets. Expect two to three days before some features activate.

Ration licenses. Freshdesk's own guide suggests giving roughly 80% of licenses to high-volume agents and 20% to agents with the slowest average first response. For a team of ten, by our arithmetic, four licenses on annual billing cost $1,392 a year against $3,480 for all ten. Live Translate is pooled at 100 translated tickets per license a month, so four licenses cover 400 translations, in about 40 languages.

Watch the license sweep. Freshdesk can automatically assign any unassigned Copilot licenses you have bought to all eligible full-time agents, after an in-product notice with a date. It is all or nothing. It never buys extra licenses, and if spare licenses cannot cover every eligible agent, none are assigned; to opt out, contact support before the date. Buy only what you will assign.

Insights, MCP and what Freddy does with your data

Insights is Enterprise only. It surfaces proactive insights weekly once it has 12 weeks of data, runs root cause analysis on negative insights, and since 1 June 2026 includes Conversational Insights, where you ask questions in plain English and get charts. It works in English only and its reports cannot be exported. The MCP server, generally available since 10 September 2026, lets tools such as Claude Code or Cursor read and update tickets with an API key, and its calls are metered as actions: 100, 500 or 1,000 a month by plan, then $15 per 1,000 actions. Any user with a valid key gets every tool within their role, so read our Freshdesk API guide before handing out keys.

Data handling needs a decision. Under Freshworks' Supplemental Terms, AI training on customer data is on unless you opt out by emailing support; processing for training stops within seven business days, and your data leaves training within 180 days. Freshworks will not use customer data to train third-party hosted models, including Azure OpenAI. Freddy keeps working after a training opt-out, though Freshworks says performance may be lower. For accounts in India, the Middle East and Australia, similar-ticket suggestions are stored in region but processed in the US.

The bottom line

Freddy AI pays off when you stage it. Clean up your published articles, run the Email AI Agent in shadow mode on real tickets, then let it send with a High confidence threshold and a daily cap. Budget sessions from your ticket history, since the 500 included sessions come once and purchased packs expire at the end of each billing cycle. Add Copilot only for the agents who handle the most volume, and decide on the AI-training opt-out before go-live rather than after.