Freddy AI is Freshworks’ AI layer, embedded across Freshsales and other Freshworks products. In Freshsales specifically, Freddy AI provides contact scoring, deal insights, activity intelligence, and — on Enterprise — conversational AI for email drafting and call summarisation. This guide covers what each Freddy AI feature actually does, how it affects sales rep productivity, what data it draws on, and what realistic expectations look like for AI-assisted selling.
The important question is not whether the AI sounds impressive, but whether it improves the practical rhythm of selling. If it helps the team prioritise better and understand deals faster, it is doing real work.
Freddy AI is Freshsales’ attempt to make the CRM more predictive and more helpful in day-to-day sales work. That includes scoring, insights, and assistance features that are meant to reduce guesswork and surface what a rep should pay attention to next.
Freddy AI Features in Freshsales
| Feature | What It Does | Plan Required |
|---|---|---|
| Contact scoring | Scores contacts 0-100 based on engagement signals; surfaces hot leads in priority order | Pro ($39/user) |
| Deal insights | Flags at-risk deals with specific reasons; shows deal health score | Pro |
| Next best action | Suggests the most effective next step for each contact or deal | Pro |
| Predictive contact scoring | Uses historical win/loss patterns to predict which current contacts are likely to convert | Pro |
| Freddy Copilot | AI email composition, meeting summarisation, suggested responses | Enterprise ($59/user) |
| Conversation intelligence | Call transcription, sentiment analysis, keyword alerts, talk-to-listen ratio | Enterprise |
| Freddy Insights | Natural language queries against CRM data (“What deals closed last month over $50K?”) | Enterprise |
Contact Scoring: How It Works
Freddy contact scoring evaluates engagement signals to assign a score from 0 to 100. Input signals include: email open and click rates, website visit frequency (requires the Freshsales web tracking script), call and meeting engagement, and recency of activity. The score updates continuously — a contact who opens three emails in a week moves from cold to warm automatically. Scores appear on contact list views, allowing reps to sort by score and prioritise the hottest contacts first.
For contact scoring to be meaningful, web tracking should be enabled. Without it, the score relies only on email and call data from within Freshsales, missing the website behaviour signals that often indicate purchase intent. Enable web tracking under Settings → Website Tracking and add the tracking script to your website.
Deal Insights: Risk Detection
Freddy Deal Insights monitors open deals and flags risk conditions. Unlike simple scoring systems that only assign a number, Freddy provides specific risk reasons: “No activity recorded in 21 days,” “Deal has been in this stage for 3x the average time,” “Contact has not responded to last 3 emails,” or “Deal value was reduced by 40%.” Each reason is actionable — the rep knows exactly what the problem is and can address it.
Deal insights learn from historical patterns in your Freshsales account. After 90–180 days of deal data, Freddy’s predictions become more calibrated to your specific sales cycle and activity patterns. In the first 30–60 days, predictions draw from Freshworks’ general model rather than your specific data.
Freddy Copilot: AI Email and Call Intelligence
Freddy Copilot (Enterprise) provides email composition assistance — reps describe what they want to communicate and Freddy drafts the email based on the contact’s history, the deal context, and the stated intent. The output requires editing but significantly cuts the blank-page friction on follow-up emails. Call summarisation transcribes recorded calls and extracts action items, follow-up commitments, and next steps — reducing time spent writing post-call notes.
Realistic Expectations for Freddy AI
Freddy AI adds the most value in specific scenarios: high lead volume where manual prioritisation doesn’t scale (scoring helps), long sales cycles where deal staleness is a genuine problem (deal insights help), and high call volume where manual note-taking is a bottleneck (Copilot’s call summaries help). For small teams with 20–50 active deals and a short sales cycle, the AI layer adds incremental value — reps can manually assess their pipeline without AI assistance.
Freddy’s predictions are probabilistic, not certain. Deal insights flag risk; they don’t predict whether a deal will close. Contact scores reflect engagement; they don’t guarantee conversion. Treat AI signals as one input into rep judgment, not a substitute for it.
Getting Consistent Value from AI Features in Practice
AI-powered CRM features deliver uneven results without sufficient data and correct configuration. Understanding what conditions produce reliable recommendations — and what signals degrade them — helps teams set accurate expectations.
How much historical data does the AI need to produce useful predictions?
Most CRM AI features need a meaningful baseline of historical activity data — typically at least 6 months of logged interactions and a minimum number of closed deals (often 50–100) to produce statistically reliable predictions. Check your vendor’s documentation for specific minimums.
Can AI features be turned off for specific users or teams?
Yes, most platforms allow AI feature visibility to be controlled at the profile or role level. This is useful during phased rollouts where you want to test AI adoption with one team before a broader deployment.
Is the AI model trained on my data alone, or shared across all customers?
This varies by vendor and is worth clarifying before you sign. Some vendors train global models on anonymised customer data for better accuracy; others build individual models per customer. Enterprise contracts often allow for dedicated model training. Verify this directly with your vendor.
How accurate are AI deal close predictions in practice?
Accuracy depends heavily on data quality and consistency. Well-configured implementations with clean, consistent data typically see 70–80% accuracy for high-confidence predictions. Poorly maintained CRM data produces unreliable predictions regardless of the underlying model quality.
What should I do when the AI recommendation seems wrong?
Most CRM AI features include a feedback mechanism — use it. Marking a recommendation as unhelpful directly improves the model over time. Accumulating this feedback also gives you data to share with your vendor if you want to raise accuracy concerns formally.
AI features are most useful when they make the CRM easier to act on. If the suggestions are disconnected from the workflow, the value drops fast.
Common Problems and Fixes
Problem: AI Recommendations Are Ignored Because Reps Do Not Trust the Model
AI-generated next-best-action suggestions only work if the sales team acts on them. Low trust — often caused by early inaccurate recommendations — leads to permanent dismissal of the feature. Fix: During initial rollout, have managers review AI recommendations alongside reps rather than expecting autonomous adoption. When the AI is right, reinforce it explicitly. Address inaccuracies with your vendor to improve the model on your data.
Problem: AI Features Require More Data Than the Team Has Logged
Machine learning models inside CRM platforms need a minimum volume of activity data to produce meaningful predictions. New implementations and small teams often fall below this threshold. Fix: Focus first on driving consistent activity logging (calls, emails, meetings) for 60–90 days before expecting accurate AI output. Most vendors publish minimum data requirements for each AI feature — check these before evaluating accuracy.
Problem: AI-Scored Leads Reflect Historical Bias Rather Than Current Ideal Customer Profile
Models trained on historical closed-won data can reinforce past patterns rather than identifying new high-value segments. Fix: Periodically review which company and contact attributes are driving your AI scores. If your ICP has shifted, work with your vendor to retrain or recalibrate the model using your most recent 12 months of closed-won deals.