Skip to content

Learn

Outreach that gets sharper with every reply.

Your agents learn from real replies, approvals and your own edits. Then they test what they learned against the baseline, so you promote changes on evidence, not hunches.

Outreach generation · Rapport phase

Illustrative data
  • Baseline

    Default rulesLive
    Reply rate
    18%
    Meeting rate
    4%

    n = 212 messages

  • Variant B

    Built from 6 learnings Leading
    Reply rate
    24%
    Meeting rate
    6%

    n = 198 messages

+6.0 pts reply rate vs baseline

Read

Every reply, read and scored

The reply analysis agent reads each response for what the prospect actually means. Warm leads rise to the top. Cold ones stop getting pushed.

  • Intent and sentiment on every inbound LinkedIn and email reply

  • Warmth updated from response speed, tone and depth of the conversation

  • An alert when someone is ready to talk, so you can step in

Daniel Osei

Head of Growth · Lumen Freight

LinkedIn

Inbound reply

Funny timing, we are rebuilding our outbound process next quarter. What does onboarding look like for a team of five?

Reply analysis

Intent
Interested
Sentiment
Positive
Warmth
0.41rises to0.78

Ready to talk. You have been alerted to step in.

The flywheel

A learning loop built into every agent

Each run is matched to what happened next. What worked feeds the next draft. What stopped working fades out.

The Wino AI learning loop1. Capture: Every agent run is logged with the phase, warmth, priority and what the agent did.. 2. Match: Replies, approvals and edits are tied back to the run that produced them.. 3. Learn: Clear wins, clear failures and your edits become learnings, each with a confidence level.. 4. Apply: The most confident learnings are added to the next run, inside a strict size budget.. 5. Experiment: Improved prompt variants are tracked against the baseline on reply and meeting rate.. 6. Promote: You promote a proven variant, and it becomes the live version for your workspace..Learnfrom outcomes01Capture02Match03Learn04Apply05Experiment06Promote
  1. 01

    Capture

    Every agent run is logged with the phase, warmth, priority and what the agent did.

  2. 02

    Match

    Replies, approvals and edits are tied back to the run that produced them.

  3. 03

    Learn

    Clear wins, clear failures and your edits become learnings, each with a confidence level.

  4. 04

    Apply

    The most confident learnings are added to the next run, inside a strict size budget.

  5. 05

    Experiment

    Improved prompt variants are tracked against the baseline on reply and meeting rate.

  6. 06

    Promote

    You promote a proven variant, and it becomes the live version for your workspace.

  1. Capture

    Every agent run is logged with the phase, warmth, priority and what the agent did.

  2. Match

    Replies, approvals and edits are tied back to the run that produced them.

  3. Learn

    Clear wins, clear failures and your edits become learnings, each with a confidence level.

  4. Apply

    The most confident learnings are added to the next run, inside a strict size budget.

  5. Experiment

    Improved prompt variants are tracked against the baseline on reply and meeting rate.

  6. Promote

    You promote a proven variant, and it becomes the live version for your workspace.

Your edits count

When you edit a draft, the agents notice

Change a line before you approve it and that edit is recorded as a learning signal. Over time, drafts drift toward the way you write.

Every draft shows which learnings it drew on, with a confidence level and how often each has been used. Nothing is a black box.

  • Your edits become learnings for your workspace
  • Highest-confidence learnings for each agent and phase go first
  • The writer also uses what the reply reader has learned

Illustrative example

Experiment

Test changes. Keep the winners.

Wino AI builds improved prompt variants from what it has learned, then tracks each one against the baseline with sample sizes you can see.

  • Variants from learnings

    New prompt versions are built from the patterns that worked for you.

  • Measured lift

    Reply rate and meeting rate for each variant, next to the baseline and the sample size.

  • You promote the winner

    When a variant proves itself, promote it to live for that agent in your workspace.

  • Old lessons fade

    Learnings lose confidence over time and are removed once they stop being reliable.

Optimize

Every campaign gets a health check

The campaign optimization agent grades each campaign A to F and gives you concrete next steps, each tied to your own numbers.

Campaign health · Q1 agency owners

Illustrative

Phase funnel

  • Contacted
  • Connected
  • Replied
  • Meetings

Recommendations

  • TimingReplies cluster Tue to Thu mornings. Shift sends into that window.
  • PhaseLeads stall in Discovery. Add a lighter question before the pain step.
  • ICPDirector titles reply more often than VPs here. Widen the title filter.

Find the bottleneck

Response, connection and meeting rates plus the phase funnel show exactly where leads stall.

Recommendations you can act on

Three to eight suggestions across ICP, messaging, timing, phase, channel and personalisation.

Insights in one view

KPIs, outreach trends, the pipeline funnel, a campaign leaderboard and which variant is live.

Guardrails

Learning you can switch off

The loop works for you, inside limits you can see.

  • Off switch

    Turn learning off at any time. Agents carry on with their default rules.

  • Cleaned before use

    Learned content is cleaned before it is added to an agent's instructions.

  • Safe fallback

    If anything looks wrong, agents fall back to their defaults.

Further reading

Guides on improving every campaign

All articles

Start the loop on your own outreach

Connect LinkedIn and email, approve your first drafts, and let real replies shape the drafts that follow.