Breakdown of Won and Lost Deals

The Grow2.ai AI agent processes every closed deal from the CRM and generates a monthly report on the reasons for wins and losses. Knowledge of why a deal was won or lost stops living only in the heads of salespeople and makes it into a document that the entire team reads — from the commercial director to product managers and marketing. Instead of manually gathering data from various sources before every retro, the team gets a structured narrative with pre-built segments and quotes.

Here is what the agent does within a single cycle:

1. Retrieves a list of deals closed during the reporting period (won and lost), filtered by pipeline and team.
2. Pulls for each deal the correspondence history, call transcripts, sales rep notes, qualification fields, stage duration, and deal owner.
3. Joins the data with behavioral metrics from the data warehouse: site visits, product events, trial activity.
4. Classifies loss reasons and win factors according to the team-agreed taxonomy — price, timing, competitor, product fit, quality of the sales process, internal political factors on the client side.
5. Extracts recurring objections, risk signals, and client phrasing as quotes with the deal and stage where they were raised.
6. Segments results by client type, lead source, product, sales rep, and cycle length.
7. Consolidates everything into a structured markdown report: a one-page TL;DR, key insights, a breakdown of individual deals, and appendices with raw data.

What automation does not do

• Does not conduct win/loss interviews with the client. A qualitative conversation with a lost client still requires a real person and interviewer skills.
• Does not replace a sales analyst or make decisions. Strategic conclusions about ICP, pricing, or product priorities remain with the team.
• Does not predict the future. The report describes what happened and why, but does not forecast the outcome of open deals or provide scoring for the current pipeline.

The document is read by the commercial director, head of sales, product managers, and marketing. The sales team gets feedback without manually digging through the CRM before every retro. Product sees signals from real conversations with lost clients. Marketing understands which segments are easier to win and where messaging or targeting falls short.

The Grow2.ai AI agent runs on a schedule (once a month or triggered by deal closure) and goes through four stages: data collection, normalization, analysis via LLM, report generation. Implementation — custom code in Python or TypeScript with orchestration via a workflow engine or a cron script, without visual no-code builders due to the complexity of prompts and data volume.

Technical flow

1. Data collection. The script calls the CRM (HubSpot, Salesforce, Pipedrive) via REST API and retrieves all deals with status closed_won or closed_lost for the period. For each deal, related entities are pulled: contacts, notes, activities, email threads.
2. Enrichment from data warehouse. Behavioral data is joined to the deal: website pages, product events, trial status. Source — Snowflake, BigQuery, Postgres replica, or BI data mart.
3. Call transcripts. If the team has Gong, Fireflies, tldv, or Otter connected — the agent pulls transcripts and notes for deals via their API.
4. Normalization and chunking. Raw data is brought to a single JSON object per deal and split into semantic chunks for the LLM.
5. Analysis via AI model. Each deal goes through several prompts: classification by taxonomy, extraction of quotes and objections, scoring of win or loss factors. Results are stored in a structured JSON.
6. Aggregation. Results across all deals are consolidated into segments, patterns are counted, and recurring themes and signals are identified.
7. Report generation. The final prompt turns structured data into a 5-10 page narrative with TL;DR, insights, a breakdown of key deals, and appendices.
8. Delivery. The completed report is published in Notion, Confluence, or sent to the sales team's Slack channel.

Implementation steps

1. Agree on the taxonomy of loss reasons and win factors with the head of sales (1-2 meetings).
2. Obtain read-only access to the CRM, data warehouse, and call recording system.
3. Collect a dataset of 30-50 closed deals from the previous period for prompt calibration.
4. Write the data collection and normalization pipeline.
5. Configure prompts for the language model and validate classification quality on a labeled sample.
6. Run the first report manually, discuss with the team, adjust the taxonomy and prompts.
7. Put the pipeline on a schedule and configure failure alerts.

Components

Report quality depends on two things: data cleanliness in the CRM (field completeness, loss reasons, call notes) and taxonomy consistency. If 40% of deals have the loss reason field empty, the LLM will not be able to fully compensate — it will recover some from correspondence and calls, but accuracy will drop. Before launch, a data completeness audit is conducted and critical gaps are addressed.

Implementation relies on access to CRM data, a data warehouse, and basic team readiness for win/loss analysis. The cleaner and more complete the input data, the stronger the output report.

Access and data

• CRM with deal history. Minimum 6 months with the basic fields filled in (stage, close date, owner, loss reason). HubSpot, Salesforce, Pipedrive connect via standard APIs.
• Data warehouse or BI data mart. Snowflake, BigQuery, a Postgres replica, or equivalent with product and behavioral data. Optional, but improves analysis quality.
• Call transcripts. Gong, Fireflies, tldv, Otter, or an internal system. Without transcripts, the report is built on correspondence and notes but loses nuance.
• Agreed-upon taxonomy. A list of loss reasons and win factors from the head of sales — 8-15 categories for each side.
• Service account with read-only permissions in the CRM, data warehouse, and call recording system.
• Report publishing platform — Notion, Confluence, or the sales team's Slack channel.

Team readiness

• Head of sales or RevOps as the project owner, ready to participate in taxonomy calibration and validation of the first reports.
• One technical team member or contractor for integrations and pipeline support.
• A sales team ready to receive and discuss findings — without a win/loss review culture, the report becomes a file no one opens.

Timeline

2-4 weeks from kickoff to the first automated report. Week one — taxonomy alignment and data audit; week two — pipeline development and prompt calibration; weeks three and four — pilot launch, team discussion, and scheduling.