What is Next Best Action and How Does It Work?
Managers spend up to 80% of their time on leads that will never convert. Next Best Action (NBA) is an ML system that analyzes deal context: CRM stage, lead activity, external signals (company news, LinkedIn updates), and temporal factors (end of quarter). Output is a single action with an explanation, e.g., "Call the sales director because the company just hired a new CTO — a good reason to reconnect."
We combine two models: LLM (GPT-4o or LLaMA 3) generates the top 5 possible actions, and gradient boosting ranks them by predicted conversion rate. This hybrid approach yields 22% higher recall compared to pure LLM generation, as it accounts for historical patterns.
What Problems Does NBA Solve — Developing an AI System
Information noise. A typical CRM dashboard contains dozens of fields but doesn't tell you what to look at. NBA aggregates deal history, lead activity, external signals, and outputs one action with justification.
Lack of personalization. Managers follow templates without adapting to the lead's style. NBA learns from the actions of top performers and suggests specific tone and communication channel.
Delayed reactions. When a lead starts actively browsing the site, you need to respond within an hour. NBA monitors events in real time and sends push notifications to Slack — reaction time drops from 4 hours to 15 minutes.
How Does NBA Incorporate Deal Context?
Each deal is in one state of a state machine. The state is determined by CRM stage, recency of contact, lead activity (site visits, email opens), company news (CEO change, funding round), and temporal factors (end of quarter). Based on this, NBA chooses an action from the library:
{
"actions": [
"call",
"send_case_study",
"offer_extended_trial",
"request_meeting_with_executive",
"propose_discount",
"ask_for_referral"
],
"scoring_model": "gradient_boosted_trees"
}
Each action has a historical success probability that is recalculated daily. LLM (GPT-4 or LLaMA 3) generates a contextual explanation: "Recommended to call because 12 days without contact, the company posted a job opening for your role, and end of quarter is in 3 weeks."
Why Is the Hybrid Model (LLM + Gradient Boosting) More Effective Than Pure LLM?
LLMs tend to generate plausible but not always statistically sound advice. We combine LLM with a boosting model on historical data: LLM suggests top 5 actions, and the model ranks them by predicted conversion rate. Thus, explanations remain natural while the choice is accurate. In A/B tests, we observed a 15–25% conversion lift within the first 2 weeks. The hybrid model is 1.5 times more accurate than pure LLM in precision@1.
Approach Comparison: LLM vs Hybrid Model
| Parameter | Pure LLM | Hybrid (LLM + Gradient Boosting) |
|---|---|---|
| Precision@1 | 62% | 78% |
| Recall@5 | 74% | 92% |
| Decision speed | 1.2 sec | 1.8 sec |
| Data requirement | any | >= 5000 deals |
The hybrid model requires more historical data but yields stable gains. We recommend it for companies with at least a year of CRM history.
What's Included in the Work
Detailed stages
1. **CRM data audit** — check quality, completeness, and relationships. Result: a report with recommendations. 2. **State machine design** — state diagram and action library. 3. **Model training** — baseline and final version with metrics. 4. **Integration with CRM and messengers** — a working prototype in a sandbox. 5. **A/B testing** — statistically significant conversion improvement (target: +20%). 6. **Documentation and training** — instructions, video, and webinar for the team. 7. **Warranty support** — one month of monitoring and refinements.| Stage | Result | Duration |
|---|---|---|
| CRM data audit | Report on quality, completeness, relationships | 3–5 days |
| State machine design | State diagram + action library | 3 days |
| Model training | Baseline + final version with precision@k, recall metrics | 1–2 weeks |
| Integration with CRM and messengers | Working prototype in sandbox | 1 week |
| A/B testing | Statistically significant improvement (target: +20% conversion) | 2 weeks |
| Documentation and training | Instruction, video, webinar for managers | 2 days |
| Warranty support | 1 month monitoring and refinements | — |
Estimated Timeline
Minimum project with ready data — 4 weeks. Complex integration with custom CRM and additional data sources — up to 8 weeks. Cost is calculated individually after an audit.
Checklist of Common Implementation Mistakes
- Using only LLM without a statistical model — recommendations are nice but ineffective.
- Ignoring personalization per manager — NBA must know that Ivan is good at demos and Petrov at price negotiations.
- Not setting up data drift monitoring — NBA degrades in 3–4 months without retraining.
- Forgetting the feedback loop: managers must be able to "ban" a specific recommendation, and that must influence the model.
Why Order Development from Us
We have deployed NBA systems in 15+ companies, from e-commerce to enterprise SaaS. Our engineers are certified in MLflow, Kubeflow, and Hugging Face. We guarantee the model accounts for your business logic and doesn't produce absurd recommendations. Order a demo project — we'll prepare an architectural draft tailored to your CRM in 3 days. Get a consultation for your case — we'll send an example architecture adapted to your CRM.
Formal definition of Next Best Action can be found on Wikipedia