Whitepaper
Always-Learning Outreach Models
Static rules age like milk. Systems that retrain on dispositions improve with every campaign.
- Static rules decay
- Two approaches
- What retraining needs
- Black-box risk
- Hybrid constraint
- Mini worksheet
- Objection table
- 30 / 60 / 90 pilot
- Evaluation questions
1 · The problem
Static rules decay
Yesterday’s best window is today’s spam window. War-room rule decks become folklore within a quarter.
Contact centers still run “best time” grids, fixed recycle counts, and scorecard cutoffs written in a launch week. Lists change. ANI reputation shifts. Consumer pick-up behavior moves with seasons and news cycles. Static rules do not—so connect rate quietly erodes while dials stay “busy.”
Who this is for: ops and product leaders who own connect rate and CPA, not vanity dials. Not for you if: you refuse to log clean dispositions or will not allow any model override process.
2 · Approaches
Two legitimate approaches
- Custom from historical call logs — data eng → features → model → validation → production → retrain on new outcomes. Best when you have volume and consistent disposition taxonomy.
- Generic + continuous retrain — start without deep history; improve on live campaign results. Best for greenfield or fragmented data—with strict guardrails against poisoned labels.
Product language worth repeating: “The machine is always learning and improving.” Learning must stay under compliance gates—models never authorize illegal dials.
3 · Data
What retraining needs
Clean dispositions, consistent outcome definitions, enough volume, and feedback that does not train on bad labels. If “sale” means three different things across teams, the model learns noise.
| Input | Why it matters | Failure mode |
|---|---|---|
| Disposition taxonomy | Labels become training signal | Same outcome, five codes |
| Contact / no-contact | Connect model backbone | AMD mislabels as human |
| Downstream sale / set | CPA-aligned ranking | Only optimizes pick-up |
| Consent / suppress flags | Learning stays legal | Model ranks blocked leads |
4 · Transparency
Black-box risk
Show rank drivers, allow overrides, report pilot lift in CPA terms—not only accuracy decks.
Managers will not trust a hopper they cannot explain to agents. Require: top drivers per score band, human override with reason codes, and holdout pilots that report connect, penetration, and CPA—not only AUC charts.
5 · Hybrid
Hybrid constraint
Models serve AI and human campaigns under the same compliance gates.
Opener / fronter / closer AI and human agents should consume the same prioritization layer. Learning improves who to call next and when—not whether consent exists.
6 · Worksheet
Mini worksheet (example numbers)
| Line | Example | Your number |
|---|---|---|
| Monthly dials | 1,200,000 | — |
| Connect rate today | 14% | — |
| Target connect after prioritization | 18% (+4 pts) | — |
| Extra connects / month | 48,000 | — |
| Close rate on connect | 4% | — |
| Extra sales / month | 1,920 | — |
| Gross margin / sale | $180 | — |
| Monthly value of lift | ~$345K | — |
Footnote: Example only—pilot / design illustration. Results vary by list, vertical, and close definition.
7 · Objections
Objection table
| Objection | Response |
|---|---|
| “We don’t have data science.” | Generic + continuous retrain + productized LeadScore-class ranking reduces the need for a PhD team on day one. |
| “Black box will scare compliance.” | Rank under consent/suppress gates; export audit of who was eligible, not just who scored high. |
| “Agents hate mystery hoppers.” | Surface drivers + override paths; coach on why a lead ranks high. |
| “What if labels are dirty?” | Fix taxonomy first; holdout pilots; don’t retrain on untrusted codes. |
8 · Pilot
30 / 60 / 90
- 30 days: Clean disposition map; baseline connect/CPA; enable scoring on one campaign with holdout.
- 60 days: Retrain cycle live; compare holdout CPA; agent override review.
- 90 days: Expand campaigns; document lift; lock retrain cadence and ownership.
9 · Evaluation
Questions for vendors
- How often do models retrain—and on what outcomes?
- Is there a holdout pilot design with CPA reporting?
- Can humans override with reason codes?
- Does learning respect consent and suppress lists at dial time?
Related: Lead prioritization · LeadScore+.