Most AI calling products ship with a single “sales agent” system prompt. Customers customize it by appending their script. After three or four edits the prompt becomes prompt-soup: contradictory instructions, late-position rules overriding early ones, and a generic compliance trailer at the bottom that nullifies half of what you wrote.
We don’t do that. DigitalCallers ships four hardened strategy templates: Hard Closer, Warm Storyteller, Lean Educator, and Consultative. Each is a structured object — not a free-form text blob — with eight tunable knobs. This post is the architecture behind them.
Indian sales teams run different playbooks for different lead types. A cold real-estate enquiry gets a different conversation than a warm patient follow-up. A loan pre-screen needs a Consultant who diagnoses before pitching. A brochure-request from a website lead needs an Educator who answers questions and exits politely.
One prompt cannot serve all four playbooks. Worse, when you try, the prompt’s instructions start fighting each other: “close hard” vs. “ask discovery questions” vs. “don’t use urgency tactics”. The model picks one at random. Calls become inconsistent.
We codify the controls that actually matter for an Indian sales conversation:
Every template is fully described by its values for these eight knobs plus the project knowledge base. The template doesn’t contain “sample dialog” or “tone description” — those are derived at runtime from the knobs.
The single most expensive AI calling failure mode we’ve seen is the agent ending a call on the wrong cue. “Aur kuch nahi” (no more questions) gets misclassified as “not interested”. The agent says goodbye. A high-intent lead — who simply ran out of questions — drops.
We codify this explicitly. Every template has two lists:
GREEN LIGHTS mean stay on the call:
RED LIGHTS mean end the call:
These lists get reinforced at three layers: the system-prompt prelude, the end_call tool description, and a final compliance trailer. Triple-redundant. We learned this the hard way after a customer’s top lead got dropped on “aur kuch nahi” in early April.
Every prompt edit is versioned. Every dispatched call records which prompt version it ran on. We can roll back the second a tweak underperforms.
This sounds boring but it solves a real problem: if a customer says “the agent was great last week, what changed”, we have a deterministic answer. Last week was v2026-05-02. Today is v2026-05-03. The diff is the empathy guidance + a banned-phrase update. We can A/B them and know which one regressed.
Concretely we ship three things on top of versioning:
The BUILD_VERSION banner came out of a real bug. We spent days iterating a prompt and seeing no change in agent behavior. The fix wasn’t the prompt — the worker was loading a stale cached profile because of a sys.path import quirk. The banner now makes that bug class impossible: the new prompt either shows up in the banner or it doesn’t.
Customers can fork any template and adjust each knob. They cannot rewrite the GREEN/RED light list directly — those are fixed by the template type. They cannot delete the banned-phrase enforcement. They can only tune within the template’s safe envelope.
This trade-off matters. We had an early customer who insisted on a custom prompt with no end-call discipline. Their conversion dropped 22 % in the first week because the agent was ending calls on neutral signals. We added the GREEN/RED safety net and got the conversion back. Constraints are features.
Today, 100 % of DigitalCallers calls run through one of the four templates (or a customer fork of one). Every call has structured analytics. Every prompt is versioned. End-call false-positives are under 1 % across our deployed customers. Calls that drop on RED lights are nearly always genuine — which is what you want.
The boring stuff matters. The strategy template architecture is what makes the AI calling product feel like a product instead of a wrapper around a model.
Want to see the templates on a real call? Book a demo and we’ll dial your phone with all four agents back-to-back so you can hear the difference.
Book the demo →