Join the waitlist

Let us know how we should get in touch with you.

Thank you for your interest! We’re excited to show you what we’re building very soon.

Close
Oops! Something went wrong while submitting the form.

AI SDR Customization: The 3-Tier Framework for Day 1 to Steady State

Austin Hughes
·
Updated on: November 4, 2026

Customize an AI SDR in three layers: qualification and message boundaries, workflow decisions, then CRM writeback and ongoing supervision. Progress when each layer behaves correctly on representative records. “Day 1,” “weeks 2–4,” and “steady state” are useful planning labels, not guaranteed implementation timelines.

The purpose of customization is to make the system follow your sales motion consistently. A longer prompt alone cannot establish account ownership, enforce exclusions, or determine what happens when two tools update the same field. Those decisions need explicit settings and a responsible owner.

The three tiers at a glance

Tier Configure Evidence required before expanding
1: Qualification and messaging ICP, exclusions, research questions, voice, unsupported claims Representative accepted, rejected, and uncertain cases reviewed
2: Workflow behavior Triggers, routing, enrollment, approvals, reply handling A complete journey including a failure and a stop condition
3: CRM and supervision Matching, field authority, connection identity, monitoring Correct readback, recoverable errors, and a maintained change record

These tiers describe dependencies. A team with an established integration may already have parts of tier 3 in place. A new team may need to resolve ownership before it can safely complete tier 1. Evaluate the actual workflow rather than treating the order as a rigid calendar.

Tier 1: Translate your ICP into testable decisions

Begin with the account, the person, and the reason for outreach. “Target growing SaaS companies” is insufficient if the operator cannot explain what qualifies as SaaS, which regions are served, or whether a hiring announcement is relevant to the offered product.

Write down required fit conditions, explicit exclusions, and how to handle missing information. Use list-building filters for structured conditions that are already available. Use research agents for questions requiring investigation. Do not ask a model to infer a definitive answer when your rule actually requires an authoritative field.

Unify research agents have a company or person record type, questions with expected response types, and optional guidance. A company agent can ask whether a business sells to a particular customer segment; a person agent can investigate role-specific information. The record type must match the records entering the qualification action.

Define specialized terms in the guidance. For example, distinguish “offers payments software” from “accepts card payments.” Test a company that clearly qualifies, one that clearly fails, and one whose public information is ambiguous. Review the supporting research, not only the final answer. Route uncertain records to review rather than converting uncertainty into a positive qualification.

Next, separate factual input from writing instructions. The sequence-personalization overview describes different ways to use context. Template variables supply structured values; Smart Snippets support conditional messaging. Neither makes an incorrect source fact true.

A useful voice specification includes an example your team approves, the recipient's likely problem, the allowed offer, and claims the message must not make. “Sound human” is too vague. “Use one relevant observation, explain why it may matter, and ask a question without claiming the person visited our website” is actionable.

A practical qualification brief

For a hypothetical vendor selling finance workflow software, a qualification brief might require a supported region, an appropriate finance role, and evidence of the relevant operating model. It might exclude existing customers and companies with an active opportunity. Missing operating-model evidence would go to manual review.

The agent should return the observed fact and its source separately from the sales interpretation. “The company announced a new finance leader” is evidence. “The company has budget for our product” is an unsupported inference. The same separation should survive into the message.

Tier 2: Customize what happens after the answer

A correct research result still needs a defined action. Use Play triggers to decide when records enter a workflow and Play actions to define qualification, prospecting, routing, and other supported steps. Inspect both qualifying and non-qualifying paths.

For example, a new-hire event could trigger research, followed by a role check, an exclusion check, and an outreach decision. It should not silently bypass the account owner simply because the event is recent. Review the engagement policy that applies to the acting user and account.

Decide where a person reviews the result. This is particularly important with agentic sequences: the entry path affects the review experience. Direct enrollment can generate messages and start the enrollment without a separate approval screen. A team's operating checklist should name the actual path, not assume that all AI-generated messages wait for review.

Read sequence enrollment states when diagnosing work that did not proceed. A blocked enrollment, a paused enrollment, and an excluded person require different responses. Re-running the entire workflow without understanding the state can create duplicate work or repeat the same error.

Reply handling needs its own owner. Use the reply workflow to verify where responses appear and who handles them. Include negative replies, referrals, absence messages, and requests to reconnect later in your test. A system that generates strong first messages but leaves ambiguous replies unattended is not operationally complete.

Tier 3: Make writeback and supervision explicit

Before enabling broad CRM writes, create a field-authority worksheet. For each field, identify the source of truth, allowed write direction, trigger, behavior for an existing value, and behavior for an empty value. The Salesforce field-mapping guide and HubSpot field-mapping guide document these choices separately from sync triggers.

In Unify's mapped writes, “Fill if empty” preserves an existing CRM value. “Overwrite always” replaces it with an available Unify value but does not clear the CRM field when the Unify value is empty. “Overwrite on manual edit,” where supported, can include clearing a value. Do not compress these into a generic promise that the “latest value wins.”

Working with CRM data in chat is another write path. Direct updates requested through chat are not the same as a background mapped sync. Verify the connected identity and request a preview when your process requires review before a consequential change.

Keep an internal record of prompt changes, qualification rules, audience definitions, and approved examples. After a change, repeat the cases most likely to be affected. The Play settings and logs help inspect executions; do not assume they constitute a complete history of every administrative change.

Evaluate customization with an acceptance set

Build a versioned set of representative accounts and people from your own motion. Include sparse websites, ambiguous company names, changed job roles, existing opportunities, missing contact routes, and conflicting CRM values. Have the sales owner assign expected outcomes before reviewing the agent's outputs.

For an internal acceptance measure, use the frozen test set as the population and the review timestamp as the snapshot. The numerator is cases whose qualification, evidence, and permitted next action all match the approved expectation; the denominator is all cases tested. Count unresolved cases separately and leave them in the denominator for this conservative measure. With no tested cases, report “not assessed,” not 100%.

This is an operational acceptance score, not a market accuracy benchmark. A pass on a small set supports a bounded pilot, not unlimited autonomous execution. Inspect the type of failure: a false positive that causes prohibited outreach may matter more than several cases correctly deferred for review.

Test case Expected behavior Owner of the decision
Clear ICP match with usable evidence Proceed through the approved path Sales owner
Missing evidence for a required condition Defer or reject according to the rule Qualification owner
Account assigned to another rep Apply the relevant engagement policy RevOps
Non-empty protected CRM field Preserve it under the approved mapping CRM administrator
Reply requesting no further contact Apply the appropriate stop and suppression process Assigned seller

What to ask a vendor before committing

Ask the vendor to demonstrate editing qualification logic, testing an ambiguous record, controlling an approval point, and investigating a failed write. Request the plan and permissions required for each capability. Unify's pricing page distinguishes plan capabilities; do not assume every integration or automation is available at every tier.

Favor controls your administrator can inspect and maintain. A vendor-managed configuration can still be appropriate, but document who can change it, how quickly a problem can be stopped, and how the team confirms the fix. Customization is useful when it makes the workflow easier to operate and explain.

Frequently asked questions

How customizable should an AI SDR be?

It should let your team control qualification, evidence requirements, messaging boundaries, permitted actions, ownership, and CRM writes. The relevant question is whether you can test and maintain those decisions.

Should every customization happen on day one?

No. Establish the essential eligibility and message boundaries first, then expand workflow and writeback behavior as the relevant cases pass review. The tier labels are planning stages, not fixed deadlines.

Does a human review checkpoint apply to every enrollment path?

Not necessarily. Review behavior depends on the product and entry path. Test the exact path your team uses rather than assuming AI-generated content always waits for separate approval.

How do you know customization is working?

Compare outcomes on a versioned acceptance set, inspect unresolved and incorrect cases, and monitor the live workflow after changes. Keep qualification quality, action correctness, and commercial outcomes distinct.