Ground the answers.
Configure the approved offer, current indicative prices, service area, FAQs and escalation boundaries. Supply relevant lead history with its provenance. The agent should identify itself and answer from approved information.
DATA SCIENCE & ENGINEERING
Understand the data.
Design the conversation.
Measure the return.
Database reactivation starts with a commercial hypothesis. We structure the data, configure AI SMS and scoped voice agents, connect the booking workflow and evaluate what the campaign adds.
Our campaign design framework. Implementation, channels and measurement are agreed in your pilot scope.
01 / DATA SCIENCE FIRST
Cleaning is the start.
Context shapes the campaign.
For analytical owners and technical teams, the useful question is how each transformation and decision connects to an economic outcome.
Define the record before scoring it.
Preserve the original export, standardise contact details and resolve duplicate people or households. Separate opt-outs, unresolved permissions, active opportunities and records outside the offer or service area.
Then inspect source, age, sales stage, previous quotes and reasons for going quiet. Keep observed facts, customer statements and hypotheses distinguishable. Missing data becomes a question.
Prioritise expected additional contribution. A high likelihood of buying does not establish that our contact changes the decision. Early targeting is a hypothesis; uplift needs a suitable comparison.
CAMPAIGN DATA CONTRACT
- Identity
- Client ID, lead ID, household/property match
- Contact eligibility
- Permission evidence, suppression, channel, service area
- Sales context
- Source, last contact, quote, stage, recorded objection
- Current suitability
- Known facts, their timestamps and unresolved questions
- Campaign assignment
- Cohort, experiment group, offer and agent version
- Outcome events
- Delivery, qualification, booking, attendance, collected payment
Collect the minimum data needed. Agree access, retention and the client's source of truth before importing records.
ILLUSTRATIVE SOLAR / BATTERY DATABASE
Fictional records & countsRead the situation behind the row.
Raw records10,000
After identity & eligibility checks5,000
Example screening removes 1,000 duplicates, 1,500 permission exclusions, 1,000 unusable/out-of-area records and 1,500 existing battery or active-deal records, sequentially without double counting.
EXAMPLE RECORD / CUSTOMER
- Previous outcome
- Solar installed in 2022
- Battery record
- Unknown · last checked in 2024
- Previous battery quote
- None recorded
- Contact permission
- Documented for this campaign
An existing relationship. A new suitability question.
OBSERVATION
The CRM records a completed solar purchase. A blank battery field does not establish current battery ownership.
TESTABLE HYPOTHESIS
Some customers may now have a relevant battery need and recognise the installer. Test whether contact creates additional profitable purchases.
NEXT QUALIFICATION STEP
Confirm the current property and battery situation, then assess energy use, system compatibility and the customer's objective.
The groups are mutually exclusive in this example. Existing battery owners may need a separately scoped expansion offer. Contact eligibility and technical battery eligibility are different checks.
02 / THE ECONOMIC MODEL
A formula you can inspect.
An outcome you can verify.
Model the funnel before launch. Where the audience supports a useful experiment, measure additional paid outcomes against business as usual.
SMS PLANNING MODEL
Expected collected contribution = eligible contacts × delivery rate × positive-interest rate × qualified-booking rate × attendance rate × paid-sale rate × contribution per sale.
Each rate is conditional on the preceding stage. Rates are assumptions until measured. Subtract full incremental campaign costs to estimate the remainder.Attribution: match a campaign conversation to a booking and payment using agreed identifiers and dates.
Incrementality: estimate how many additional outcomes the campaign caused. Randomise comparable households and observe both groups through the agreed sales cycle.
INCREMENTAL ROI LAB
DBR group
Include failed deliveries and non-replies.
Count once per household in this example.
Business-as-usual group
Randomised from the same eligible audience.
Observe over the same sales window.
Client economics
A$
After equipment, installation and direct job costs.A$
Fees, messaging and extra operating effort; avoid double counting.Editable illustrative figures. These are not conversion benchmarks, Vaultar prices or a forecast. AUD; exclude GST where applicable.
DBR purchase rate1.8%
Baseline purchase rate0.6%
Estimated additional contribution after campaign cost$6,000
A point estimate before other overheads and tax. The control comparison estimates purchases that would have happened anyway.
- Purchase-rate uplift
- 1.2 percentage points
- Estimated additional paid purchases
- 6
- Additional contribution before campaign cost
- $12,000
- Return on incremental campaign cost
- 100%
- Break-even purchase-rate uplift
- 0.6 percentage points
Small sales counts can leave substantial uncertainty. A positive point estimate alone does not prove uplift or repeatability. Plan sample size and report uncertainty before drawing a causal conclusion.
Inspect the formulas and assumptions
Uplift = (DBR paid purchases ÷ DBR assigned households) − (control paid purchases ÷ control assigned households).
Additional purchases = uplift × DBR assigned households. Additional contribution = additional purchases × contribution per purchase. Remainder = additional contribution − full incremental campaign cost.
ROI = remainder ÷ incremental campaign cost. Break-even uplift = campaign cost ÷ (DBR group size × contribution per purchase). ROI is undefined at zero campaign cost; break-even uplift is undefined at zero contribution per purchase.
Use random assignment within comparable cohorts, a stable household identifier and the same observation window. Keep business-as-usual activity comparable and record other outreach. The example assumes at most one purchase per household and equal contribution per purchase in both groups. When margins vary, compare realised contribution per assigned household directly. Fractional additional purchases are estimates, not billable events. No statistical significance or cash-flow timing is calculated here.
Define the primary outcome, minimum useful uplift, sample size and observation window before launch. A small first batch can test operations while leaving sales uplift unresolved.
Explore the pilot planning calculator03 / NLP & CONVERSATION ENGINEERING
Useful conversations.
Defined actions.
Natural language processing helps interpret how people actually reply. Campaign rules decide the appropriate next step, with qualified attendance as the commercial objective.
Understand the reply.
Recognise interest, objections, timing, wrong numbers and opt-outs. Extract useful facts such as location, existing system and preferred time. Ask one relevant question when a necessary detail is missing or contradictory.
Choose the next step.
Apply qualification rules, contact limits and handoff requirements. Validate structured fields before CRM or calendar actions. Use short, relevant responses; measure progression to qualified attendance alongside opt-outs and complaints.
What agent “training” means here: configuring instructions, approved knowledge, example conversations and evaluation cases around your business. Fine-tuning model weights is a separate implementation choice requiring suitable data and a defined scope.
SIMULATED NLP READING
Scripted examples · no live AI callSame campaign. Different intent.
CUSTOMER REPLY
“Still keen, but we already have solar. No battery yet.”INTERPRETED INTENT
Interest + current setup
STRUCTURED INTERPRETATION QUALIFY
- solar_present
- Yes · customer stated
- battery_present
- No · customer stated
- assessment_ready
- Not yet verified
Ask the next missing qualification question. Existing solar alone does not establish technical suitability or a confirmed booking.
AI SMS
Asynchronous, contextual follow-up.
Design for brief messages, delayed replies and conversation continuity. Recheck suppression before follow-ups, stop on replies that change the next action, and pass qualified interest to booking or a salesperson.
- Evaluate
- Delivery, positive interest, qualification accuracy, qualified attendance and contact fatigue.
- Operating boundary
- Approved audience, offer, cadence, spend and escalation rules.
AI VOICE · SCOPED SEPARATELY
A conversation that can listen and hand off.
Voice adds speech recognition, spoken responses, turn-taking and interruption handling. Configure permitted callbacks or agreed inbound use, identity disclosure, qualification and human transfer paths before connecting the calendar.
- Evaluate
- Speech accuracy, response latency, interruptions, safe handoffs and booking outcomes.
- Operating boundary
- Channel permission, calling hours, recording requirements and available integrations.
04 / CALENDAR & SALES HANDOFF
A booking needs
a confirmed calendar record.
Connect the agreed CRM calendar or supported booking integration. Validate availability, write the appointment and attach the context your team needs.
- 01
Qualify against the written criteria
Confirm relevant need, service-area fit, decision-maker status, timing and indicative price acceptance where required. Technical suitability and final quotes remain with your team.
- 02
Check the correct live calendar
Use the selected event type, salesperson, time zone, duration, buffers and working hours. Offer only slots returned as available by the configured integration.
- 03
Validate and create the appointment
Confirm the chosen slot and required details, recheck availability where supported, and prevent repeat requests from creating duplicate bookings.
- 04
Confirm only after a successful write
Require the calendar's success response and appointment identifier. A timeout, unavailable slot or failed request becomes a retry or human handoff.
- 05
Track attendance and the eventual outcome
Send agreed reminders, sync changes, record attendance evidence, and match paid outcomes and refunds. Apply the written billable-event definition.
05 / DEVELOPMENT & EVALUATION
Build. Test. Observe.
Improve with evidence.
The release process should make the behaviour reviewable, the outcomes traceable and the next improvement specific.
Design
Agree the cohort, offer, primary outcome, qualification criteria, channels and appointment capacity.
Configure & test
Version instructions and approved knowledge. Test interest, objections, opt-outs, ambiguity, unsupported claims and booking failures.
Launch a bounded batch
Verify delivery and routing, inspect a sample of conversations, and check the calendar workflow before expanding volume.
Review & iterate
Reconcile attendance and payment, examine cohort results, and test a focused change against the current baseline.
| Layer | Measure | Decision it informs |
|---|---|---|
| Data quality | Identity conflicts, missing fields, permission evidence, exclusions | Whether the cohort is ready for contact. |
| Conversation quality | Reviewed intent errors, grounded-answer errors, human interventions | Whether agent configuration needs correction. |
| Qualification | Qualified outcomes checked against agreed criteria | Whether appointments meet the actual sales requirement. |
| Booking operations | Failed writes, duplicate appointments, wrong time zones, attendance | Whether the integration and handoff work reliably. |
| Client economics | Collected contribution, incremental costs, uplift with uncertainty | Whether continuation or expansion is justified. |
| Contact experience | Opt-outs, complaints, frequency, unsupported responses | Whether to change the campaign or pause it. |
Follow the event, not just the latest CRM stage.
Use stable client, lead and appointment identifiers to reconcile message status, replies, qualification, booking changes, attendance, invoices and refunds. Keep event timestamps and agent versions so duplicate callbacks, missing outcomes and configuration changes can be investigated.
Technical review: the questions we work through
- Identity and retries: what prevents duplicate contacts, cross-client joins and repeated booking writes?
- Missing outcomes: which system owns attendance and collected payment, and who completes incomplete records?
- Experiment validity: is assignment stable, are both groups observed consistently, and could other outreach or limited appointment capacity distort the comparison?
- Agent releases: which examples are held out for evaluation, what triggers escalation, and can a problematic version be rolled back?
- Baseline: does the AI workflow improve contribution or reduce operating effort compared with a simpler approved workflow?
- Repeatability: do results survive another cohort, and do new eligible records replenish the opportunity over time?
Record permissions and keep access scoped to the client. Agree retention, deletion, recording and audit requirements before launch. Suppression updates must reach scheduled outreach as well as the contact record.
METHOD & IMPLEMENTATION REFERENCESExperiment design Message-status events Contact consent Battery eligibility
BRING YOUR COMMERCIAL QUESTION
Let’s design a pilot
your team can inspect.
Start with rough database counts, your sales process and contribution per paid job. We’ll identify the data, conversation and measurement requirements together.
Book a database review Review the solar pilot scope No customer export needed for the first conversation.