Oliver — Test Build Requirements

Purpose

Everything needed to build Oliver for the first Dynamo Dave test and to prove a measurable result against the current system. Materials can arrive in layers — nothing has to come at once. Once the first layer lands, tuning Oliver to test-ready takes a few days, a week at the outside.

1. The baseline to beat (the priority)

This is what turns a demo into a test. Without it, we can show Oliver performs well — but not that he beats what you already have.

• Current reactivation rate on drop-offs (~25% was mentioned) — confirmed as the figure to beat.
• The dataset behind that figure: size, source, how recent.
• The drop-off journey: channels in use (email warm-up → SMS → WhatsApp), the landing page / form the leads hit, and the point at which a drop-off is counted.
• Per-claim economics.
• Definition of success for the test, and the target delta.

2. Domain knowledge to ingest

• Source material on the PCP claim: eligibility, the process, what the consumer must do, the redress timeline (including the 2027 deferral).
• The compliance boundaries for B2C financial claims (FCA financial-promotion rules) — what may and may not be said.
• The specific claim flow, FAQs, and common objections for this partner / law firm.
• Accepted formats: Markdown, Word, RTF, plain text, or text-based PDF. (No scanned / image-only PDFs.)

3. Behavioural rules for this vertical

The governance layer — what Oliver must and must not do.

• Never: promise or quote a payout figure · give legal advice · apply pressure · accuse anyone · hide that he is AI · withhold the free / alternative route.
• Always: disclose he is AI honestly · surface the free route · reassure · offer a human · book the call.
• Any off-limits topics, the preferred wording when Oliver doesn't know something, and whether he may close an unproductive conversation.

These correspond to controls already built into the platform and validated against ~300 adversarial test cases. A short confirmation that this rule-set is right for PCP-via-insurance is all that's needed.

4. A sample of the target data

• A sampled or obfuscated slice of real records, so Oliver knows who he's speaking with (name, insurance product held, whether car finance is present).
• Anonymised or synthetic data matching the real demographic distribution is fine — what's needed is the shape, so we can test "Oliver knows I'm this person, in this situation."

5. Personality & resonance

• The demographic median of the Dynamo Dave list (age band, region, profile) so the persona lands.
• Preferred tone, and the WhatsApp cadence — human-paced replies, a "typing…" feel, not instant-encyclopedia. Text-only assumed for the test.

6. Routing & outcomes

• Where a qualified / converted lead should go (e.g. straight to the law firm, or captured and handed off).
• Who receives the conversation summaries / briefs.
• Escalation conditions — when Oliver should hand off to a human.

Timeline

• First materials → Oliver test-ready: a few days to a week after the first layer arrives.
• Target for the first test tranche: 1 July.