Talking points — draft

Working draft for Thu 2026-04-30 demo. Two parts: "Walk Us Through Your Process" answer + questions for the panel.

Walk Us Through Your Process

1. OpenAI flows failed silently. Pivoted to Azure to find anything that worked.

Built AI Enrichment flows on OpenAI per the brief. Active toggle green, scheduled, triggered manually — History Logs stayed empty. No error message. Spent ~45 minutes cycling through models, context fields, scheduled vs. trigger-based, before trying Azure. gpt-5.3-chat worked the first try.

That experience IS Signal 2 ("not sure if this is a data quality issue or if we set something up wrong") — but lived as the candidate. With Claude as a sounding board, no production data on the line, and time to brute-force my way through configurations. The customer with $185K expansion pending doesn't have that runway.

2. Support docs were stale. UI was sometimes adversarial. Nothing was immediately intuitive.

The 6 linked support PDFs are 2–5 years old — including the History Logs article, the one explicitly recommended for verifying AI enrichment output. References to features that aren't in the dropdown (Perplexity Sonar). The model lineup in the docs doesn't match the UI.

So I learned by clicking. Some of it was figure-out-able with patience (the AI Enrichment step config panel, Trigger Settings). Some of it was hostile — the Scheduled Flow row showing "Deleted" status while the Active toggle stayed green, or field dropdowns offering 100+ options with 7 different "Account Name" variants and no contextual hints.

Same opacity Signal 6 describes ("we're spending a lot of time explaining classifications we can't fully explain ourselves") — visible upstream, in the docs themselves.

3. UI lag stacked on cognitive overload.

Custom fields didn't appear in list view pickers until I hard-reloaded. The AI Enrichment configuration panel had loading delays. Field dropdowns rendered slowly with the long option lists.

Each is minor in isolation. Stacked with the cognitive overload of 100+ option dropdowns, every small lag is one more chance to fat-finger an option I didn't mean to pick and not realize for 20 minutes.

How I used AI on this assignment

I used Claude throughout as a thinking partner. The dynamic that worked: I brought directionality, Claude brought comprehensiveness and pushback.

Three concrete examples from this prep:

• The "TC + D&B as a trust bridge" framing came from me looking at the underlying data model and recognizing that LLM speed and D&B verification are complementary, not competing. Claude pushed me to structure that into a property-by-property comparison (speed vs. verification, deterministic vs. probabilistic, real-time vs. analyst-curated) and pressure-tested whether the metaphor actually held up at each axis.
• The feedback loop as a customer asset (rather than just "training data") was my framing — corrections from sales/RevOps are equivalent in kind to D&B's analyst-curated work, except scoped to the customer's own knowledge. Claude surfaced risks I hadn't named: cold start, capture bias, GDPR rights of departed employees, the messaging tension with the stickiness story. None changed the recommendation; all of them mean I'm not blindsided in this room.
• "Subtract, not add" started as my frustration after a day clicking around the sandbox. Claude helped me name that frustration as a design principle that ties together the cognitive overload, the silent failures, and the data the system wasn't capturing — three observations into one thesis.

What I deliberately didn't outsource: which observations matter, what the recommendation should be, what to scope in vs. defer. Claude is good at more — more options, more risk angles, more comparisons. I'm still the one deciding what's load-bearing.

The failure mode I tried to avoid: letting AI generate the answer and then editing it. That produces something polished but not mine, and in a panel environment, you can hear it.

Questions for the panel

Q1: Why is AI Enrichment under "Assignment Flows" instead of "Clean" or "Match"?

I expected to find AI Enrichment under "Complete Clean" (data quality) or near "Internal Match" (entity classification). Instead it lives inside "Assignment Flows" — a category I associate with routing and territory work.

A RevOps user thinking "I need to clean up the industry classifications on my accounts" might not look in Assignment Flows for this. Was this a deliberate IA choice, or has the product evolved past its original menu structure?

Q2: I tried adding an Anthropic key. It didn't work.

The Integrations panel surfaces Anthropic alongside OpenAI, Perplexity, and Google. I added an API key, then went to configure an AI Enrichment flow. The provider dropdown only accepts OpenAI and Azure auth. So the Integrations panel advertises Anthropic, but the AI Enrichment auth handling doesn't actually accept it.

That same shape — UI advertising more than the product actually delivers — showed up a few other places this week:

• The 6 linked support PDFs reference Perplexity Sonar as a model option (not in the dropdown)
• The brief describes 8 industry categories that don't appear in either object's standard picklist or in the model's free-text output
• OpenAI flows showed Active = green in the UI while History Logs stayed empty (no error, no log entry)
• "Re-run Traction → complete" dialog tells the user nothing about what actually changed

My actual question is the pattern, not Anthropic specifically: does the gap between what the UI advertises and what the product actually delivers map to what Signal 2's customer is reacting to ("not sure if this is a data quality issue or if we set things up wrong") — and is that a connection the team is actively working through?

If Anthropic specifically is on the roadmap, Claude has a strong enterprise/compliance posture (relevant to Signals 1, 3, 5, 8), Bedrock could unlock AWS-shop customers, and it tends to perform well on structured classification — what has the trade-off looked like internally?

Notes for the draft (Andrea)

• Each story above is ~30s spoken. Trim as needed for time.
• Q1 is a soft / IA-flavored question. Reads more design-thinking, will land well with Scott.
• Q2's primary work now is surfacing the advertise-vs-deliver coherence pattern — anchored in lived experience (tried adding the Anthropic key, didn't work) plus 4 concrete examples. The Anthropic-specific reasons are a fallback if they want to engage there. Risk: could land as a candidate calling out product gaps in front of the CPO. Mitigation: the question is grounded in their customer's pain (Signal 2), not in candidate critique. If it still feels too pointed in dry run, demote the meta-question to a follow-up and lead with the simpler version: "I tried adding an Anthropic key, it didn't work — has Anthropic been on the roadmap?"
• The "How I used AI" section is ~45-60s spoken — slightly longer than the others. Brief explicitly calls for authentic AI use, so worth the extra time. If trimming, drop the third example ("subtract, not add") — the first two carry the dynamic on their own.

Q&A handling for AI usage questions:

If asked organically about AI use (likely from any of the three panelists, especially given the brief's emphasis), the directionality-vs-comprehensiveness framing transfers as a one-liner: "I bring the directionality, Claude brings comprehensiveness and pushback. The decisions are still mine."

If pressed on specifics ("show me an example"), have one ready that's small enough to land in 30s:
- The risk-surfacing on the feedback loop is the cleanest — I had the "this is a customer asset" insight, Claude listed five risks I hadn't considered, I kept four (cold start, bias, GDPR, messaging) and discarded one as out-of-scope. That's the dynamic in miniature.

If asked something pointed like "how do we know this isn't just Claude's output?" — answer with what was deliberately NOT outsourced: which observations matter, what the recommendation is, what to defer. Plus the sandbox time was real and unprompted — Claude wasn't with me when I was clicking through field dropdowns at 11pm Tuesday.