Vertical AI's evidence trade: cited playbooks in law, gap-showing diligence in pharma, and the insight gap in CX
Three vertical AI players made the same bet this week: the product is not the answer, it is the evidence trail behind the answer.
- Harvey's Playbook Builder converts existing contract playbooks into structured rules applied in Microsoft Word, categorizing clauses as acceptable, needs review, or unacceptable, with citations to source documents for each rule. Self-reported m:…8659
- Formation Bio says its Phyla tool uses 90 underlying tools across 27 sources, responds 24% faster than frontier models, and cited its statements with 99.7% consistency in research report generation tasks. Self-reported m:…0490
- Phyla is delivered through Formation Bio's internal AI system, ARK, and systematically queries biomedical datasets while making gaps in evidence explicit. Self-reported m:…0490
- 41.2% of CX operators take no action when they find a problem, and 55.1% of those who do act never see the insight implemented, according to a Parloa and Metric Sherpa study. Self-reported m:…2626
- 61.4% of CX operators want to evaluate every conversation rather than a sample, while less than half of executives ever see a CX insight. Self-reported m:…2626
Law: the playbook as product
Harvey's latest pitch for legal AI is not a smarter redline — it is a discipline. A contract playbook, in Harvey's telling, defines preferred positions, fallbacks, and escalation paths for a specific contract type, and the company's Playbook Builder converts existing guidelines into structured rules that are applied to drafts inside Microsoft Word. The categorizations are blunt by design: acceptable, needs review, unacceptable. The detail that matters most is the provenance one: the builder maintains citations to source documents for each rule, so every automated call can be traced back to the firm's own guidelines. m:…8659
The interesting move is ownership. Harvey argues that in-house legal teams typically own and apply contract playbooks to counterparty paper, and the tool is built so teams can modify fallbacks, liability caps, and escalation paths as deals change — the playbook is a living document updated from negotiation results. That is the anti-black-box pitch: the AI does not substitute the firm's judgment, it operationalizes it. Whether lawyers will actually maintain the living document is the unasked question; Harvey's blog does not say. m:…8659
Pharma: diligence with citations
Formation Bio launched Phyla, an AI tool for drug asset diligence built to survive the failure modes that kill AI in pharma: hallucination and missing negative signals. The company says Phyla uses 90 underlying tools across 27 sources, responds 24% faster than frontier models, and — the number it clearly wants you to remember — cited its statements with 99.7% consistency in research report generation tasks. It is delivered through the company's internal AI system, ARK, and is used by over half of Formation Bio, which acquires and develops clinical or IND-stage drug assets. m:…0490
The design choice worth noting is what Phyla surfaces: it systematically queries biomedical datasets, standardizes outputs, and makes gaps in evidence explicit. In diligence, the gap is the finding — an AI that tells you what it could not verify is more useful than one that confidently papers over it. Every one of these numbers is Formation Bio's own, so file them under marketing until independently tested. But the architectural bet is right for the vertical: in drug diligence, the evidence trail is the product. m:…0490
CX: the insight that dies
Parloa's contribution to the week is a study with Metric Sherpa that names the failure vertical AI keeps tripping over: findings that never become actions. 41.2% of CX operators take no action when they find a problem, and of those who do, 55.1% never see the insight implemented by the business. Less than half of executives ever see a CX insight at all. The diagnosis is a language barrier: 46.4% of executives ranked financial and business impact linkage among their top three priorities for CX measurement, while operators are still drowning in conversation samples. m:…2626
The fix operators want is scale, not summaries: 61.4% of CX operators want to evaluate every conversation rather than a sample, and Parloa argues AI-driven observability can do that by analyzing full conversation datasets into business-relevant insights. The study's numbers are Parloa's own survey product, so treat them as a vendor's framing of its market. But the pattern is real enough to be the week's throughline: evidence without an action path is just expensive reading. m:…2626
Why it matters
Three verticals, one bet: citations in law, gap-showing diligence in pharma, full-conversation observability in CX. The common thread is that vertical AI's differentiation is no longer the model — it is the evidence discipline wrapped around the model. Harvey cites the source document for every rule; Phyla makes the evidence gap explicit; Parloa wants every conversation evaluated so insights can survive the trip to the executive suite. For founders building in verticals, the lesson is plain: the buyer is not purchasing an answer, they are purchasing a defensible trail of how the answer was reached. m:…8659 m:…0490 m:…2626
What is still unresolved
- Whether legal teams will actually maintain playbooks as living documents — Harvey describes the capability, not the behavior.
- Whether Phyla's 99.7% citation consistency holds on external assets rather than internal research tasks.
- Whether evaluating every conversation actually closes the CX implementation gap — the study documents the gap, not the cure.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.