An advisory firm that works with senior executives at private equity backed companies runs engagements built on interviews with a leader's colleagues, delivered as a confidential feedback report organized by theme and supported by exact quotes. Producing one report meant days of reading transcripts, sorting hundreds of quotes into themes, anonymizing them and drafting the analysis, and the firm's standing rests on every quote being accurate and every theme being complete. A general-purpose chat tool had been tried, but when asked to theme and quote in one pass it paraphrased, and roughly ninety percent of its quotes failed a word-for-word check. New Clarity built a staged analysis engine that takes raw transcripts and produces a complete draft in the firm's own report template, with every quote verified against the transcript, every claim in the summary traced to the passages it came from, every consultant edit captured and turned into rules for future runs, and an independent benchmark that scores each draft against the firm's own reports.
The reports are read by the executive being assessed and by the private equity sponsor, so an invented or altered quote is not a minor error, and it can end the engagement and the client relationship. Accuracy and completeness are the product. The firm needed automation that met that standard rather than approximated it, and it needed a system that moved closer to the firm's own judgment and writing with each engagement, because the firm's method is the asset.
The engine runs each engagement through fixed stages rather than one model call: confirm the engagement goals, extract themes, verify quotes, anonymize, draft recommendations, write the executive summary. Confirming goals first keeps every later stage focused on what the engagement is about. At the stages where accuracy and completeness matter most, theme extraction and quote verification, two models analyze the material in parallel and a third merges their output, so a theme one model misses is caught by the other and a theme one model invents is rejected by the other. Every quote is checked by exact matching against the source transcript; low coverage is rejected and any theme without verified quotes is dropped. The summary's claims are tied to the exact passages they came from, and an audit stage reruns generation until every claim traces to evidence. Interviews about other subjects in the same engagement are scanned for relevant context but used as reference only, so attribution stays clean. The output maps into the firm's report template, so the draft arrives finished rather than as raw text to be reformatted.
Nothing reaches a client without a consultant's review, and every finding is traceable to its source. Every edit the consultant makes, accepting or rejecting a finding, rewording a quote, reassigning a speaker, is recorded with before-and-after values and distilled into rules applied on future runs, so the firm's judgment accumulates in the system instead of being re-applied by hand each time. A separate benchmark scores each AI draft against the firm's own hand-written reports using an independent model as judge, with those reports kept out of every prompt so the score measures the engine rather than its ability to copy.
The engine is in production. Consultants now start from a complete, verified draft and spend their time on judgment, over engagement material that can run to roughly eight hundred thousand characters. The verification stages produce a report that meets the firm's accuracy standard by construction, not by proofreading, and the edit capture means each engagement improves the next. For the firm this turns its method into an asset it can apply at scale, and for any firm whose product is a confidential written judgment, the same design, verified sources, traceable claims, review before release, and edits that teach the system, is what makes AI drafting acceptable at all.