Architecture review boards run on preparation: a pre-read the members actually read, the objections anticipated before they are voiced, and — after any workshop — a synthesis the attendees recognize as what happened. AI compresses all three from hours to minutes; validation by the humans in the room stays mandatory.
Tier: Foundation · Hands-on time: ~20 minutes
Meeting preparation and synthesis are low-risk, high-frequency AI work — summarization-shaped, human-checked, and repeated weekly for anyone who owns or faces an ARB. The compounding trick is the knowledge base: an ARB's past findings are its personality, and retrieval over prior decisions (pb-18) turns objection anticipation from guessing into pattern-matching against what this board has actually said. Vikram, who chairs a BFSI ARB, uses the same playbook from both chairs: preparing submissions to face the board, and preparing the board's own pre-reads.
Skip AI when the review is informal and the design fits on one page — over-produced preparation for a hallway decision reads as defensiveness.
Draft the pre-read as a decision instrument, not a summary. Prompt with the artifacts and the board's template: "Draft the ARB pre-read: the decision requested (one sentence, first), context in five lines, the options considered with the matrix summarized, the recommendation with its top three reasons, risks with mitigations, and what approval unblocks. Length limit [board's]. Every claim must trace to the attached artifacts — no new claims." The pre-read's job is to let a member arrive already convinced or already precise about their objection; both outcomes beat a cold room.
Mine past findings for likely objections.
Query: the attached pre-read proposes a managed-agent platform with
an external LLM API dependency for a document-processing workflow.
From the ARB findings corpus (retrieved excerpts attached), list:
1. Past findings this board has raised on similar submissions —
verbatim finding + which submission + outcome.
2. The 5 most likely objections to THIS submission, each grounded in
a cited past finding or a stated board standard.
3. For each objection: what evidence in our artifacts answers it, or
NOT COVERED if we have no answer staged.
Do not invent board positions — every predicted objection cites its
precedent.
The NOT COVERED list is the real product: each entry is either evidence to stage before the meeting or a concession to make gracefully in it.
Rehearse the two hardest objections. Pick the two you least want to hear and have the model play the objector in character — sustained, with follow-ups, using the cited precedent. Answer out loud or in writing. This is presales-grade preparation applied to governance: the difference between "good question, we'll take that away" and an answer with evidence attached, delivered calmly.
Run the meeting with capture in mind. Whoever scribes, capture decisions verbatim as they are made ("agreed: X, subject to Y, owner Z") — synthesis quality is set during the meeting, not after. For workshops, photograph every board and collect the sticky-note walls before they are dismantled.
Synthesize within 24 hours, structured for validation. "From the transcript, notes, and whiteboard photos: produce (1) decisions taken — verbatim where possible, with any stated conditions; (2) open questions with the owner who took each; (3) actions with owner and date; (4) risks raised but not resolved; (5) ATTRIBUTION UNCERTAIN — anything you cannot confidently assign to a speaker or an outcome. Do not resolve ambiguity by choosing the likelier reading; list it." The uncertain bucket is what makes the draft validatable instead of quietly wrong.
Validate with attendees before circulation — always. Send the draft to attendees (or the chair plus one dissenter, minimum) with a deadline: "correct anything that is not what happened." Only after that pass does the synthesis circulate as the record, and decisions in it flow onward — ADR drafts via pb-14, actions into the tracker, unresolved risks into the register. An unvalidated synthesis circulated as minutes is how organizations remember things that did not occur.
ARB materials and workshop records are internal-confidential at minimum; client review boards add MSA constraints. Run drafting and synthesis on the approved tenancy (enterprise, or the client's Bedrock/Azure OpenAI subscription for engagement boards); meeting transcripts especially should never transit consumer tools — they carry unguarded statements by named people. For BFSI-grade environments like Vikram's, a local Ollama model handles synthesis structure well since the transcript supplies all content. Workshop photos with client whiteboards count as client data; check the pb-03 register before any multimodal upload, and prefer transcribing the boards to text locally first.
The pre-read earns the meeting, the anticipated objection defuses it, and the validated synthesis is the only version of the meeting that should survive. AI drafts all three; the attendees' confirmation is what makes any of it the record.