The Coherence Gap: Catch Strategic Drift With a Weekly Doc-Reading Agent

Most organizations discover strategic misalignment the expensive way. Two teams ship contradictory features. A platform migration collides with a launch. An architecture decision quietly invalidates three months of roadmap planning. Every one of those incidents was preventable — the conflict was sitting in documents nobody compared side by side.

The coherence gap is the distance between what separate teams believe to be true about shared goals, timelines, and constraints. It doesn't grow from explicit disagreement — that gets aired in planning. It grows from semantic drift: the same strategic vocabulary, used in good faith, decoded into different operational definitions. One team reads "scalable" and pictures horizontal autoscaling. Another reads the same word and plans a monolith refactor with bigger machines. Both are internally consistent. Neither knows the other exists.

A 2025 survey by Mural and The Martec Group, covering 350 marketing, sales, and R&D professionals across the US and Europe, found that 85% of GTM teams report regularly working toward different goals — despite the same 85% feeling confident about their collaboration.^[9] The paradox is exact: high confidence, high misalignment, coexisting in the same room.

A weekly document-scanning agent closes the gap before it costs a quarter. Not by resolving conflicts — by making them specific enough that the right humans cannot pretend they did not see them.

Semantic drift is the operational definition that two teams never reconciled because they were never in the same document at the same time. Same vocabulary. Different mechanics underneath.

A real shape: Team Alpha's PRD calls for "multi-tenant isolation" in the billing service. Team Beta's ADR pins "tenant-level data separation" as a Q3 milestone. The platform OKRs measure success by "customer onboarding velocity." All three documents touch the same surface. None defines where tenant isolation ends and onboarding speed begins. The billing team implements strict row-level security. Every API call costs an extra 200ms. The onboarding team's latency targets break the day the migration ships.

No standup catches this.^[7] It lives in the assumptions between paragraphs of documents written weeks apart by people who never sat at the same table.

Zalando's engineering team discovered the same pattern when they applied AI to two years of postmortem data: the most persistent failure mode wasn't missed dependencies recorded in tickets — it was implicit assumptions embedded in narrative text that no automated check ever read.^[11] Their AI pipeline found that root cause analyses systematically reflected team-level assumptions that went unspoken in the actual incident record.

The coherence agent runs on a fixed weekly cadence. It pulls the latest versions of strategic documents from each team, normalizes them into a shared semantic representation, then performs cross-document comparison to identify assumption conflicts. Output: a coherence brief delivered to leadership and team leads.

The agent surfaces. Humans resolve. Resolution requires judgment, context, and organizational authority. The agent's job ends at making conflicts visible with enough specificity that the right people can't route around them.

The key technical challenge is not retrieval — pulling documents from Notion or Confluence is straightforward. It's extraction: getting from "what the document says" to "what the document assumes." Explicit claims are easy to compare. Implicit ones require a prompt architecture designed around the specific ways your organization's documents embed undeclared constraints.

The hardest problem in coherence detection is the gap between what a document states and what it takes for granted. A PRD might state "the recommendation engine will use collaborative filtering" (explicit claim) while assuming "the data pipeline delivers fresh user events within 5 minutes" (implicit dependency). If the data team's ADR pins batch processing on a 6-hour cycle, the implicit assumption is already broken. No keyword search catches it.

Reliable detection requires a claim taxonomy the agent can extract against. The taxonomy teaches the model where implicit assumptions live — which sentences are load-bearing and which are decoration. Chain-of-thought prompting works substantially better than simple structured extraction for this task: asking the model to reason step-by-step about what a document takes for granted before listing explicit claims surfaces assumptions that direct extraction misses.^[13]

Building this taxonomy is the calibration work that most teams skip. Generic prompts produce generic results.

A direct prompt — "list all assumptions in this document" — fails for a predictable reason: the model treats the instruction as a request for what is unusual, not for what is taken for granted. Assumptions that are deeply embedded in an author's mental model produce no signal in the text that marks them as assumptions. They look like facts.

Zalando's two-year AI postmortem project discovered the equivalent failure mode in incident analysis: LLMs would attribute root causes to technologies mentioned prominently in the text, regardless of actual causality. Their fix was a multi-stage pipeline where separate stages handled description, causality classification, and recommendation — preventing the model from short-circuiting reasoning by picking the most salient noun.^[11]

The same architecture applies to assumption extraction. Build stages that are solely responsible for a single extraction task:

1. Description stage — summarize what the document explicitly claims
2. Gap analysis stage — for each claim, ask: "what must be true for this to hold that the document does not state?"
3. Classification stage — categorize each extracted gap using the claim taxonomy
4. Cross-team comparison stage — only now compare across boundaries

Keeping these stages separate prevents the model from anchoring on surface similarity and missing mechanically divergent assumptions buried underneath.

An agent that flags every conflict equally is a noise generator. The product is the ranking. Blast-radius estimation scores each conflict by how many teams, timelines, and customer commitments will absorb damage if it stays unresolved.

Three factors drive blast radius. Dependency depth — how many downstream teams rely on the assumption. Time proximity — how soon the contradiction materializes in production. Reversibility — whether the decision can be undone cheaply or requires a rewrite.

What we got wrong on the first build: the original scoring weighted semantic divergence highest. That produced a brief where the top-ranked conflict was a definitional difference that two teams had already verbally resolved but hadn't updated in their documents. The actual critical conflict ranked lower — a timeline assumption that would break three production services in six weeks. Blast-radius scoring only became useful once we weighted time proximity above semantic distance. Semantic difference tells you the conflict exists. Blast radius tells you which one to fix this week.

The software engineering field uses blast radius for exactly this kind of risk containment: not every change has the same risk profile, but without explicit scoring, small and large-impact conflicts go through the same process pipe.^[14]

Each weekly brief carries three sections: a summary of the coherence score trend with contributing factors; a ranked list of detected conflicts with full source context; and a proposed resolution format for each one.

The resolution format is not decoration. Not every conflict warrants a meeting. Some need an async clarification in a shared document. Others require an ADR revision. The high-blast-radius, low-reversibility conflicts need a sync with decision-making authority physically in the room. Picking the wrong format is how good detection turns into wasted hours.

The brief also tracks recurring misalignment patterns. A conflict that reappears week over week — even after resolution — signals a structural cause: a shared glossary that doesn't exist, a dependency nobody owns, or a planning process that generates ambiguity by design. These patterns are worth more than any single conflict finding.

Pairwise comparison stops scaling fast. Ten teams produce 45 pairwise relationships. Thirty teams produce 435.^[8] Comparing every document against every other document burns tokens and surfaces noise. The agent needs a smarter topology.

Topical clustering with boundary detection is the strategy that holds up. The agent first groups documents by the strategic topics they address — billing, onboarding, data platform, observability — then concentrates comparison on documents that straddle topic boundaries. The highest-value conflicts live at those boundaries, where two teams touch the same system from different angles with different assumptions. Comparing inside a cluster catches drift the cluster owners would have caught themselves. Comparing across clusters catches drift nobody is currently watching for.

This means the dependency graph is structural input, not just metadata. You need to know which teams share systems before you can know which document pairs deserve full comparison. Build the graph once; the agent uses it every week to filter pairwise candidates before any embedding comparison runs.

ROI for coherence detection lives in conflicts caught before they materialized as incidents. Three numbers prove value. Conflicts surfaced per weekly brief. Resolution time from detection to decision. Production incidents and delivery delays that trace back to assumption mismatches — month-over-month, before and after the agent shipped.

Organizations running coherence analysis have reported reductions in cross-team coordination failures. Those figures depend heavily on document quality, team participation, and calibration effort. Treat them as directional. The actual mechanism isn't that the agent is smarter than humans. It's that the agent is more consistent. It reads every document, every week, and never assumes two teams have already talked.

For a team with $10M in payroll, the research on organizational misalignment suggests wasted effort in the $2M–$3M range annually.^[10] A coherence agent running weekly costs roughly $5–15 per run in LLM inference for 50–100 documents — under $1,000 annually. The infrastructure cost is negligible against a single avoided misalignment incident.

The coherence gap isn't a failure of communication. It's a failure of comparison infrastructure. Teams produce the signal — it's in the documents they write every quarter. The agent's only job is to read those documents in the same week, across the same team boundaries, and produce a ranked list of the conflicts nobody has yet named out loud. The brief lands. Humans decide. The quarter ships without the drift.