Discover.
Weeks of context curation now automated and repeatable.
Every snippet, diagram, quote, and number on this page comes from an actual run of the Aurora L01 Workflow Intelligence automation against a synthetic Keystone Analytics engagement. Layer 01 consolidates context the organization already holds — operator notes, transcripts, Slack threads, architecture diagrams, process documentation — together with focused interviews conducted by Northbeam. It synthesizes the true current state across both process and architecture, surfaces operational deviations that live only in tribal knowledge, names every gap on the record, and produces a signed Future-State Blueprint with a ranked AI Opportunity Register and safety ratings. Once it starts, it isn't a team pulling information together. It runs.
Discovery doesn't fail because there's no information.
It fails because no one reconciles what's already there.
Every organization already holds the answer in scattered form. The work isn't gathering — it's triangulating, surfacing contradictions, and converting evidence into a signed commitment about what to build.
The diagram doesn't match the system.
The architecture doc says one thing; the production code says another. The gap is six months old and nobody noticed because nobody opens the dashboard anymore.
The "official" workflow isn't the real workflow.
The CS team has been operating from a personal Notion page maintained by an analytics engineer over Slack DM. Eight weeks of cadence. Zero documentation. This is where decisions actually get made.
Every team has a different number.
The CRO says churn is 11.4%. CS Ops says 14.1%. Both are technically defensible. The board memo is due in two weeks. Nobody is positioned to give the CEO the "real" number.
AI opportunities get assessed last — and emotionally.
By the time someone asks "is this safe to automate?", the deck is half-built and the budget is committed. Safety becomes a constraint to argue around, not a rating that gates work.
Drop what you already have.
No new fieldwork. No new interviews to schedule. We work from the context the organization has already produced.
Sample run shown above. Production engagements accept whatever the organization already produces — Loom recordings, Granola transcripts, Confluence pages, Linear tickets, Gong calls, Box folders, support ticket exports. The methodology is format-agnostic.
The full operational surface — mapped in one pass.
L01 doesn't stop at the dashboard. The Keystone run surfaced thirteen systems across four operational layers — sources, models, delivery, supporting — and located every one of them in the documented vs. discovered map. The downstream charter (L02) and build (L03) inherit this exact system surface.
- Salesforce
- Gainsight
- Segment
- Zendesk
- Snowflake
- dbt Cloud · 340 models
- marts.customer_health_v2 ⚠
- Hex dashboard · abandoned
- Hightouch → Gainsight + Slack
- Gainsight writeback
- Metabase embeds
- Datadog
- Linear · data-eng queue
- Notion · ops docs + shadow
- Slack · #data-customer-health
Every system appears in the L01 dossier with: documented state, discovered state, ownership, last-modified timestamp where knowable, and inclusion/exclusion verdict for the workstream register. This is the surface area L02 acceptance criteria reference and L03 integration proofs cite.
The raw context, before synthesis.
Three real snippets from the Keystone inputs. Messy. Partial. Each one captures information that exists in the organization but isn't in any single document.
marts.customer_health_v2 was last modified 8 months ago. It uses a Salesforce account-id mapping that broke in November when Diego restructured CRM ownership. Three of its five inputs return nulls now. Nobody's caught it because nobody actually opens the Hex dashboard anymore."Three different sources. Three different vocabularies. Same underlying problem. Without L01, this stays scattered. A team would spend weeks reconciling these and still leave dependencies tribal. Aurora L01 reads all of them in one pass, triangulates the claims, names the gap on the record, and produces a workstream ranked against value.
A signed blueprint —
with every claim citable to its source.
A single integrated artifact with four parts: workflow + architecture maps, AI opportunity register with safety ratings, value model with confidence intervals, and a dependency graph. Every assertion carries a citation. Two signatures lock it.
Workflow and architecture, rendered together. Documented connections in cyan. The shadow path the CS team actually trusts in violet. The broken connection in red. All three live in one render so the engagement starts with the same picture in every head.
Every AI opportunity gets a safety rating —
before specification starts.
The methodology refuses to defer this question. Three ratings. Clear meaning. Operational consequences in every layer that follows.
In the Keystone run: seven opportunities surfaced. All seven cleared to specification with appropriate ratings; one candidate ("fully-autonomous customer outreach calls") was reviewed and refused at L01 before reaching the charter. The refusal is on the record.
Eight specialists.
Each enforcing a discipline the others can't override.
Ingest
Heterogeneous handlers normalize every input — markdown, transcripts, diagrams (vision-extracted), structured exports — into evidence-bearing claim units. No source goes in without provenance.
Extract
Every claim is paired with its source quote, line number or thread reference, extraction confidence, and the agent that surfaced it. The audit trail is built before the first conclusion is drawn.
Triangulate
Single-source claims are flagged. Cross-source claims are confirmed or contradicted. The Triangulator's job is to attempt and fail before allowing a single-source claim to influence ranking.
Gap Analysis
Documented state versus discovered state. Process doc says X; production system does Y. Every delta gets a gap ID, evidence, and an inclusion-or-exclusion decision for the workstream register.
Rank
Every candidate workstream scored on cost avoided, productivity gained, revenue protected — with confidence intervals tied to evidence. Plus AI suitability (Green / Yellow / Red), feasibility, and strategic fit.
Boundary Discipline
The Boundary Sentry names what's not in scope and why. Refused AI opportunities get their refusal logged. This is the audit trail your security and audit teams will want when the program ships.
Assemble
Workflow map, architecture map, opportunity register, value model, dependency graph — assembled into one render. The same picture in every stakeholder's head.
Pre-Signoff Audit
Citation integrity verified before the artifact reaches signers. Broken citations don't make it to the boardroom. Two signatures required — one operational, one executive — to advance.
Weeks of discovery
now automated and repeatable.
The compression isn't from cutting rigor — it's from removing the steps where humans were doing what evidence triangulation can do: cross-reference, contradiction-surface, citation-track, gap-name. The judgment calls stay in human hands. The wiring underneath runs automatically. Once the context is dropped, the synthesis isn't a project plan — it's an agent run that produces a signed artifact at the end.
Time to gather inputs is a separate task. Once inputs are in hand, Aurora L01's runtime is measured in hours, not weeks. The same agent run against the same inputs is reproducible.
L02 takes the signed blueprint
and turns it into a binding spec.
See how the methodology turns "this is what we discovered" into "this is what we will build — and what we refuse to."