SUP-928
Respond to API outage
Resolved · Trace impact and verify recovery.
- People
- Models
- Cost
- $2.84
CortexThe Architecture of High-Yield Inference
AI is metered in tokens. Humans care about outcomes.
Cortex reconstructs work from scattered AI conversations, cross-application artifacts, and token streams so organizations can value, govern, and improve the outcomes.
The semantic engine · animated
Signals become attributed work.
Cortex observes activity, links the surrounding context, and infers Work Units that retain their people, agents, applications, and cost.
Observed activityObserved
Linked contextLinked
Labeled workLabeled
Complete semantic flow:
- Observed activity from Slack, Cursor, Jira, Claude, GitHub, and Notion.
- The Semantic Engine links the surrounding context.
- Labeled Work Units retain people, agents, applications, and cost.
- Related Work Units form the Customer escalation responses Work Stream.
Work StreamCustomer escalation responses
The work model
The system adapts to the organization.
Work topology
Units become streams. Streams advance the roadmap.
Every Work Unit retains its source records. Work Streams stay linked to the roadmap items they advance, alongside Google Workspace, Jira, Notion, and conversation evidence.
Work Units
3 of 38 shown6 of 38 shownSUP-913
Resolve billing dispute
Resolved · Validate charges and the approved adjustment.
- People
- Models
- Cost
- $1.62
SUP-902
Find missing evidence
Reopened · Reconcile the incomplete case record.
- People
- Models
- Cost
- $0.94
+35 more Work Units
Work Stream
Customer escalation responsesResolve urgent customer issues
38 Work Units22 min baseline · 91% accepted
- People
- 12 contributors
- Models
- 126 model calls
- Cost
- Total across 38 WUs · 30d$73.10
Roadmap item & linked records
Google WorkspaceCustomer experience roadmap · Q3JiraSUP projectNotionEscalation runbookSlack#support-escalationsWork Units
2 of 17 shown4 of 17 shownKB-144
Publish incident runbook
Published · Turn the response into reusable guidance.
- People
- Models
- Cost
- $1.76
KB-139
Review escalation playbook
Reviewed · Check steps against recent outcomes.
- People
- Models
- Cost
- $0.88
+15 more Work Units
Work Stream
Support knowledge maintenanceKeep support guidance current
17 Work UnitsMonthly cadence · Maya Voss
- People
- 8 contributors
- Models
- 49 model calls
- Cost
- Total across 17 WUs · 30d$28.70
Roadmap item & linked records
Google WorkspaceService quality roadmap · Q3NotionSupport knowledge baseGitHubRunbook repositoryProduct principles
Quiet by default. Explicit when it matters.
Cortex runs in the background, learns the existing workflow, and surfaces only when it can help or when a decision needs attention.
- 01Observe quietlyLearn from supported capture paths in the tools people already use.
- 02Intervene at decision pointsRoute, approve, restrict, or block only on configured managed paths.
- 03Explain every actionKeep the reason, policy, and supporting evidence inspectable.
Atlas
Atlas maps work, learns its baseline, and ranks what matters.
Repeated Work Units reveal the expected pattern, meaningful variance, and the highest-value place to investigate next.
Cortex workspace · Work Stream
Customer escalation responses
38 related Work Units · 30-day learned baseline · SUP-928 selected
Review owner
Maya Voss · Support operations
Linked evidence
- Slack18 messages
JiraSUP-928- NotionP1 playbook
- GitHubPR #1842
01 · Expected baseline
Evidence-linked response
Expected pattern
- Ticket, playbook, and mitigation are linked before drafting.
- Owner verifies the evidence before the customer send.
02 · Observed variance
41 min observed- Context rebuilt manually across four source systems.
- Two edits added evidence already present in the Work Unit.
03 · Outcome feedback
91% accepted- Positive: owner-approved, linked response accepted.
- Negative: reopened, edited, or missing-evidence response.
04 · AI Opportunity Map
Ranked from observed work
- 01Reduce context reconstruction72/100
- 02Automate evidence verification61/100
- 03Tune the escalation route47/100
Foundry
Foundry intervenes and optimizes.
Atlas supplies the constraint, success measure, and safety boundary. Foundry selects the intervention it can defend.
Foundry decision receiptAtlas context · Customer escalation responses
SUP-928 · customer escalation response
Foundry decision receipt
Work Unit requirements
Constraints travel with the work- Policy
- Linked evidence
- Release gate
- Owner approval
- Quality floor
- ≥ 0.92 verifier
- Cost ceiling
- ≤ $0.18 / call
- Latency target
- ≤ 3.2s p95
Candidate routes
Private eval replay · 250 cases · Jul 28, 2026- SelectedClaudeClaude Sonnet 5
- Policy
- Eligible
- Quality
- 0.94
- Cost
- $0.14
- Latency
- 2.8s
OpenAIGPT-5.6 Terra- Policy
- Eligible
- Quality
- 0.90
- Cost
- $0.11
- Latency
- 2.4s
GeminiGemini 3.6 Flash- Policy
- Eligible
- Quality
- 0.88
- Cost
- $0.07
- Latency
- 1.9s
Route selected
Semantic routing selected
Claude Sonnet 5
Why this route
- ✓ Linked evidence makes the route policy-eligible.
- ✓ Clears quality, cost, and latency requirements.
- ✓ Preserves the owner approval gate.
Replay snapshot
0.94 quality · $0.14 · 2.8s p95
Single model invocation per case; customer send still requires owner approval.
Foundry suite
No silver bullets. Use the right lever.
Governance and security surround every intervention; supported and configured paths define the enforcement boundary.
Use when
Model cost and quality vary by task
Semantic routing
Choose against the Work Unit requirements.
Use when
The request is missing approved context
Just-in-time prompt enrichment
Bring linked evidence into the request at runtime.
Use when
Repeated work has private verifier data
Model distillation
Hill-climb against proven private outcomes.
Use when
An agent exceeds appropriate autonomy
Agentic governance
Apply the right policy and approval boundary.
Use when
Sensitive context approaches an unsafe destination
Semantic security
Control the configured path using full work context.
Map your first Work Stream.
Bring one important workflow. We will trace the work, its outcomes, and the constraints that should govern it.