Overview / Platform / Cockpit
Operations made visible

See what AI costs
and what it delivers.

AI without a cockpit is a black box on the bill. Nexograph makes every euro, every workflow, and every action visible: costs per model, department, and operation, agent workflows from prebuilt prompts, and a complete audit trail — from budget to individual model call.

01 / Cost Control

Every euro has an address.

The usage cockpit breaks down every AI output — by model, department, and operation, over 7, 30, or 90 days. No "the AI was expensive," but rather: which AI, for whom, in which process.

Usage drill-down · Tenant: acme-group Live · last updated 2 min ago
Period
Break down by
128.430
Events
▲ 6.2 % vs. previous month
41,8 M
Input Tokens
Ø 326 / Event
9,1 M
Output Tokens
Ø 71 / Event
50,9 M
Total Tokens
across 6 models
€ 2.847
Costs (30 d.)
▼ 18 % after routing

Top by token volume — per model

30 days · 6 models
GPT-4o-miniAzure · small/fast
€ 312 / 22,4M
Claude Sonnet 4.5Reasoning
€ 1.044 / 14,2M
GPT-4oAzure · Standard
€ 921 / 8,6M
Llama 3.1 70BPrivate · self-hosted
€ 0 / 4,1M
embedding-3-largeVectorization
€ 41 / 1,3M
Claude Haikucheap/Batch
€ 529 / 0,3M
Details — by model
ModelEventsInputOutputTotalCosts
GPT-4o-mini71.24020,1 M2,3 M22,4 M€ 312
Claude Sonnet 4.522.88011,9 M2,3 M14,2 M€ 1.044
GPT-4o14.1107,2 M1,4 M8,6 M€ 921
Llama 3.1 70B (Private)12.6403,6 M0,5 M4,1 M€ 0
embedding-3-large6.3101,3 M1,3 M€ 41
Claude Haiku1.2500,2 M0,1 M0,3 M€ 529
Total128.43041,8 M9,1 M50,9 M€ 2.847

Top by token volume — per department

30 days · 5 workspaces
Sales workspaceQuotes · CRM
€ 988 / 18,1M
ServiceSupport responses
€ 612 / 12,9M
ComplianceReview · Evidence
€ 701 / 9,4M
ControllingReporting
€ 358 / 6,6M
ProcurementTender · Contracts
€ 188 / 3,9M
Details — by department / workspace
WorkspaceActivityEventsTotalCosts
Sales workspaceQuotes, CRM research, cross-selling46.12018,1 M€ 988
ServiceSupport responses from knowledge base33.48012,9 M€ 612
ComplianceEvidence review, source checks21.9009,4 M€ 701
ControllingAnalyses, report drafts16.7406,6 M€ 358
ProcurementTenders, contract analysis10.1903,9 M€ 188
Total128.43050,9 M€ 2.847

Top by token volume — per operation

30 days · 6 operations
graphrag:queryKnowledge query
€ 1.121 / 19,8M
chatStandard chat
€ 486 / 11,4M
agent-mode:queryAgent run
€ 798 / 8,7M
document:ingestDocument ingestion
€ 211 / 5,8M
summarizeSummary
€ 190 / 3,9M
embeddingVectorization
€ 41 / 1,3M
Details — by operation
OperationEventsInputOutputTotalCosts
graphrag:query52.30016,4 M3,4 M19,8 M€ 1.121
chat38.9109,3 M2,1 M11,4 M€ 486
agent-mode:query9.6407,1 M1,6 M8,7 M€ 798
document:ingest14.8805,5 M0,3 M5,8 M€ 211
summarize6.0903,2 M0,7 M3,9 M€ 190
embedding6.6101,3 M1,3 M€ 41
Total128.43041,8 M9,1 M50,9 M€ 2.847
Budget guardrails

Limit per request & workspace

Every logical request has a token/cost budget — including all retries. If a workspace runs hot, the limit kicks in before the bill does.

Routing saves real money

Cheapest suitable model

The multi-LLM router sends simple tasks to small models and sensitive data to the private LLM. In the example: 18 % lower costs at the same quality.

Cause-based allocation

Cost center assignment

Every event carries department, project, and operation. The analysis can be exported — as a basis for internal chargeback and ROI proof.

02 / Agent Workflows & Prompts

From prebuilt prompt
to orchestrated workflow.

Nobody has to program agents. Tested prompt templates describe the task in plain language — Nexograph turns that into an agent, connects several into a workflow, and executes them traceably step by step.

STEP 1 — SELECT PREBUILT PROMPTS
Research AgentTemplate
“Search the data room for all contracts with customer X. Extract terms, notice periods, and special conditions. Back up every statement with a source.
Tools: GraphRAG · Documents
Analysis AgentTemplate
“Assess the findings for risks & red flags. Prioritize by business relevance. Ask the human about critical points.”
Tools: Reasoning · Human-in-the-Loop
Report AgentTemplate
“Create a structured memo from the results — following the company template, with bibliography and open points.”
Tools: Document export · Custom API
STEP 2 — AGENTS ARE CONNECTED INTO A WORKFLOW
Workflow Supervisor Agent-as-Tool Approval
Supervisororchestrates · 3/3 Research AgentGraphRAG · Documents Analysis AgentReasoning · Risks Source ToolAgent-as-Tool Report AgentMemo · Export Appr-oval
STEP 3 — EVERY EXECUTION IS LOGGED STEP BY STEP
Execution · Research workflow
7 steps18.420 tokens00:42 duration completed
#1 Reasoning
3,1 s
#2 RagQuery
5,8 s
#3 ToolCall
2,4 s
#4 SubAgent
4,9 s
#5 Reasoning
1,9 s
#6 ToolCall
2,8 s
#7 Final answer
memo.pdf
Reasoning RagQuery ToolCall SubAgent Final answer
  • From plain language, not code: tested prompt templates per use case — customizable, versioned, reusable.
  • Controlled connections: supervisor, agent-as-tool, parallel and conditional paths — with approval steps at critical points.
  • Nothing runs in the dark: every reasoning and tool step is visible with input/output, duration, and token costs.
03 / Audit Trail

Nothing happens without a trace.

Every security-relevant action writes an immutable audit event — who, what, when, with which model. From enabling agent mode to the individual tool call, everything remains traceable and auditable.

TimeLevelAudit eventActor
11.06. 09:14:02InfoAUDIT:DataRoomAccessGranted — Compliance workspace, data room “Contracts”m.weber@acme
11.06. 09:14:48InfoAUDIT:AgentExecutionStarted — Research workflow · model Claude Sonnet 4.5m.weber@acme
11.06. 09:15:21InfoAgentShell▶ exec=7f3a cmd=grep exit=0 stdout=2,4 KB duration=2380 msagent:researcher
11.06. 09:15:55WarnAUDIT:HumanApprovalRequested — Red flag in contract #4821, approval pendingagent:analyst
11.06. 09:17:10InfoAUDIT:HumanApprovalGranted — Approval granted, workflow continuedm.weber@acme
11.06. 09:17:44InfoAUDIT:AgentExecutionCompleted — 7 steps · 18.420 tokens · Export memo.pdfagent:reporter
11.06. 14:02:09WarnAUDIT:TenantAgentModeDisabled — Agent mode disabled for tenantadmin@acme
Who

Authenticated user or agent — every line carries the responsible actor.

What

The specific action — access, run, tool call, approval, configuration change.

When & With What

Timestamp plus model, data room, and cost of the operation.

Documented

Every AI response carries its sources — auditable down to the document.

IMMUTABLE

Append-only

Events are only appended, never overwritten — manipulation stands out.

EXPORTABLE

For audits & supervision

Complete history per tenant — as evidence for GDPR, internal audit, and financial auditing.

OBSERVABLE

Telemetry included

Spans and metrics per run — anomalies such as rate limits or rejected tools are visible.

Costs & control

Your AI budget deserves a cockpit.