Skip to content

Deployment method

Deploying WorkGraph, step by step.

Two graphs, built from your systems, then joined. The sources, connections and processing, for your IT lead.

  1. Step 1 · Work GraphBuild the graph of work.
  2. Step 2 · AI GraphTrace AI compute to processes.
  3. Step 3 · Value GraphJoin the two graphs.

ScopeA cross-functional process, such as lead-to-cash, or a whole business unit. The method does not change: the unit of analysis is the process. Scope only sets how many sources to connect.

Step 1 · Work Graph

Build the graph of work.

Organisation, documentation and volumes come from your systems. The questionnaire fills in what no system sees.

  1. Your systems

    • HRIS
    • Process maps, SOPs
    • CRM, ticketing, ERP
    • Guided questionnaire

  2. WorkGraph processing

    • Taxonomy
    • Semantic mapping
    • Time allocation
    • Reconciliation

  3. Result

    • Processes → activities → tasks → roles
    • FTE, loaded cost, volume, unit cost
    • Source and confidence per figure
  • HRIS: Workday, SAP SuccessFactors, Lucca, PayFitConnection ·REST API or scheduled CSV export (SFTP)Access needed ·Read-only service accountData read ·Position, role, entity, reporting line, working time, loaded cost band
  • Repositories: Signavio, ARIS, Confluence, SharePointConnection ·BPMN 2.0 export (XML), filesAccess needed ·Export or uploadData read ·Documented processes, SOPs, job descriptions
  • Business tools: Salesforce, HubSpot, Zendesk, ServiceNow, Jira, ERPConnection ·Read-only REST API, or CSV exportAccess needed ·Read-only token, listed objectsData read ·Object counts by type, status and month; created and closed dates
  • Existing process mining: Celonis, Power AutomateConnection ·OCEL 2.0 or CSV exportAccess needed ·ExportData read ·Event logs, if they exist
  • WorkGraph questionnaireConnection ·Web applicationAccess needed ·SAML 2.0 or OIDC SSOData read ·Tasks, triggers, frequencies, durations, tools

Steps

  1. 1Organisation ingestion

    Roles, headcount and average loaded costs loaded into PostgreSQL. Every employee ID is replaced with an HMAC whose key stays with you.

  2. 2Taxonomy

    The APQC framework (levels 3 and 4) is the skeleton, merged with your BPMN maps. An LLM extracts candidate activities from SOPs; the sponsor signs off.

  3. 3Collection

    A 20-minute guided questionnaire behind your SSO. Structured answers: task, trigger, input, output, tool, frequency, duration.

  4. 4Mapping

    Embeddings and an LLM map every answer to an activity. Duplicates are merged, time normalised to 100% per person.

  5. 5Volumes

    Connectors count business objects per activity. Implied FTE = volume × unit duration ÷ hours worked.

  6. 6Reconciliation

    Self-reports, manager estimates and the volume model are compared. Above a 30% gap, validation in the app; every decision is logged.

Storage: a relational graph in PostgreSQL, with nodes, typed edges and sourced estimates. OCEL 2.0 export available.

Work Graph

Step 2 · AI Graph

Trace AI compute to processes.

Every type of AI leaves a different trace. We connect the right source to each, read-only, without content.

  1. Your systems

    • Assistant admin consoles
    • Agent telemetry
    • LLM gateway
    • Billing APIs

  2. WorkGraph processing

    • Pseudonymisation
    • Process attribution
    • Cost calculation and reconciliation

  3. Result

    • AI spend by process, day, model
    • Attribution mode and confidence
    • Unattributed share, shown
  • ChatGPT EnterpriseConnection ·Workspace analytics, Cost APIAccess needed ·Workspace admin keyData read ·Usage and credits per user, product, model
  • Microsoft 365 CopilotConnection ·Microsoft Graph Reports API, Office 365 Management Activity API (Purview audit)Access needed ·Entra ID app: Reports.Read.All, ActivityFeed.ReadData read ·Usage per user; app, context and agent of each interaction, no text
  • Claude Enterprise, Claude CodeConnection ·Analytics API, OpenTelemetry exportAccess needed ·read:analytics key, OTEL_* variablesData read ·Tokens and cost per user per day; per prompt for Claude Code
  • GitHub Copilot, CursorConnection ·Usage metrics API, Admin APIAccess needed ·Organisation token, metrics readData read ·Usage per user, team, repository
  • Internal agents and applicationsConnection ·OpenTelemetry GenAI SDK to a collector on your sideAccess needed ·Deploy a collector containerData read ·Spans: model, tokens, tools, workgraph.process_id
  • Uninstrumented API trafficConnection ·LLM gateway: LiteLLM, Cloudflare AI Gateway, or yoursAccess needed ·Route calls through the gatewayData read ·Tokens, cost, virtual key, tags
  • Agent platforms: Copilot Studio, Agentforce, n8n, ZapierConnection ·Consumption reportsAccess needed ·Admin exportData read ·Consumption per agent
  • Invoices: OpenAI, Anthropic, AWS Bedrock, Azure, VertexConnection ·Usage & Cost APIs, cloud billing exportsAccess needed ·Admin key, billing exportData read ·Amount billed per project, workspace or tag

Steps

  1. 1Inventory

    AI tools in place, billing mode (seat or usage), admin access available. That is the 30-minute audit.

  2. 2Admin connectors

    Daily extraction of assistant usage and invoices. User identity pseudonymised on intake, joined to the role through the HRIS.

  3. 3Agent telemetry

    An OpenTelemetry collector in your environment. The process tag is set at the workflow entry point and inherited by every span. A processor strips prompts and responses before export.

  4. 4Attribution

    Explicit when the tag is present. Otherwise by mapping: an agent, key, project or repo → process table, set once. By allocation for assistants: by the role’s time in the Work Graph.

  5. 5Cost

    Estimated cost = tokens, cache included, × a versioned price table. Reconciled nightly with the amount billed per project or workspace.

  6. 6Unattributed

    Anything no mode covers stays shown as unattributed, in euros and as a share. Never spread out to look complete.

Setting the process tag
# Agents: at the workflow entry point (Python)from opentelemetry import baggage, contextctx = baggage.set_baggage("workgraph.process_id", "lead_to_cash")ctx = baggage.set_baggage("workgraph.activity_id", "lead_qualification", ctx)context.attach(ctx)  # every child span inherits the tag # Packaged tools (e.g. Claude Code): per deploymentOTEL_RESOURCE_ATTRIBUTES=workgraph.process_id=customer_support

Storage: aggregated facts by day × AI asset × model × role × process. OpenTelemetry GenAI conventions (gen_ai.*) plus a workgraph.* namespace.

AI Graph

Step 3 · Value Graph

Join the two graphs.

No new connection: the Value Graph is computed from the first two, plus your software and external costs.

  1. Your systems

    • Work Graph
    • AI Graph
    • Accounting export: software, contractors

  2. WorkGraph processing

    • Join by process
    • Scoring
    • Baseline

  3. Result

    • Operating statement per process
    • Priorities: value × transformability
    • Value identified → realized

Steps

  1. 1Join

    By process and month: FTE and human cost, AI cost by mode, software, external, volume, unit cost.

  2. 2Additional costs

    Attributable software and contractors, from an accounting export or the contracts.

  3. 3Prioritisation

    A value score and a transformability score, with weights set with you and shown.

  4. 4Baseline

    Unit cost and throughput at the start date. Every opportunity carries its mechanism: cost-out, cost avoidance, revenue, quality.

  5. 5Tracking

    Monthly connector refresh. Realized value is signed off by your financial controllers.

Storage: one view per process and month, recomputed on every validation. CSV export and API.

Value Graph

Security and data.

Four principles, for every connection.

Read-only
No writes to your systems. No agent on devices.
No content
No prompts, responses, emails or tickets. Counts and metadata.
Pseudonymisation
HMAC on intake, key held by you. Reported by role, never below 5 people.
Hosting
European Union, one database per client, encryption at rest and in transit, access log.
Everything about privacy

A first 30-minute call.

Your scope, your data, your timeline. You leave with a pilot proposal.

Your request
Headcount

We only use these details to reply to your request.