Structured input, constrained output, human validation.

Shared operating principle

Both workflows follow the same shape: structured input and context, constrained processing and generation, validation, then human review. The differences are what goes into the model and how evidence is handled.

1. Step 01

   Structured input and context

   Each workflow starts from what you deliberately provide: a fixed form, selected documents, or both. The Workbench does not scrape your systems or retrieve outside information about your organization. Organization-specific context comes from what you submit, while generated assumptions and inferences still require validation.

   You must be signed in. Unauthenticated requests are rejected before any model or embedding call is made.

2. Step 02

   Constrained processing and generation

   Model calls run on the server. The browser never receives provider credentials and never contacts a model provider. Generation asks for a specific structure rather than open-ended advice.

   Embedding calls for Organization Discovery also run server-side, against document chunks prepared after upload.

3. Step 03

   Validation before anything is treated as a result

   Responses are parsed and checked against the workflow schema. Organization Discovery also validates citations against the evidence that was retrieved, and applies grounding and advisory checks.

   A response that fails required schema or blocking checks is treated as a failure rather than shown as a finished result. Advisory checks are recorded with the result.

4. Step 04

   Human review

   Each workflow makes uncertainty visible in a form suited to the result. Shelter surfaces assumptions and validation questions. Organization Discovery shows evidence, citations, and gaps in the available evidence. The Workbench does not apply changes to your organization.

   Telemetry records how a run behaved; feedback lets you mark whether a result was useful. Neither replaces human judgement.

Shelter AI Assessment

A form-based assessment for animal shelters and animal welfare organizations. It does not retrieve uploaded documents.

1. Shelter 01

   Twelve form inputs

   You describe intake, status tracking, adoption, volunteers, foster support, repeated questions, documents and software in use, repetitive tasks, sensitive data, and human-owned decisions. All of it is free text.

2. Shelter 02

   Server-side model call

   Those answers are sent from a route handler to the configured model provider. No uploaded documents are retrieved for this workflow.

3. Shelter 03

   Structured assessment and controls

   The draft includes a workflow map, opportunities, pilot recommendation, roadmap, assumptions, validation questions, a do-not-automate list, and guardrails. Schema validation runs before persistence.

4. Shelter 04

   Telemetry and feedback

   A successful run stores the assessment under your organization, records run metadata for traceability, and accepts result-level feedback.

See a saved fictional example: Smiley Animal Welfare Center.

Organization AI Opportunity Discovery

Evidence-grounded opportunity discovery from organization context and selected operating documents. Documents are organization-scoped. Generation uses your submitted context and the retrieved excerpts from the documents you selected for that run, not every file your organization has ever uploaded.

1. Organization 01

   Organization context and selected documents

   You complete a context form and choose which of your organization’s uploaded documents are in scope. Documents you do not select are not searched for that run.

2. Organization 02

   Extract, chunk, and embed

   After upload, the server extracts machine-readable text, splits it into chunks, and embeds those chunks so they can be retrieved later. Embedding uses the configured embedding provider.

3. Organization 03

   Themed semantic retrieval and evidence assembly

   At report time, the Workbench retrieves relevant chunks from the selected ready documents, assembles an evidence set, and builds a grounded generation prompt from that evidence plus your form context.

4. Organization 04

   Structured report, citations, and checks

   The model returns a structured opportunity report. Citations are validated against the retrieved evidence. Grounding and advisory checks run before the report is persisted with its evidence snapshot.

5. Organization 05

   Telemetry and feedback

   Run metadata, evidence counts, and related diagnostics are recorded for traceability. You can mark whether the report was useful.

Opportunities can be classified as AI, deterministic automation, or process improvement, with an evidence basis of documented, inferred, or insufficient evidence. A single report is not required to contain every category.

See a saved fictional example: EveryWorks Supply.

Versions and traceability

Authenticated results can show which workflow, prompt, schema, and (for Organization Discovery) retrieval version produced a draft. Those labels exist so runs are comparable over time. They are not a claim that the output is correct.

• Shelter: shelter-workflow-v1, shelter-assessment-v3, shelter-assessment-v1
• Organization: business-workflow-v1, business-report-v4, business-report-v1, business-retrieval-v1

Why the output is deliberately conservative

A model asked an open question about an organization it has never seen will produce confident, plausible, unfalsifiable advice. Constraining the output to a schema, requiring assumptions, requiring a do-not-automate list where the workflow asks for one, and requiring citations against retrieved evidence are ways of forcing the result to expose its claims, so a reader can disagree with a specific statement rather than with a wall of prose. That is not the same as exposing the model's internal chain of thought.