WORKFLOW BUILDER · BETA GUIDE

Turn a process into reviewable work.

The builder helps invited users configure read-only document workflows. It combines a saved process, a fixed business vocabulary, supplied evidence, and your recorded review. It does not make regulated decisions or execute accounting actions.

1. Sign in and choose a workspace

For a first test, download the synthetic vocabulary JSON and either the billing evidence or document-review evidence. These contain invented information and are safe starting examples.

Use Demos & sign-in to accept your invitation or sign in. Return to My workflows. Your definitions and runs are private to your account inside the selected workspace. Workspace administrators do not automatically receive access to your private documents.

2. Describe the outcome

For example: “Review the engagement log and identify evidence missing before billing.” Choose a starting point: claims case preparation, a healthcare template, or Build from my description. The description is placed into an editable template; the beta does not generate a custom graph with AI.

Healthcare templates. Healthcare payer: prior authorization preparation drafts a cited case summary for a clinical reviewer: the member, the requested service, the criteria version in force on the date of service, the member's own earlier requests and the documentation found or missing. Healthcare provider: denial appeal preparation drafts a cited appeal packet: the denied claim line and reason code, the payer policy version in force, the supporting record, missing documents and the deadline basis. Both include a starter healthcare vocabulary and three synthetic test cases, and require those cases to pass before other runs. Neither approves, denies or decides medical necessity, coding or billing; a qualified person does. Open them directly at /builder?template=prior-auth or /builder?template=denial-appeal, then copy and adapt them. Every workflow you save is your own; the templates are starting points.

Regulatory reporting template. Regulatory reporting: FR Y-9C to FR Y-14Q loan reconciliation drafts a cited reconciliation memo for the regulatory reporting controller: FR Y-9C commercial and industrial loans (BHCK1766) against the FR Y-14Q utilized exposure total (CLCOG075) for one as-of date, the difference, the reconciling items documented for that date, any break over the team's tolerance and any adjustment proposed without support. Its vocabulary uses Federal Reserve MDRM codes, which you can browse and extend in Ontology Studio's Industry models. Its three test cases use an invented company and invented conventions, not real filings or Federal Reserve rules. It never adjusts figures, says a report is ready to file or attests; the controller does. Open it at /builder?template=y9c-y14q-reconciliation.

3. Configure the steps

Edit the purpose, department, expected output, and rules. Add, remove, and reorder extraction or drafting steps. A human review always remains last. Uploaded text and step instructions cannot grant permissions. Each drafting step sees the frozen sources and previous accepted step outputs.

The Guided setup checklist at the top of the editor shows the recommended order: attach an approved ontology release, map inputs and outputs, describe the output and rules, choose review controls, add a test case, then save and do a sample run. Go jumps to each part.

Review controls. Choose whether the person who runs the workflow reviews it, or a named colleague (a separate review: the person who runs it cannot approve it). Optionally set a review due time in hours; runs then show waiting, due or overdue. Choose what happens when a live step reports evidence gaps: continue and list them, or stop for a person to resolve. Process view and BPMN export shows the flow with its review decision, exception path and rework loop, and downloads a BPMN 2.0 file for process tools; it describes the process and is not an executable engine definition.

4. Import terminology from Ontology Studio

In Ontology Studio, approve your terms, then either publish a release (Use & export → Publish a release → Download release) or export Knowledge JSON with drafts excluded. Import the file under Business vocabulary and inspect the definitions before saving. The importer accepts up to 100 approved, active concepts and keeps their identifiers, definitions, term kinds, aliases, scoped meanings and approved relationships; a release also pins its version and fingerprint, shown next to the vocabulary. It discards persona instructions and raw corpus content. Importing a new version shows its impact first: added, removed and redefined terms, and the mapped fields that depend on them. Approval labels are user-entered assertions, not verified reviewer authority.

When you import a release, the builder also lists your other saved workflows that pin an earlier version of the same vocabulary, with the terms whose meaning changed or was removed and the number of test cases to rerun. Only the latest version of each workflow is listed.

Example: define “eligible work” in Ontology Studio, then use that vocabulary when reviewing engagement evidence. A definition does not establish that a particular work entry is eligible. The source documents must supply that evidence.

Saving creates an immutable workflow version containing the selected vocabulary snapshot. Later ontology changes require another import and workflow version. No cross-product account or live synchronization is assumed.

5. Reuse a knowledge collection or ontology publication

The Knowledge & ontology library stores reusable, immutable revisions. Create a publication, give it a name, and upload either an approved-only Ontology Studio JSON export or up to five TXT, Markdown or CSV documents. Publishing stores content but makes no AI request. Private is the default. Workspace sharing requires an explicit acknowledgement and gives every active member access to that revision, including members invited later. Publish only material you are authorized to share with that audience.

Choose Use ontology while editing a workflow, then save a new workflow version. Choose Add collection to run after saving a workflow. Inspect each document before running. The library revision and workflow version are separate: an ontology may call itself version 1.0.0 while the library has revision 2. Neither changes automatically in a saved workflow or run. This is a library of imported publications, not a live connection to your Ontology Studio account or an independent verification of its approvals.

Publishers can revise their own material or withdraw a revision from new selection. A new private revision does not make older shared revisions private; withdraw those separately. Withdrawal cannot recall copies already pinned to workflows, runs, exports or backups. A selected collection is checked again when starting a new run. Existing runs retain their snapshot. Previously saved ontology definitions remain part of their workflow.

6. Connect business concepts to the work

After selecting vocabulary, add a concept mapping and choose its direction. A required input mapping needs at least one document assigned to that concept under Inspect document and assign input concepts. This validates the assignment, not whether the document actually contains sufficient evidence. An output mapping requests a text value from the source. Each returned value must appear inside its exact quoted passage; required fields must be present. These are text-extraction checks, not typed financial calculations, relationship reasoning, eligibility decisions or authorization rules.

Example: publish the bundled engagement vocabulary and engagement evidence as separate synthetic library items. Open the billing template at /builder?template=billing, select the vocabulary, and add a required input mapping called Eligibility evidence using Eligible work. Optionally add an output mapping called Missing terms using Evidence gap. Save, add the evidence collection, and assign its document to Eligible work. A live draft may quote that the engagement terms were not included; it must not infer that billing is approved. Inspect every result. Reference mode does not extract mapped fields, so leave output fields optional for a reference walkthrough. A required output cannot be accepted without a supported extracted value. Missing evidence is a reason to clarify or reject, not to invent a value or weaken a necessary business requirement.

7. Supply demonstration documents

Save the workflow, then upload a PDF or DOCX up to 4 MB (text is extracted with page or paragraph markers) or UTF-8 TXT, Markdown, or CSV up to 400 KB. Up to ten documents are allowed, each with no more than 100,000 characters and 300,000 combined, about 75,000 tokens per step. CSV is treated as text, not a spreadsheet calculation. Original files are not retained separately; extracted text is stored in the run. Websites, audio and enterprise repositories are not ingested.

Scanned PDFs. If a PDF has no text layer, the builder offers Read scanned pages on this device. The pages are read in your browser with an open-source OCR engine served by Workflow Studio itself: the file is not uploaded or sent to any AI provider while it is read. The first time, the reader downloads about 9 MB. Up to 40 pages are read; pages that were hard to read are counted in the message. Check the extracted text before running, because quotations must match it exactly, and the patient-data check runs on it like any other document.

8. Run and inspect

Reference extraction copies sample passages and makes no AI request. Live drafting appears only when the owner enables it and the workspace has an AI provider connection (see Limits and responsibilities). Before each live run, acknowledge that the supplied documents and vocabulary will be sent to the workspace's configured provider. Use demonstration material you are authorized to process.

Patient and personal data check. Before anything is sent to a provider, the builder checks the documents and test cases for text that looks like a Social Security number, a medical record or patient number, a date of birth, a Medicare beneficiary identifier or a payment card number. If it finds one, nothing is sent and the step reports which document and which kind, never the value. Synthetic identifiers such as member ID M-5520 or claim CL-9055 pass. The check is a safety net, not a guarantee: do not upload real patient or customer records.

After the library migration, new live runs retain the acknowledgement's provider, notice version, requesting user and time alongside the frozen evidence. It appears in the run and its export. Older runs do not receive an invented acknowledgement. This is a user-authored start record, not independent confirmation that processing was legally authorized.

Create a saved run, then choose Execute next step. Each click executes one step. Live steps run in the background for up to 14 minutes: the page checks every few seconds and you can leave and come back. A step is never retried automatically; if it cannot finish, the run shows why (for example model_timeout) and becomes uncertain. Check an uncertain run before starting a new one, which may incur another provider charge.

Every finding must cite an exact quotation from a supplied source. Each finding and mapped field shows its quotation highlighted inside the surrounding source text. This validates the quotation's presence, not the truth of the source or the correctness of the interpretation. Inspect the summary, findings, and gaps yourself.

Test cases. Save synthetic documents with expected results (mentions, mentions one of, does not mention) and rerun them after changing steps, rules or the ontology. Run test starts a run from the saved documents and shows each check. The claims template includes two cases from the book: the claimant's own prior claims, and the procedure in force on the policy date. The healthcare templates include three cases each. Reference runs only quote the documents, so use a live run for a meaningful result. When test cases are required, other runs start only after every case passes on the saved version.

Comparing providers. If an administrator has added a comparison provider, choose Run on comparison provider to run the same test cases there. The comparison shows, for each side, the provider and model, how many cases passed and the tokens used per case. Comparison runs are kept apart: they never satisfy the release gate, which reads only the main provider.

Scorecard. Each saved workflow shows a scorecard built from its live runs on real documents, across every version: cases run and reviewed, the share accepted, corrected by the reviewer and returned for rework, the median time from start to decision, and tokens per accepted case. Test-case and comparison runs are left out. Add your own estimates (minutes per case today, reviewer minutes with the workflow, and the main provider's prices) to see AI cost per accepted case and estimated hours saved. The estimates stay in your browser. Start-to-decision time includes waiting, and hours saved is capacity, not cash.

9. Record a review and export

Provide a reason for accepting or rejecting the completed packet. With a named reviewer, only that reviewer can decide, and the record shows a separate review; otherwise it is your own disposition. Neither is independent finance approval. A named reviewer receives an email with a link when the owner has enabled alerts, and the run also appears under Waiting for your review, with waiting, due and overdue labels. Before deciding, the reviewer can flag a finding as incorrect or correct a mapped value, with a reason; the original draft is kept and the correction is shown beside it. The reviewer can also return the run to an earlier step for rework: later results move to the run history and the step runs again. The person who owns the workflow can then choose Add as a test case beside each correction: a corrected field becomes a check that the right value appears and the wrong one does not, and a flagged finding becomes a “does not mention” check to narrow to the wrong statement. Save a new version to keep it. Runs whose documents exceed test-case limits (five documents of 16,000 characters) can't be converted. Export the evidence JSON for an offline record of the saved workflow, documents, results, and review. Provider token counts appear when available; monetary cost is not inferred.

10. Reuse, revise, and delete

Open any saved version to rerun it with new documents. Edits create another version; earlier runs keep their original version. Delete a run to remove its stored documents and results from the active application database. This does not delete exports, provider records, or database backups. Workflow definitions remain until an authorized maintenance deletion; avoid including personal or confidential information in definitions.

Connecting your systems (planned)

Workflow Studio does not connect to core systems, CRM or AI agents yet. The planned integration surface is a governed REST API with signed webhooks, for your integration team to connect through your own gateway, and an MCP server over the same operations for approved AI agents. Results leave only after a person has reviewed them. Read the integration design or ask about integration as part of an engagement.

Limits and responsibilities

The library retains up to 200 revisions per workspace/publisher, including withdrawn revisions. The list shows the latest 200 accessible revisions. Library collections keep the five-document, 16,000-character per-document limits; a run combining collections and uploads follows the run limits in section 7. No automatic retrieval, vector index, file malware scanner or enterprise connector is added by this library. Use authorized synthetic text for the pilot.

AI provider for this workspace. A workspace administrator adds the provider key under AI provider for this workspace: choose the provider (Anthropic, OpenAI, Azure OpenAI, Google Gemini, Amazon Bedrock, Mistral, Cohere, xAI and others), paste the key, list the models it can use, pick one and choose Check and save. A main provider is used for every run; an optional comparison provider is used only for comparison test runs. The key is checked, then stored encrypted for this workspace only. Members see the provider, the model and the last four characters, never the key, and only administrators can replace or remove it. Providers other than Anthropic are marked not yet tested until a live run has been verified.

Workspace administrators can choose Check provider connection. This checks access to the configured model without submitting documents or generating text. A successful check does not establish billing readiness or successful workflow execution. Complete an intentional synthetic live run, inspect its quotations and token usage, and export the evidence before declaring live acceptance complete. Never retry an uncertain run automatically.

The beta permits 20 new runs per user per day, 100 retained runs per user, and 200 workflow versions per workspace/user. These are usage limits, not a monetary budget. Users control the material they submit and remain responsible for decisions. Administrators manage access, provider configuration, retention procedures, and incident response. Do not upload regulated or confidential client data until the applicable processing terms and controls have been agreed.

Saved records are user-authored and are not signed execution attestations. External writes, scheduled background operation, automatic corpus retrieval, independent reviewer queues, and regulatory approval are outside this builder increment. Existing billing/QuickBooks controls remain separate.

Open the builder →