Join the Waitlist

Use CasesPharmaceutical

Northwind Therapeutics

Evidence, everywhere it is needed.

A mid-cap therapeutics company running trials and launches across 14 markets, where every piece of content has to survive a regulatory reviewer before anyone outside sees it.

workflows shown here
3
AI agents behind them
4
runs seeded in the tenant
8
human gates on every asset
2

Nothing Northwind publishes reaches a prescriber without passing an AI claims check and two named human sign-offs. Dave is the thing that makes that chain provable rather than promised.

Dave tenant
demo-pharma
Workspace
Northwind Review Hub
Sector
Pharmaceutical, global clinical and commercial operations
Who reads their work
Prescribers, trial investigators, patients and carers, and the regulatory affairs teams who sign off before any of them read a word.
  • Clear
  • Careful
  • Warm
  • Accountable

How does AI fit into pharmaceutical MLR review?

As the first pass, before the people, and never as a decision. In this demonstration tenant every promotional asset goes through an AI label claims check that returns numbered findings against the approved label, each graded minor, major or critical and tied to the label section that would have to support the claim. The agent is told never to approve: findings only. Two named human gates follow, a medical reviewer and then regulatory affairs, and both have to approve before an asset is cleared for use.

Can one AI agent check another agent's work before a person sees it?

Yes, and Northwind's submission workflow is built that way. An Agent Review node sits between the drafting agent and the medical writer: it votes on the draft, and a rejection loops the draft back to be rewritten with no human touched at all. The drafting agent is also required to leave a data-required placeholder wherever a value was not supplied, rather than inventing one.

The workflows, node by node

Not a summary of them. These are the graphs as they are seeded: every step in order, its type, the agent on it, who it is assigned to, and where a rejection sends the run back to.

Guide: Build your first workflow

Promotional Material MLR Review

Medical, legal and regulatory review as one graph. The AI reads the material against the approved label first, so the two humans further down spend their time on judgement rather than on catching the obvious.

6 nodes

  1. Start

    Start

  2. User Interaction

    Submit Promotional Asset

    Submitter

    Asset title, the product it covers, and a summary. All three required.

  3. Agent Interaction

    AI Label Claims Check

    Label Claims Checker

    Numbered findings against the approved label, each with a severity of minor, major or critical and the label section that would have to support it.

  4. Human Review

    Medical Review

    Medical reviewer

    First gate. A rejection returns the whole asset to its author, not to the agent.

    Reject returns to Submit Promotional Asset

  5. Human Review

    Regulatory Sign-off

    Regulatory affairs

    Second gate, a different person and a different role. Both must approve.

    Reject returns to Submit Promotional Asset

  6. End

    Approved for Use

Regulatory Submission Section Prep

A submission section drafted from briefing notes, then checked by a second agent before a medical writer ever opens it. Where data is missing the drafter is required to leave a marker instead of filling the gap.

6 nodes

  1. Start

    Start

  2. User Interaction

    Provide Briefing Notes

    Submitter

    Which section, and the notes to write it from.

  3. Agent Interaction

    AI Draft Section

    Regulatory Section Drafter

    Formal CTD style, numbered headings, no marketing language, emitted as an HTML artifact.

  4. Agent Review

    AI Consistency Check

    Regulatory Section Drafter

    An agent voting on another agent's output. Rejecting loops the draft back to be rewritten, with no human touched yet.

    Reject returns to AI Draft Section

  5. Human Review

    Writer Finalization

    Medical writer

    Reject returns to AI Draft Section

  6. End

    Section Ready

Adverse Event Literature Screening

A batch of abstracts screened for safety signals. The agent is instructed to be conservative, so an uncertain case becomes a human review rather than a closed file.

6 nodes

  1. Start

    Start

  2. User Interaction

    Set Screening Scope

    Submitter

    Which product.

  3. Agent Interaction

    AI Abstract Screening

    Literature Screening Agent

    For each abstract: relevance, suspected adverse-event terms, seriousness, and whether follow-up is warranted. Writes a yes or no on signal detected.

  4. Routing

    Signal Triage

    Suspected signals require pharmacovigilance review.

    • Signal DetectedPharmacovigilance Review
    • No SignalScreening Complete

  5. Human Review

    Pharmacovigilance Review

    Medical reviewer

    Reject returns to AI Abstract Screening

  6. End

    Screening Complete

The agents, and what each one is forbidden to do

An agent in Dave is a configured worker, not a chat box: a written prompt, a pinned provider and model, and its own context. The line that matters most in each of these prompts is the constraint.

Guide: Create an agent

Label Claims Checker

Compares promotional claims against approved label language and flags unsupported claims.

GuardrailNever approve. Findings only, each tied to the label section that would have to support the claim.

Regulatory Section Drafter

Drafts regulatory submission sections from briefing notes in agency-appropriate style.

GuardrailWhere data is not provided it must insert a data-required placeholder rather than invent a value.

Literature Screening Agent

Screens literature abstracts for potential adverse event signals requiring pharmacovigilance review.

GuardrailBe conservative. Uncertain cases are flagged for human review, not closed.

Expense Policy Checker

Validates expense submissions against company policy: limits, categories, receipts and justification quality.

GuardrailAssess against policy. It does not approve or reject; the routing rule and the manager do that.

Who is in the workspace

Five seats, and the Dave role each one carries. Role decides what a person can reach; the workflow decides what lands in their inbox.

  • Director, Regulatory AffairsAdmin
  • Process Excellence LeadCurate
  • Medical WriterUse
  • Medical Reviewer (MD)Use
  • Quality Assurance AnalystReporting

What is in flight

The runs seeded into this tenant, and for each one still moving, the step it is waiting on.

Guide: Watch a run and read its timeline
  • RunningMLR: Claridem XR Congress Booth PanelsWaiting on Medical Review
  • RunningJuly Literature Screen: NorvexaWaiting on Pharmacovigilance Review
  • CompletedMLR: Norvexa HCP Leave-Behind (v3)Approved at both gates
  • CompletedCTD 2.5 Clinical Overview: Norvexa 25mg sNDAFinalized
  • CompletedJune Literature Screen: Claridem XRClosed, no signal

What one of these agents actually produced

An excerpt from a real run in this tenant. Read what it does and does not do: it grades every finding and hands the decision to the next node.

claims-check-claridem-booth.txtAI Label Claims Check, by the Label Claims Checker
LABEL CLAIMS CHECK: Claridem XR Congress Booth Panels 1. [CRITICAL] Panel 2 headline implies disease modification; label supports symptomatic treatment only. Must be reworded.2. [MAJOR] Panel 3 efficacy graphic uses post-hoc subgroup data without the required limitation statement.3. [MINOR] Panel 4 safety panel font size below fair-balance readability guideline. Summary: 1 critical, 1 major, 1 minor. Routed to Medical Review with findings attached.

The MLR pipeline on the canvas

The same workflow diagrammed above, as it is actually built in the Dave workflow designer.

The Dave workflow designer with a promotional material review open on the canvas. A node palette on the left lists Start, End, User Interaction, Human Review, Agent Interaction, Agent Review, Routing, API Call, Information and Safety. On the canvas a Start node connects to Submit Promotional Asset, then to an AI Label Claims Check, then on to Medical Review and Regulatory Sign-off, each review node offering an approve and a reject path.
Northwind's MLR review in the designer: submit, AI label claims check, medical review, regulatory sign-off.

The agents, and the work waiting on people

Four configured agents on the left. On the right, the review steps in flight with the person who owns each one.

The Agents screen listing four AI agents: Expense Policy Checker, Label Claims Checker, Regulatory Section Drafter and Literature Screening Agent. Each row shows the provider and model that agent is pinned to, its status, its tags and how many versions it has. A second tab offers a Prompt Library.
Northwind's four agents, each pinned to its own provider and model.
A task inbox listing work assigned to people. Each row names the task, says whether it is a human review or a user interaction, and shows its status, the workflow it belongs to and the person it is assigned to. Some rows are assigned and some are completed.
The review steps waiting, with the people who own them.

Every run, inspectable node by node

A single instance opened up: its status, its step counts, and the execution timeline underneath.

The detail view of a running workflow instance. Tiles show its status, when it was created and started, and how long it has run, followed by counts of all, actionable, active, pending and complete steps. Below that an execution timeline and a list of node executions, with Start complete, Submit Expense assigned, and the AI Policy Check, Approval Triage, Manager Approval and Expense Processed steps still pending.
One run, node by node: what finished, what is assigned, what has not started.

An agent reviewing another agent's work is a real step, not a trick. The consistency check on the submission workflow votes approve or reject and loops the draft back on a reject, so a whole class of rewrites never reaches a medical writer at all.

Every fact on this page is read from this tenant's seed data (packages/backend/prisma/demo-packs/pharma.ts, with templates/expense-report.ts, demo-packs as of 2026-08-29) or from the firm's own brand pack. Nothing here is a customer story.

The other two

Meridian Capital GroupFinancial servicesHartwell & Associates LLPLegal

Frequently asked questions

Is Northwind Therapeutics a real company?

No. Northwind Therapeutics is a fictional company we invented as a demonstration tenant. It is not a customer of HelloDave.ai, it has no website, and this page is not a case study. The workflows, agents, roles, runs and screenshots on it come from the seeded demo-pharma Dave tenant.

Does a person still have to approve every asset?

On the MLR workflow, yes, and there is no path around it: a medical reviewer approves first, then regulatory affairs, and only then is the asset cleared for use. A rejection at either gate returns the whole asset to its author rather than to the agent. The literature screening workflow is different by design: a routing rule sends only a suspected safety signal to pharmacovigilance review.

What stops the AI from closing a case it is unsure about?

The instruction it runs under. The Literature Screening Agent is told to be conservative and to flag uncertain cases for human review rather than close them, so an unclear abstract becomes a pharmacovigilance review instead of a filed result.

Now build yours.

Same canvas, same node types, your process. Put the first version on it and watch one run end to end.

Join the WaitlistAll three use cases
Pharma Use Case: MLR Review and Regulated Content Workflows | Dave