The tool was broken. The question was wrong.

Seven workflows shipped. 10,523 hours reclaimed annually.

The rollout: 100+ users, no mandate

I launched Perplexity Enterprise Pro at WebMD! Senior managers across health, legal, automotive, and contractors came in first and their teams followed, reaching 100+ active users across seven business units — Legal, Editorial, Product, Program Management, Medscape Education, Compliance, and Data Science. I validated HIPAA and SOC2 compliance with the VP of Engineering before a single license went out.

On paper that was the win. In practice the real work started immediately after. In breakout sessions I found that a significant number of users thought Perplexity was interchangeable with ChatGPT and were using it to rewrite emails and generate images. They weren't wrong, but they were using a research engine to do what any chatbot could. We had headcount, not changed behavior.

So I stopped presenting the platform in the abstract and started asking one question: what's slowing you down this week? Repeated across dozens of sessions, that question produced the workflows below.

SSO Tear-Down: the autonomous M&A due-diligence agent

A prototype, and the most technically involved thing I built. The product management team was evaluating acquisition targets' SSO architecture, data migration readiness, compliance posture, and transactional email systems. Historically that was six weeks to six months of manual research per target.

I built an autonomous agent pipeline on seven persistent skill files inside a Perplexity Space running on computer use. It scrapes public endpoints, enumerates subdomains, inspects server configurations, and scores 13 technical dimensions against a set of auto-fail triggers. A single run takes about 3 hours. I demoed it once, against an active acquisition target. The product manager who owned the evaluation judged the findings material. It hasn't been run since, so treat this as a prototype that worked rather than a workflow in use.

Interactive Quiz Design

Writers on the Medscape Interactives team were spending 4 to 16 hours per quiz researching recent peer-reviewed guidelines, writing five clinical questions, sourcing each answer, and formatting the output. I built three interconnected persistent skill files for Perplexity's computer use feature: one governing a structured nine-phase flow from research through post-session capture, a second tracking factual error severities and style preferences as a living feedback database, and a third defining overrides for pharmacotherapy-specific content. The system reached roughly a 90% completion rate, two quizzes passed editorial validation and went live, and the organization retired its external quiz vendor.

Book Index Citation

Editors on Nolo's legal editorial team were manually indexing 500-to-600-page legal textbooks, compiling 400 to 600 main entries and 800 to 1,200 subentries per volume across multiple days per book. I built a multi-agent workflow in a dedicated Perplexity Space that splits manuscripts into alphabetical chunks to avoid context-window overflow, then processes each chunk against uploaded style guides and term reference files. One editor pioneered an edition-by-edition technique that feeds the previous index in and tracks only the page shifts in the new manuscript. A human verification layer closes it out: editors spot-check 30 to 40 entries for page alignment and audit for encoding errors.

Fact-Check Tool for Medical Health Content Summarization

Reviewers on the MedAffairs team were spending 12 hours a week verifying AI-generated clinical content against source documents, then hunting exact page citations by hand. I built an accuracy scorer that takes a passage, checks it sentence by sentence against uploaded internal documents, and returns a structured report: verbatim source quotes, dual page numbers for both PDF index and printed page, and a confidence score. I disabled web search entirely so the model could not reach outside sources, and cross-tested the production prompt across three models before finalizing.

Rural Health: fifty states, tracked by hand

A task force at WebMD Ignite was prioritizing all 50 U.S. states for a $50 billion federal rural health transformation program. That meant tracking roughly $10 billion in annual funding, hunting state-level RFPs scattered across inconsistent government websites, scoring each for relevance, and keying the results into Airtable by hand. One stakeholder was tracking state activity with tally marks on paper. Missing a single RFP had direct financial consequences.

Nobody assigned me the project. I volunteered and delivered an MVP in three days. The pipeline calls the Perplexity Search API once per state, filtered to .gov domains, batched into ten calls to stay inside rate limits. Each returns structured JSON: state, RFP title, source URL, publication date, snippet, and a confidence flag for downstream triage. Results key straight into the Airtable the task force was maintaining by hand. For the sites where search alone couldn't reach the document, I piloted Perplexity's computer use feature across ten state portals to navigate to the RFP page and pull the PDF without manual browsing.

Legal Spaces: research outside Lexis

A stakeholder on the legal team tested Perplexity against Lexis, a purpose-built platform with a proprietary case law library, and concluded the tool was broken. They were right that Perplexity lost to Lexis on case law. That just wasn't the question worth asking. It's the most common failure mode I see: test a new tool against the one task where the incumbent is strongest, declare it settled, and walk away.

I didn't defend Perplexity's case law retrieval. So I reframed the question. Lexis should do Lexis work; what research lives outside Lexis and eats a week? That produced Legal Spaces: custom Spaces with source isolation hardcoded to official government domains, structured prompt frameworks per research type, and a task-isolation strategy that splits complex queries into sections. Editors had been spending full weeks hand-building multi-state comparison charts, things like 50-state minimum wage matrices. I ran a four-week onboarding sprint with sequential capability demos.

Health Accreditation Space

A stakeholder was manually reading digital health articles against NCQA compliance guidelines, documenting which sections passed and which failed. I built a Perplexity Space around her evaluation criteria. By the fifth version the Space encoded her reading of the NCQA guidelines, not just the guidelines themselves.

Results

What was hard

Nothing fell over in production. The difficulty was upstream, in getting the agents to behave.

Source restriction was the first problem. A research engine wants to search everything, and most of these workflows are only valid against a narrow set of accredited sources — official government domains for legal research, uploaded internal documents for fact-checking. Hardcoding source isolation into each Space, and disabling web search outright for the fact-check tool, is what made the output usable.

Matching the golden sets was the second. Getting an agent to produce output that consistently matched known-good examples took repeated iteration rather than one good prompt. The Health Accreditation Space took five passes, and each pass surfaced evaluation criteria the stakeholder hadn't thought to say out loud. By the fifth version the Space encoded her reading of the NCQA guidelines rather than the guidelines themselves.

Stack