Why Generative AI matters

in process automation

Generative AI is transforming business automation in ways traditional tools never could. This article explores how large language models (LLMs) enable unprecedented capabilities for processing unstructured documents, holding natural conversations, and handling judgment-laden workflows that older automation simply couldn't touch.

Jump to:

For years, Robotic Process Automation (RPA) tools have promised to eliminate tedious manual tasks by mimicking human interactions with digital systems. While RPA excels at rigid, rule-based processes, it falls dramatically short when facing the messy reality of business operations.

Traditional automation stumbles when it encounters:

  • documents in varied formats—punch lists from every consultant, lien notices from every state, COIs from every broker
  • nuanced rules around decision-making
  • communication that requires context understanding
  • dynamic workflows that change based on circumstances

This is where generative AI shifts the paradigm entirely, as evidenced by our flagship product, Bex. Harmonic Mean's LLM-powered solution doesn't just follow predefined paths—it understands context, interprets meaning, and responds intelligently to novel situations.

The business world runs on unstructured data. Observation reports arrive as long PDFs with annotated photos. Lien notices arrive as scanned forms, dense legal filings, or plain business letters with the key facts buried in a paragraph. Certificates of insurance show up in a fresh layout from every broker. Humans process these subtleties effortlessly; traditional automation breaks down entirely.

Consider the Bex Punch module. Every consultant has their own format—multi-page PDFs with photos and annotations, Word documents with tables, spreadsheets with color-coded rows, sometimes hand-marked drawings. A skilled project engineer can chew through this variety; a rigid template-matching system can't. Bex handles all of it natively, extracts every issue, picks the right trade and assignee, and writes the data to your CMS like Procore, ProjectSight, even just spreadsheets. Bex chooses the proper data fields to keep you organized, and preserves annotations on photos.

Or consider Bex Liens, where the input might be a Texas Notice of Claim with structured fields, a recorded lien affidavit with a county clerk's stamp, or a plain business letter with the same key facts buried in prose. Bex pulls roughly 22 fields out of any of them, fuzzy-matches the project, and classifies the document type so a release closes its matching record automatically and a filed lien triggers a more urgent escalation path.

These are layouts that people handle effortlessly but that have prevented automation prior to LLMs.

Just as revolutionary is the ability of generative AI to engage in meaningful dialog. Bex isn't a form to fill out—it's a colleague to email.

When a Bex Punch reviewer wants to make corrections, they don't open a UI. They reply: "Set the type to Deficiency on all rows. On item 1 the trade should be Waterproofing. Reassign item 4 to Vasquez." Bex understands "row 6", "all items", "the drywall stuff", and "everything in Building B" the way a person would. It re-runs the relevant matching, sends a fresh approval request, and waits.

When a subcontractor's COI is short on coverage, Bex Insurance composes the follow-up email itself: "Your CGL shows $1M / $2M, but the project requires $2M / $4M with our entity named additional insured on a primary and noncontributory basis. Please provide an updated COI or a confirming endorsement." Plain English, specific, no template language to misread. The sub replies, Bex re-assesses, and the loop continues until the sub is compliant—usually without anyone on your side touching it.

When a subcontractor finishes a punch item, they reply with a photo and a sentence. Bex updates the observation in your CMS and confirms back. No portal, no app, no friction.

Bex isn't a one-model wrapper. Every LLM call has its own configuration: which provider (Anthropic, OpenAI, Azure OpenAI), which model, what temperature, what tools. Extraction tasks where image fidelity matters use one provider; transcription tasks use another; structured-output fuzzy matching uses whichever model is currently best at it. You don't pick or tune any of this—your dedicated customer success manager does, continuously, as the model landscape shifts.

For high-stakes extraction (notably Bex Liens), Bex runs multiple LLM passes per document and votes per field. Critical fields require consensus or Bex refuses to commit and escalates. The full consensus log is preserved per record. It's the cleanest answer we've found to "but how do you know the AI got it right?"—we don't trust a single pass either.

And where math has to be deterministic—Bex Risk scoring, for instance—Bex runs actual Python, not LLM-improvised arithmetic. Use the LLM where its strengths apply; use code where reproducibility matters.

Productivity gains are expected from any automation solution. What's different about generative AI automation is that employees actually enjoy the experience.

Reviewers report that working with Bex feels collaborative rather than constraining. The natural language interface eliminates the frustration of learning complex systems, and the intelligent responses create a genuinely helpful experience. Line employees love not having to do the drudgery, while also not having to learn yet another web app and maintain yet another username/password combo.

This enjoyment factor significantly improves adoption rates compared to traditional automation tools, accelerating ROI and organizational transformation. (See real conversations and results.)

Generative AI represents a fundamental shift in what's possible with business automation. By understanding unstructured documents, holding natural conversations, and handling judgment-laden workflows with intelligence—and by knowing when to fall back on deterministic code or multi-pass consensus—these systems deliver capabilities that were unimaginable just a couple of years ago.

Harmonic Mean is at the forefront of this revolution, bringing automation capabilities once considered science fiction into business reality. Is your organization ready to embrace them?