The most compelling value proposition of B2B LLM applications is workflow automation. In nearly every function, headcount costs are an order of magnitude larger than software spend. Outside of cost savings, AI automation can improve the speed and quality of the work performed. An AI-powered customer service bot means a customer will never have to wait for an agent. An AI-powered security analyst can work around the clock, and won’t get bored after looking at 100 nearly identical alerts.

But Robotic Process Automation (RPA) tools have been around for decades – both general platforms like UiPath and Zapier, as well as function-specific ones like Workday (human resources) and Demisto (security). So what’s changed?

Historically, RPA just hasn’t worked well enough to automate end-to-end workflow. Their “if this, then that” type functionality can tackle some small tasks. But they’re too cumbersome and brittle to handle the vast majority of work that people do on a daily basis.

LLMs change this equation. They enable systems that can actually accomplish human work, instead of just moderately increasing efficiency. This will result in a new age of automation software.

Business workflows require human-like decisions and judgment

While many business workflows follow a consistent high-level process, you can’t blindly follow a sequence of steps robotically. Humans running the process make many decisions and judgments along the way.

For example:

• Someone reading an email that mentions “invoice” must determine if the email is sending an invoice, requesting it, or referencing a previous one.
• An employee entering a purchase order will quickly notice if the quantity is off by a factor of 10 compared to previous orders, and flag it for review.

It’s clear why traditional automation platforms fall short. Enumerating all of the different possibilities would require hundreds or thousands of rules and workflows, with exception handling for every possible edge case. Any change to the business process likewise requires rewriting all of these rules. Many of the judgments require semantic understanding and human-like logic, which isn’t possible with rules-based systems.

Integrations and interfaces are complicated to build

Most valuable business workflows require read and write access to different systems and software. Many involve both structured and unstructured data, with text inputs via email, forms, or chat interfaces.

RPA platforms provide integrations to many other pieces of software. But for most, these interfaces aren’t easy to use. They require a developer to implement. They are limited in functionality by the API of the third-party platforms. Any access to data requires customized queries and transformations. Unstructured data is almost impossible to work with.

When any of these things changes, the whole process breaks.

For example:

• When security operations centers (SOCs) investigate alerts, they need to pull data from a variety of systems – SIEM, NDR, EDR, threat intelligence feeds, etc. Each one of these has its own data structure, querying language, and field parsing.
• Supply chain communications use the Electronic Data Interchange (EDI) format, but it is a weak standard whose structure varies by company and even employee – for example, someone might put information in a “product” field, others in a comment.

It doesn’t matter how simple it is to design a workflow in a GUI: If the pipelines and integrations to power it take months of specialized work to build and maintain, it makes many projects infeasible.

Modern LLM systems solve these problems

LLM systems solve both of the challenges above.

LLMs can make many small judgments and decisions effectively, using semantic understanding plus basic logic/reasoning. They can categorize a message or decide if a question has been answered.

LLMs can also be used for ETL and interface, e.g. extracting information from an unstructured email or writing queries to access different data sources. This is still an emerging capability, but is rapidly improving.

We do not need fully-autonomous AI agents to realize the benefits of LLM-based automation. We expect the majority of applications will still be built with a concrete process description and guardrails designed with expert knowledge.

What the LLM unlocks is the human-like reasoning that allows a system to navigate the branching points, reason, and provide flexibility that is not possible with rules-based systems alone.

Examples where LLMs will transform workflow automation

Security operation centers (SOCs)

SOCs staff teams of analysts and engineers to triage, investigate, and resolve alerts.

Security Orchestration, Automation and Response (SOAR) platforms promised to automate security workflows. But the rules-based playbooks are only good at simple, repetitive tasks like “send a Slack notification when you receive a high-severity alert from X system” or “email Y users daily until they change their password”. They cannot perform the actual investigation, data gathering, and analysis that humans do to determine if incoming alerts are real or false positives.

With LLM-based automation, platforms can replicate analyst workflows like a human, querying data from different systems, interpreting it, and executing actions to mitigate the issue. These systems can respond much more quickly to alerts, work 24/7 without alert fatigue, and allow humans to focus on higher complexity issues.

Supply chains

Supply chains are built on complex, legacy infrastructure. Large enterprise ERP systems like SAP and NetSuite can take months and millions of dollars just to implement.

Because of this, a huge portion of supply chain data is exchanged via EDI and unstructured documents like emails, PDFs, and spreadsheets. These have limited the opportunity to build automation.

LLMs make data extraction and transformation much more robust. With these capabilities, companies can automate end-to-end workflows – not just to save labor costs, but to drive better supply chain operations and provide better service for customers.

—––

If you’re building a company around LLM-based workflow automation, we’d love to hear from you at info@theory.ventures!