Can Employee Action Logs Be Turned into Meaningful Workflows?

Hello. This is Keito Inoshita from Affectosphere Group.

“We want to understand what our employees are actually doing” — it’s one of the most common frustrations in organizations.

In practice, what teams can see is system-level logs and form data. Who did what with what intent, and how exactly they did it, tends to stay invisible.

Browser histories, learning platform click data, internal AI usage logs — this kind of data is probably already accumulating in your organization. The problem is that these are low-level action records. They don’t map cleanly onto meaningful units of work.

A study published on arXiv in June 2026 (Gaurav Verma, Scott Counts, arXiv:2606.14654) proposes a framework called WorkflowView that uses LLMs to close exactly this gap.

---

Three takeaways for today

1. WorkflowView is a generalizable framework that uses LLMs to abstract low-level behavioral logs from different domains into interpretable, high-level workflows.
2. It achieved semantic similarity of 0.91 for reconstructing task descriptions from browser logs, and weighted F1 of 0.90 for predicting learner dropout from educational platform interactions.
3. The approach has strong potential for business process analysis, internal AI usage visualization, and knowledge management applications.

---

① What WorkflowView actually does

The core idea behind WorkflowView is simple.

When people work, they’re always executing purposeful sequences of actions. But what systems record are low-level events: URL clicks, button presses, text input. There’s a gap between those raw events and the meaningful activity they represent.

WorkflowView bridges that gap by abstracting sequences of low-level actions into coherent, interpretable units of activity — using an LLM as the interpretation engine.

The research validates this approach across three different domains.

First, browser activity logs. Given a sequence of URL accesses with timestamps, the system reconstructs a natural-language description of the task the person was working on.

Second, educational platform interaction logs. Given a learner’s behavioral history on a learning platform, the system predicts whether that learner is at risk of dropping out — with strong performance even in low-data settings.

Third, AI feature usage in Microsoft Word. Given a record of where and how a user engaged with AI writing assistance, the system visualizes document structure and the editing process in a privacy-respecting way.

---

② The numbers

The results reported in this arXiv study are notable.

For reconstructing task descriptions from browser logs, semantic similarity reached 0.91. That’s close to “essentially the same content was recovered.”

For predicting learner dropout from educational platform behavior, weighted F1 was 0.90. Importantly, this was achieved in a few-shot setting — implying the system doesn’t require large volumes of labeled data to function.

The Word AI usage visualization is harder to compare with a single metric, but the study demonstrates that document structure changes can be tracked in a privacy-preserving way.

The authors conclude that “LLM-based abstraction is a robust and efficient means of transforming low-level behavioral data into high-level, interpretable, and actionable insights.” Having this demonstrated across three distinct domains is what makes the claim compelling.

---

③ Where to start: practical applications

From a business implementation standpoint, internal AI tool usage visualization is probably the most accessible entry point.

Many organizations have deployed tools like Microsoft Copilot or ChatGPT Enterprise. But “who’s using it, for what, and how” tends to stay opaque. Usage logs exist, but they don’t tell you much on their own.

A WorkflowView-style approach could interpret AI tool interaction logs at the activity level: “This employee primarily uses AI for drafting document templates.” “This employee’s main use case is summarizing meeting notes.” That’s actionable insight that raw log data doesn’t provide.

Here’s one concrete scenario. A DX team building an internal AI adoption dashboard could pull Copilot for Microsoft 365 API logs periodically, apply WorkflowView-style abstraction to classify activity by business use case, and track KPIs like AI adoption rate, distribution of use case types, and penetration by department. That’s a more meaningful picture than login rate alone.

HR applications follow naturally. Learning platform interaction logs from onboarding could be used to predict early dropout risk for new hires — in few-shot settings, which matters when you don’t have years of labeled retention data to train on. The weighted F1 of 0.90 the paper reports maps directly to this kind of use case.

For knowledge management and process standardization, browser logs and internal tool activity from high-performing employees could be used to extract implicit workflows — turning undocumented expertise into something that can be standardized and shared.

---

④ Things to think through before you start

A few considerations worth working through before moving to implementation.

Privacy is the most important. Behavioral log collection and analysis requires careful thinking about employee privacy and applicable labor law. The paper’s Word usage analysis explicitly notes a privacy-preserving design — that framing matters, and it should inform how any real-world implementation is designed from the start.

Abstraction accuracy will vary by domain and task type. LLM-based abstraction is powerful, but performance can degrade when the behavioral logs are noisy, sparse, or outside the distribution the approach was designed for. A small pilot to validate accuracy in your specific context before broader rollout is the pragmatic approach.

The framing matters for adoption. Systems that analyze behavioral logs can feel like surveillance if positioned that way. Designing the use of WorkflowView-style analysis around employee benefit — helping people reflect on their own workflows, identify bottlenecks, surface best practices — makes it more likely to be used and trusted rather than resisted.

---

From “logs accumulating” to “meaning readable”

Most organizations are already sitting on behavioral data. The bottleneck isn’t collection — it’s interpretation.

What WorkflowView demonstrates is that LLMs can bridge low-level action sequences to interpretable, high-level workflows across meaningfully different domains. That’s the capability gap that’s been missing.

The most realistic path to implementation is to start narrow: one domain, one use case, a small pilot to validate accuracy. “Visualize internal AI tool usage logs as workflow-level activity summaries” is a well-scoped starting point — concrete enough to measure, contained enough to move quickly.

That’s it for today!

---

Reference

1. Gaurav Verma, Scott Counts (2026). Abstracting Cross-Domain Action Sequences into Interpretable Workflows. arXiv preprint.

* This article was written in part with AI assistance and may contain inaccuracies.