Workflow Intelligence

Overview

Workflow Intelligence is a concept extension of the Workflow Builder I designed for HCL BigFix.

The original product helped IT administrators turn repetitive endpoint-management tasks into reusable no-code workflows. It introduced a visual canvas, configurable nodes, branching, failure handling, deployments and a JSON editor for power users.

After completing that project, I wanted to explore the next problem. A visual builder reduces manual execution, but creating a dependable workflow still requires considerable product and technical knowledge.

Administrators must know what should be automated, which actions to use, how to configure them, where a workflow could fail and what should happen when it does.

Workflow Intelligence explores how AI could support that complete journey without hiding the workflow logic or taking control away from the administrator.

Relationship to the original project

The original canvas remains. The extension adds support before, during and after workflow creation.

My objective

Explore how AI could help administrators discover, create, validate and maintain workflows while preserving the reliability, accessibility and control expected from an enterprise product.

The remaining problem

The first version made repeated workflows easier to execute, but creating them still required administrators to translate an operational goal into system logic, know the available actions and conditions, configure technical fields, define branches and failure behaviour, identify missing permissions and invalid combinations, understand failures, review execution data and decide what should improve over time.

An empty canvas is flexible, but it assumes that the user already knows what the final workflow should look like.

The challenge

AI can reduce setup effort, but endpoint-management workflows may affect hundreds or thousands of devices. A wrong suggestion can create more than a usability problem.

The redesign therefore had to help users move faster, keep the system’s decisions understandable and increase human control as the consequence of an action increases. The manual Workflow Builder also needed to remain fully usable when AI was unavailable or when an administrator preferred to build without it.

Goals

Project status

This is an independent concept extension based on the constraints and learnings from the original Workflow Builder project. It was not shipped or tested as part of the HCL BigFix product.

The outcomes shown later in the case study are measurement goals, not production results.

Looking at the automation market

I reviewed how established automation products were adding AI to their workflow experiences.

• Zapier Copilot
• n8n
• n8n human-in-the-loop
• n8n evaluations
• Power Automate Copilot
• Power Automate Process Mining
• Workato Recipe Copilot

What I took from the benchmark

The benchmark changed the scope of the redesign. AI could support more than the first draft of a workflow. It could help across four connected areas: build, run, repair and improve.

The opportunity

How might we help administrators discover, build, validate and improve automations with AI while keeping the workflow understandable and the administrator in control?

Design principles

From Workflow Builder to Workflow Intelligence

I added AI support around the original workflow canvas rather than replacing it.

Core journey

The journey does not end when AI generates a canvas. A dependable enterprise experience must help users understand, validate and maintain what was created.

Discover and build with AI

The first journey starts with evidence, then moves into a structured creation experience.

Choosing how to start

I kept more than one creation path because speed and control are not the same for every user or workflow. Build with AI suits a clear, low-risk task. Guide me is more appropriate when the workflow is unfamiliar or sensitive. Experienced users can continue with the blank canvas or JSON editor.

Describing the outcome

The AI entry point is a way to start building, not a generic chatbot dashboard.

Clarifying missing information

An incomplete prompt should not quietly become an automation. When information affects the result, the system asks before it builds.

Generated workflow

AI generates normal Workflow Builder nodes. The result is not trapped inside a conversation. Every node can be inspected, moved, edited or replaced using the existing canvas.

Selective editing

A user should not need to regenerate the entire workflow to correct one decision. AI changes should remain local, visible and reversible.

Checkpoints

AI makes experimentation faster, which also makes recovery more important. Checkpoints let users explore a different structure without losing a dependable version.

More useful than one confidence score

I initially considered showing one confidence percentage for the generated workflow. I dropped that direction because a single number combines several different questions.

A workflow can be technically valid while relying on weak evidence. AI can be confident about a recommendation that still carries a high business consequence. I separated confidence into four visible states.

Explaining a decision

The explanation focuses on information the administrator can verify. It does not expose raw chain of thought or use generated reasoning as evidence.

Adding an AI Agent

Some endpoint events cannot be handled through fixed conditions alone. An alert may contain unstructured logs, incomplete signals and patterns from previous incidents.

For these cases, I introduced an AI Agent node that can interpret information and produce a bounded, structured recommendation. The administrator defines its objective, inputs, tools, output and authority.

Autonomy levels

Autonomy belongs to an individual action, not the entire workflow. A workflow may automatically collect evidence while still requiring approval before isolating a device.

Human approval

High-consequence actions remain behind explicit approval, with the expected effect, evidence, missing information and audit context visible together.

Test before activation

Traditional validation checks whether the workflow can run. AI evaluation also checks whether output remains useful and consistent across different inputs.

Previewing impact

The preview helps administrators understand the likely effect before the workflow reaches real devices.

Diagnose and repair

The assistant diagnoses the failure, shows the evidence and proposes a repair. It does not silently modify a live workflow.

Repair comparison

Repairs are compared in context, tested and saved as a draft before they reach a live workflow.

Learning from workflow history

The system should help administrators recognise patterns, but it should not convert every correlation into an automatic change. Insights remain recommendations until someone reviews the underlying cases.

Designing the way back

The manual builder is the fallback for every AI state. An unavailable model should never prevent an administrator from accessing or editing the workflow.

AI cannot become the only interface

The original Workflow Builder treated keyboard interaction as a primary path rather than a fallback. The AI redesign continues that approach.

A conversational input can make creation faster, but it cannot replace structured controls, predictable focus and an accessible representation of the workflow.

AI status is never communicated through colour alone. Generated content is announced without stealing focus. Users can skip to the first issue, cancel generation or testing, and use structured controls alongside natural-language input.

What I would test next

Because this is a concept extension, I would validate the riskiest assumptions before increasing its scope.

The first question is not whether users like the AI interface. It is whether they can understand, correct and safely approve what it creates.

Expected product impact

The concept would be successful if it reduced the effort required to create and maintain workflows without increasing unsafe or misunderstood automation.

I would expect it to improve time to first valid workflow, discoverability of automation opportunities, completion of technical configurations, recovery from failed executions, understanding of AI-generated decisions, reuse of approved workflow patterns and accessibility for users who find a visual canvas difficult to operate.

These are hypotheses for future validation, not measured production outcomes.

From building workflows to understanding them

The original Workflow Builder focused on making repeated endpoint operations easier to automate.

This redesign pushed the problem further. It explored how administrators could discover what to automate, describe the outcome they want, test the resulting workflow and understand what happens when AI becomes part of the execution.

The most important design decision was not adding a prompt to the canvas. It was defining the boundary between an AI suggestion, a system-verified configuration and a human-approved action.

I kept the original action library, canvas, keyboard interaction and JSON editor because dependable manual control still matters. AI adds another way into the product, but it should never become the only way through it.

The first release made workflow automation usable. This concept explores how it could become easier to discover, build and maintain without making its decisions harder to understand.

This project is an independent design exploration and was not shipped as part of HCL BigFix.