MANThLE

GlossaryUpdated 2 min read

Intelligent workflow automation

Intelligent workflow automation is a workflow in which software makes some of the decisions a person used to make, such as classifying a request, choosing the next step or drafting a reply, inside limits the business sets. A classic workflow routes work between people. An intelligent workflow decides part of the work itself and escalates the rest.

Workflow software has moved cases between people for decades. A ticket arrives, a rule assigns it to a queue, a person reads it and decides what happens. Intelligent workflow automation puts a model at some of those decision points. The workflow reads the ticket, decides what it is and what should happen, and either does it or hands it to a person with a recommendation.

What makes a workflow intelligent?

The test is whether the workflow makes a decision that depends on the content of the case. Three kinds of decision come up again and again:

  1. Classification. What is this request? A refund, a dispute, a change of plan, a complaint that needs a supervisor.
  2. Next step. Given the account, the policy and the history, which action applies? Approve, ask for a document, escalate.
  3. Generation. Draft the reply, the case note or the adjustment memo for a person to send or for the workflow to send itself.

Routing by a fixed field ("all invoices over a threshold go to finance") is ordinary workflow automation, however modern the tool.

How does it run in an operation?

Consider a parcel carrier whose drivers report failed deliveries in free-text notes and photos. An intelligent workflow reads each report, matches it to the shipment and the customer's delivery terms, and classifies the problem: a redelivery the terms allow, a return to the depot, a damage claim. The clear cases are resolved by rule. The ones where the model is unsure, or where the parcel's value crosses a limit, go to a customer service specialist with the shipment record, the driver's note and the photo side by side and a suggested resolution.

The specialist spends the day on the cases that need a specialist. The workflow keeps a record of every decision, so a manager can check a sample each week and tune the thresholds.

What guardrails make a deciding workflow safe?

A workflow that decides needs more control than one that routes. The working set:

  • Confidence thresholds that send low-confidence decisions to a person.
  • Action limits on what the workflow may change on its own, by type and by value.
  • Approvals for actions that are hard to undo.
  • An audit trail of the input, the decision, the action and the outcome.
  • A stop switch that returns all cases to people.
  • Sampling of automated decisions, reviewed by someone who knows the work.

The NIST AI Risk Management Framework (NIST AI 100-1, January 2023), which NIST intends for voluntary use, organizes AI risk management into four functions: Govern, Map, Measure and Manage. It is a reasonable checklist to hold an intelligent workflow against. The AI guardrails entry goes further into limits and approvals.

How does it relate to intelligent automation and process redesign?

An intelligent workflow is usually the orchestration layer inside intelligent automation, the part that carries a case from step to step. It works best after a process redesign: decide which decisions the model should own and which stay with people, then build the workflow around that split. Teams that want one engineered and operated for a specific process can look at AI workflow automation.

Should every step get a model?

Only the steps where the decision depends on reading unstructured content. A tool that makes it easy to add a model everywhere invites overuse, and each model decision adds a failure mode and a monitoring burden. Keep plain rules wherever a rule will do.

Questions

An intelligent workflow is a sequence of steps where some decision points are handled by AI models, with rules and a person covering the decisions the models should not make alone.

Regular workflow automation moves a case to the next person or system according to fixed rules. Intelligent workflow automation reads the case content and makes a judgment about it before acting.

Confidence thresholds, limits on what actions the workflow may take, approval steps for high-risk actions, a full audit log and a way to stop the workflow quickly are the minimum set.

Routine cases can run without a person. Unusual or high-risk cases should always reach someone with the context to decide them.

Start with one workflow where most cases follow clear patterns, a minority need judgment and delays have a visible cost to customers or revenue.

Sources

  1. NIST, AI Risk Management Framework
  2. NIST AI 100-1, Artificial Intelligence Risk Management Framework (AI RMF 1.0), January 2023
SolutionsAI workflow automation for one critical processAn AI workflow engineered around one critical process: models on routine decisions, a person on the exceptions, hardened for production and run with your team.GlossaryIntelligent automation (intelligent process automation)Intelligent process automation joins rule-based bots, AI that reads and decides, and a person on the exceptions. How it works and where vendor definitions fail.GlossaryProcess redesignProcess redesign changes the steps, handoffs and roles of a process. How it differs from reengineering and automation, and how to design for people plus models.GlossaryAI guardrailsGuardrails in operations go past content filters: permissions, action limits, approval gates and audit logs that keep an AI agent within what a business allows.GlossaryHuman in the loop (HITL)Human in the loop as an operating design: which AI decisions go to a person, which person gets them, what context they see, how fast they act, what it teaches.