# AI for Process Optimization: A 5-Phase Rollout > Published 2026-11-10 > Source: https://viktor.com/blog/ai-for-process-optimization You implement AI for process optimization in two halves: first use AI to see how a process really runs, from the event data your systems already record, then use AI to remove the delays it finds by running the waiting, chasing and rework steps itself. Most optimization projects fail on the first half. Teams redesign the process they think they have, not the one in the data. AI changes that by reading timestamps from the CRM, ticketing, ERP and chat tools, measuring where cases wait and loop, and predicting which ones will be late. This guide gives a five-phase rollout framework, the metrics to track and the tools for each phase. The framework in brief: measure (pull event logs and compute cycle time, waiting time and rework per step), diagnose (let AI find the bottleneck variants and their causes), predict (flag cases likely to miss their deadline), act (hand the waiting and chasing steps to an AI system with approval rules) and monitor (track the same metrics weekly and catch drift). Measuring comes first because it decides everything after it. ## What AI process optimization is AI process optimization is the use of AI to analyse how business processes actually perform and to improve them, either by recommending changes or by carrying out steps directly. It builds on two older ideas: - **Process mining**, which reconstructs a process from system event logs (case ID, activity, timestamp) to show every path cases really take. - **Process automation**, which runs rule-based steps in software. Our explainer on [what process automation is](https://viktor.com/blog/what-is-process-automation) covers the basics. AI adds the ability to read unstructured data (emails, chat, documents), to explain causes in plain language and to handle steps that need some judgement, such as drafting a reply or deciding who to chase. If you want a general method for redesigning a process with workshops and mapping, see [streamline business processes in five steps](https://viktor.com/blog/streamline-business-processes). This guide is about the data-led, AI-run version. ## Phase 1: Measure the process from its data Pick one process with a clear start and end and enough volume: order-to-cash, procure-to-pay, support ticket resolution, employee onboarding or quote-to-contract. Pull the event data. Each record needs at least: ```json:object { "_key": "k36", "_type": "table", "rows": [ { "_key": "k31", "_type": "tableRow", "cells": [ "Field", "Example" ] }, { "_key": "k32", "_type": "tableRow", "cells": [ "Case ID", "Order 10482, ticket 5531" ] }, { "_key": "k33", "_type": "tableRow", "cells": [ "Activity", "“Quote sent”, “Approved”, “Invoice issued”" ] }, { "_key": "k34", "_type": "tableRow", "cells": [ "Timestamp", "2026-11-03 14:22" ] }, { "_key": "k35", "_type": "tableRow", "cells": [ "Owner or team", "Sales, Finance, Dana" ] } ] } ``` Compute the baseline: median and 90th percentile cycle time, waiting time between steps, rework rate (steps repeated), and the share of cases that follow the main path. **You should see:** a few variants handle most cases, and a long tail of variants causes most of the delay. ## Phase 2: Diagnose the bottlenecks Ask AI to find where time goes and why: - Which step has the longest waiting time before it? - Which variants have rework loops, and what do they have in common (customer segment, product, owner, deal size)? - Where do cases wait for a person rather than for a system? - What do the emails and chat threads around slow cases say? Most bottlenecks turn out to be waiting: for an approval, a missing document, a reply or a handoff between teams. Waiting is exactly what AI can remove. ## Phase 3: Predict late cases With history in place, a simple model can score open cases by their risk of missing the deadline: a quote waiting on legal for three days, an invoice disputed twice, a ticket bounced between teams. Flag them to the owner early, while there is still time to act. ## Phase 4: Act on the waiting steps Now hand the waiting and chasing to AI, with clear rules: ```json:object { "_key": "k70", "_type": "table", "rows": [ { "_key": "k64", "_type": "tableRow", "cells": [ "Bottleneck", "AI action", "Approval" ] }, { "_key": "k65", "_type": "tableRow", "cells": [ "Approval waiting", "Remind the approver with the summary they need; escalate after a set time", "None for reminders" ] }, { "_key": "k66", "_type": "tableRow", "cells": [ "Missing information", "Ask the customer or teammate for the exact item, then update the record", "Customer messages approved at first" ] }, { "_key": "k67", "_type": "tableRow", "cells": [ "Handoff delay", "Create the next task with full context the moment a step finishes", "None" ] }, { "_key": "k68", "_type": "tableRow", "cells": [ "Rework from errors", "Check the record against rules before it moves on", "Exceptions to a person" ] }, { "_key": "k69", "_type": "tableRow", "cells": [ "Status questions", "Answer “where is my order?” from live data", "None for internal; review for customers" ] } ] } ``` Start with internal steps, then customer-facing ones once quality is proven. ## Phase 5: Monitor and keep improving Track the Phase 1 metrics weekly. Watch for drift: a new product, team or system can bring back old delays. Re-run the diagnosis each quarter and move to the next process when gains level off. ## Metrics that show real improvement - Median and 90th percentile cycle time. - Waiting time per step. - Rework rate. - On-time completion rate. - Touch time per case (hours of human work). - Exceptions per 100 cases. Report the 90th percentile, not just the average. Customers remember the slow cases. ## Tools for each phase ```json:object { "_key": "k100", "_type": "table", "rows": [ { "_key": "k95", "_type": "tableRow", "cells": [ "Option", "Phases", "Limits" ] }, { "_key": "k96", "_type": "tableRow", "cells": [ "Viktor (AI employee)", "Measure from your tools’ data, diagnose in plain language, flag late cases, run the chasing and handoffs, and report weekly, all from Slack or Microsoft Teams", "Not a dedicated process-mining suite for very large event logs" ] }, { "_key": "k97", "_type": "tableRow", "cells": [ "Process mining (Celonis, UiPath Process Mining, Microsoft Process Advisor)", "Measure and diagnose at enterprise scale", "Cost, setup and data engineering" ] }, { "_key": "k98", "_type": "tableRow", "cells": [ "RPA (UiPath, Automation Anywhere, Power Automate)", "Rule-based execution on fixed screens", "Breaks when interfaces change" ] }, { "_key": "k99", "_type": "tableRow", "cells": [ "Workflow automation (Zapier, Make, n8n)", "Simple triggers between apps", "Each flow is built by hand" ] } ] } ``` Viktor is an [AI employee](https://viktor.com/ai-employee) who lives in Slack and Microsoft Teams and connects to 9,800+ tools, including Salesforce, HubSpot, NetSuite, Zendesk, Jira and Google Sheets. An operations lead can write: “Pull the last six months of orders from HubSpot and NetSuite, show me where orders wait longest between won and invoiced, and from next week chase the approvers on anything waiting more than two days.” He builds the analysis, posts the findings and runs the follow-ups on a schedule. Read [what an AI employee is](https://viktor.com/blog/what-is-an-ai-employee) for how he works. ## Common mistakes - **Optimising from opinion.** Measure first; the slow step is often not the one people complain about. - **Automating every variant.** Fix or remove rare variants instead of automating them. - **Averages only.** The 90th percentile shows the cases that hurt. - **No owner after launch.** Someone must watch the metrics and the exceptions. ## FAQ ### How is AI used for process optimization? AI analyses event data to find where processes wait and loop, predicts which cases will be late, and runs the waiting, chasing and handoff steps that cause most delays. ### What is the difference between process mining and AI process optimization? Process mining reconstructs and measures a process from event logs. AI process optimization uses that insight and goes further, explaining causes, predicting delays and carrying out steps. ### Which process should I optimize first? One with high volume, a clear start and end, data in your systems and visible pain, such as order-to-cash or support ticket resolution. ### How long does it take to see results? Measuring and diagnosing one process takes days with connected systems. Removing waiting steps shows up in cycle time within a few weeks. ## Find the bottleneck, then remove it Viktor works in Slack and Microsoft Teams, connects to 9,800+ tools and turns process data into faster cases. [Start free with up to $100 in credits, no credit card](https://app.viktor.com/signup), and ask him where your orders wait.