#ai#project management#workflow automation#predictive planning

AI in Project Management: Probabilistic Forecasts, Admin Automation, and What to Avoid

Use Monte Carlo forecasts to answer "when will it be done?" honestly, automate PM admin, and avoid per-person velocity tracking, high-risk under the EU AI Act.

📅 July 25, 2024✏️ Updated: September 27, 2026⏱ 6 min read✍ Web3 Listicle Editorial Team

Project management team collaborating on interactive, AI-driven dashboard displays showing predictive Gantt charts and resource indicators.

Projects run late far more often than plans admit. Bent Flyvbjerg's database of more than 16,000 big projects shows, as a review of his 2023 book How Big Things Get Done summarizes it, that only 8.5% came in on both budget and schedule, and just 0.5% delivered on budget, on schedule, and with the benefits promised. The database is weighted toward large and megaprojects across many fields, so read it as a warning about plans made without an outside view, not as a prediction for any one team.

AI can help in two ways. It can take over the administrative work that consumes a project manager's week, and it can support better forecasts by making the uncertainty visible instead of hiding it in a single date. This guide covers both, plus one use to avoid.

Forecasting: ranges from your own history

The question every stakeholder asks is "when will it be done?" A single date answers it badly, because it hides how likely that date is. Teams using agile methods have a better option that needs no specialist AI at all: a Monte Carlo forecast based on their own throughput.

A worked example

A team has 60 backlog items left. The method is the one behind flow-metrics tools such as ActionableAgile, which answer "when will it be done?" with probabilities instead of a date. Over the last 12 weeks it finished 3, 5, 2, 6, 4, 4, 7, 3, 5, 1, 4, and 6 items, an average of about 4.2 a week.

The naive forecast divides 60 by 4.2 and says about 14.4 weeks. A Monte Carlo simulation instead samples a random week from that history, adds it up until 60 items are done, and repeats 100,000 times. The results:

Confidence Weeks to finish 60 items
50% 15
85% 16
95% 18

The same method answers the reverse question. How much can the team finish in the next 8 weeks? The median is 33 items, but there is an 85% chance of finishing at least 28. If the release needs 32 items, the honest answer is "probably, but not certainly," and the conversation can move to which items are essential.

Two caveats. The forecast assumes the next few months look like the last few, so recalculate when the team or the kind of work changes. And backlogs grow as work is discovered; tracking how much the backlog has grown in past projects and adding that to the simulation makes it more realistic.

This is where AI tools add value on top of the method: pulling throughput data from Jira or similar tools automatically, running the simulation weekly, and explaining in plain language what changed since last week's forecast.

Predictive Gantt charts and risk timeline indicators illustrating how AI anticipates project bottlenecks.

For large, one-off projects

When there is no throughput history, as with a new building, a system migration, or a first-of-its-kind product, use reference class forecasting: look at how similar past projects actually turned out and adjust the plan toward that distribution. Flyvbjerg's work is built on this, which grew out of Kahneman and Tversky's "outside view". AI can help find and summarize comparable projects, internal or published, but the adjustment is a judgment call.

Automating project admin

A large share of a project manager's week goes to work that language models handle well:

  • Meetings. Transcripts turned into decisions, action items with owners, and open questions, posted to the right place.
  • Status reports. A first draft built from ticket movement, completed milestones, and blockers, which the PM edits rather than writes.
  • Threads and documents. Summaries of long discussions for people joining late (Jira Service Management, for example, can summarize a work item's comments with Atlassian's AI features); first drafts of requirements, risk logs, and release notes.
  • Backlog hygiene. Flags for duplicate tickets, tickets with no activity for weeks, and items missing acceptance criteria.
  • Scope checks. New requests compared against the approved scope document, with anything that does not map to it flagged for a decision.

These are low-risk because a person reviews the output, and the time saved is easy to measure. Start here.

What to avoid: individual velocity tracking

Some tools offer to measure each developer's speed and assign work accordingly. There are good reasons not to.

It breaks the measure. Once individual output is tracked, people optimize the number: splitting tickets, avoiding hard problems, skipping reviews and mentoring that do not show up in their count. This is Goodhart's law, and it reliably makes teams slower overall.

It is now regulated in the EU. Annex III of the AI Act lists AI used to allocate tasks based on individual behavior or personal traits, and AI used to monitor and evaluate workers' performance, as high-risk. After the 2026 Digital Omnibus on AI (Regulation 2026/1744), those obligations apply from 2 December 2027 for stand-alone systems: risk management, data governance, human oversight, transparency to workers, and documentation. Works councils and employee representatives in several EU countries also have consultation rights over monitoring tools.

Forecast at the team level, as in the example above. It is more accurate anyway, because team throughput is more stable than any individual's.

A multi-variable resource allocation matrix matching developers to tasks based on historical skill profiles.

Data and security

Project tools hold code, architecture documents, customer names, and internal plans. Connect AI assistants only through enterprise plans that exclude your data from model training, respect the permissions of the underlying tools (an assistant should not show someone a ticket they cannot open), and keep audit logs. Our guide to AI governance frameworks covers approved-tool policies.

Getting started

  1. Automate meeting notes and status report drafts for one team, and measure the hours saved after a month.
  2. Pull 12 weeks of throughput data and run a Monte Carlo forecast for the current release. Share the range, not a single date.
  3. Recalculate weekly, and track whether actual delivery lands within the forecast range; that is how you learn to trust it.
  4. Add backlog hygiene and scope checks once the basics work.
  5. Keep evaluation of people out of the AI tools.

For broader automation of business processes, see workflow automation. For forecasting methods beyond projects, see AI financial forecasting and AI for strategic decisions.


This guide is for informational purposes only. Evaluate tools and workplace monitoring rules with qualified technology, HR, and legal advisors.

Frequently Asked Questions

The most reliable method is not a complex model but a Monte Carlo simulation on the team's own history: sample past weekly throughput thousands of times to see how long the remaining work might take. The output is a range with probabilities, such as 50% likely by week 15 and 85% likely by week 16, which is more honest than a single date.
Rarely, for large ones. Bent Flyvbjerg's database of more than 16,000 projects, described in his 2023 book How Big Things Get Done, found only 8.5% delivered on both budget and schedule, and 0.5% on budget, schedule, and expected benefits.
The administrative ones: meeting notes and action items, status report drafts from ticket data, summaries of long threads, first drafts of risk logs and requirements, and flagging tickets that look stalled or duplicated. These save real time with low risk because a person reviews the output.
It is a bad idea for two reasons. Individual velocity numbers are easily gamed and damage collaboration, and in the EU, AI used to allocate tasks based on individual behavior or to monitor and evaluate workers' performance is classified as high-risk under the AI Act, with obligations applying from December 2027. Forecast at the team level instead.
It can help by comparing new tickets and requests against the approved scope document and flagging items that do not map to it. The decision on whether a change is in scope, and what it costs, stays with the project lead and sponsor.

Share this article