AI in Project Management: Case Studies and What They Prove
Compare documented project work at Discovery and JE Dunn, plus an engineering planning example from NASA JPL. Read the evidence carefully before borrowing a method for your own project.
Updated 6 min read
AI in project management case studies show how a particular team applied a method to a defined problem; they do not prove that any AI tool will improve every project. The most useful cases explain the input, human review, observed result, and limits well enough for you to design a small test of your own.
Below are two workplace cases, test preparation at Discovery Limited and construction progress review at JE Dunn, plus an engineering comparison from NASA JPL’s ASPEN. Tools like fenbs preserve the lesson you take from a case and the separate human decision to try it. None of these organizations is presented as a fenbs customer.
What kind of evidence does each case provide?
Separate a published account of use from a controlled comparison. A practitioner story can reveal a workable process without establishing how much of the result came from AI. A technical project description can explain a mechanism without reporting an independently measured business outcome. Both can be useful when you keep that distinction visible.
| Documented example | Evidence you can inspect | Question it leaves open |
|---|---|---|
| Discovery Limited | PMI’s account of a company AI-adoption campaign and a test-preparation example. | Would the same method improve accepted test coverage in your project? |
| JPL ASPEN: engineering comparison | The engineering team’s description of modeled goals, resources, and planning. | How much work would your own constraint model require? |
| JE Dunn and Buildots | A supplier’s case study describing progress review and trade coordination. | Would earlier progress warnings change your team’s next action? |
Use what AI project management means for the broader definition. Here the question is narrower: what can these particular accounts support, and what should you verify before acting on them?
What does Discovery’s test-preparation example actually show?
PMI’s Discovery Limited case study (opens in a new tab) describes a South African financial services organization encouraging AI use through training, shared prompts, examples, and a project management office campaign. One highlighted example used Microsoft Copilot to help prepare user acceptance tests, the checks people use to decide whether delivered work meets their needs. The account emphasizes that people still reviewed the work.
This is evidence of a specific adoption method and a reported drafting use case. It is not an independent trial showing that projects finished sooner overall. The transferable idea is to give people an approved starting process and a practical example they can inspect, then measure the accepted output rather than the size of the first draft.
For a hypothetical US software team, a useful pilot would take one approved requirement and ask AI for candidate test cases. A reviewer would compare them with the requirement, reject invented behavior, add missing failure paths, and record revision effort. That proposed pilot is our recommendation, rather than a result claimed by Discovery.
What can project teams learn from ASPEN’s planning approach?
NASA JPL describes ASPEN (opens in a new tab) as a framework using AI for planning and scheduling. Its documented problem is translating high-level goals into valid low-level commands for systems such as spacecraft, rovers, and communications antennas. The system represents operational knowledge, resources, and timing constraints rather than simply generating a written schedule.
This is an engineering comparison, rather than a deployment of a workplace project-management assistant. The page explains applications and technical components without establishing a universal business productivity improvement. Its value here is the design principle: a useful plan needs a representation of what makes an action possible, beyond a fluent description of tasks.
For example, a hypothetical installation project might require a site inspection before ordering equipment and a qualified crew before installation. Before asking AI to suggest a plan, write those dependencies explicitly. Then have the responsible people check whether the proposed sequence respects them. Do not assume a general chat assistant performs the constraint checking described in the NASA system.
What does JE Dunn’s construction example show?
Buildots’ JE Dunn case study (opens in a new tab) describes a US contractor using its platform on a high-rise residential project. The supplier reports that the work expanded from quality checks into forecasting and coordination with trade partners. It describes comparing actual progress against the schedule to identify potential bottlenecks and help teams consider changes to staffing.
This is a vendor-published account of a named customer’s workflow, with direct relevance to project delivery. It is not an independent experiment isolating AI’s effect on completion time. Read it for the documented process: make progress visible, discuss the warning with the relevant trade, and decide what action the evidence supports.
For a hypothetical construction pilot, choose one work area and compare an AI-assisted progress review with the current review process. Have a qualified person check whether flagged delays are real, actionable, or already known. Record false alarms as well as useful warnings, and track whether the accepted action actually changes delivery.
Preserve the confirmed result and next action as Where we left off in fenbs for a useful handoff. Keep site evidence in its approved system; the shared project record needs the reviewed conclusion and remaining work.
Which conclusions should you refuse to draw?
- A faster draft proves a faster project. Review, correction, integration, and rework still need to be measured.
- An engineering prototype proves a capability is deployed everywhere. Check the project description and distinguish demonstrations from operational use.
- Success in one organization predicts your result. Different data, staff, permissions, and failure costs can change the outcome.
- More recorded AI activity means more value. Count work people accept and use, with its errors and review effort.
Ask what the account does not say: the starting quality, selection of tasks, reviewer experience, missing attempts, and follow-up period. Missing details do not make a case worthless. They determine which statements you can reasonably make when asking colleagues to try the approach.
Avoid translating a company account into a legal or professional assurance. The NIST AI Risk Management Framework (opens in a new tab) provides a voluntary way to organize trustworthiness questions across AI work. It is a framework for your review, rather than evidence that one of these cases meets your organization’s requirements.
Industry-specific evidence belongs in its own assessment: see AI use cases in wealth management and AI in HR case studies for those separate questions.
Turn a case into a testable project lesson
Save the source link and a short lesson in fenbs: what happened, what you learned, and what you would do differently. Keep your interpretation separate from the reported result. A person can then decide whether to authorize a pilot and record the limits of that choice.
Create a free accountFree for one person · No credit card · Works with Claude, ChatGPT, Cursor and more
How can you run a pilot that produces useful evidence?
- Choose one repeated task with a clear reviewer and an output you can compare with the current method.
- Write the starting process and acceptance checks before using AI. Include the errors that would make the output unusable.
- Use approved information and comparable examples. Record which tool, instructions, and source material were used.
- Measure drafting, review, and correction together. Keep rejected results and explain why they failed.
- Ask the accountable person to continue, change, or stop the pilot based on the evidence. Save the result for the next team.
An initial pilot can teach you how to measure without proving a lasting improvement. Repeat it on different examples before expanding access or automating consequential actions. The guide to using AI assistants in project management covers the wider working process once the pilot has a clear purpose.
How to do this with fenbs
- Create a free account and add a project for one pilot. Write the question you want the pilot to answer.
- Add the case source to Links and listings and record its useful lesson. Label your own hypothesis as something still to test.
- Record the person’s approval and limits in Decisions and rules. Add tasks for preparation, review, and measurement.
- Ask a connected assistant to read the context before each session and save progress afterward, including failures and open questions.
fenbs keeps the shared record and human decisions; it does not run controlled experiments or supply a specialist scheduling engine. It is not the right fit if your only need is automated measurement inside an application. For one person it is free with no credit card.
Start with one transferable method and one result you can check. Keeping the source, interpretation, and local outcome separate makes the next decision easier to defend and gives your team something concrete to contribute to its AI at work practice.