Gains:
- Ability to understand analogical, parametric and three-point (PERT) estimation methods and use artificial intelligence to produce time and resource estimation drafts.
- Ability to interpret forecast uncertainty (optimistic/possible/pessimistic scenarios, confidence interval) and base planning on a range rather than a single number
- Ability to recognize that the AI forecast is based on historical patterns and assumptions, and that the final budget and time commitment must be validated with team data and expert judgment
Estimation is the most difficult and risky task in project management. "How long will this take?" and “how many people does it take?” The answers to these questions form the basis of the schedule, the budget and the promise made to the customer. Poor estimating will sink even a well-managed project: too optimistic a time will lead to delays and penalties, too pessimistic a time will lead to losing the bid. Artificial intelligence is an attractive assistant in this field; gives a time in seconds. That's exactly why this is where it's most dangerous: AI can produce a confident number without knowing your team's actual speed. In this unit, we will cover forecasting methods, how to use AI safely in forecasting, and how to validate each forecast.
Basic forecasting methods
Let's define a few methods. Analogous guessing means looking at a similar job in the past and saying "that took 10 days, this one is similar"; It's fast but rough. Parametric estimating multiplies a unit cost/time by the quantity (e.g. "0.5 days per page × 40 pages = 20 days"); is based on data. Bottom-up estimation estimates and aggregates each work package separately; It is the most accurate but slowest. Three-point estimation (PERT) acknowledges uncertainty: for each task, the optimistic (O), most likely (M), and pessimistic (P) duration are estimated and the weighted average taken.
The PERT formula is: Expected duration = (O + 4M + P) / 6. For example, for a task, if optimistic 4 days is the most likely 6 days, pessimistic 14 days is expected duration = (4 + 24 + 14) / 6 = 7 days. The standard deviation is (P − O) / 6 = (14 − 4) / 6 ≈ 1.7 days; This shows the margin of uncertainty. Instead of saying a single "6 days", a much more honest plan is to say "expected 7 days, deviation ±1.7 days".
In addition, two concepts are important. Effort is the amount of man-days required for the job (e.g. 10 man-days). Duration is how many days this job will take on the calendar. Two people can finish 10 man-days' work in 5 days; but not every job can be divided infinitely (nine women cannot give birth to a baby in one month). AI can confuse the two; you must separate it.
Method
When is it appropriate?
speed
accuracy
Role of AI
analogical
Early phase, little information
high
low
Suggests similar project pattern
parametric
If there is a measurable unit
high
medium
Generates formula and unit outline
bottom up
If there is detailed WBS
low
high
Collects packets of drafts
Three-point (PERT)
If uncertainty is high
medium
medium-high
Suggests O/M/P scenarios
Step by step: Safe prediction with AI
- Put the team data first. Give the AI your own past realizations: “Similar tasks took this long on our team.” So the prediction is based on your data, not generalization.
- Select the method. If uncertainty is high, ask for PERT; parametric if there is a measurable unit.
- Ask for three dots, not single numbers. Ask for O/M/P values and assumptions for each package.
- Make assumptions visible. Print the assumption underlying each estimate (e.g. "test environment ready", "customer returns in 2 days"); A wrong assumption is a wrong guess.
- Add source and share. Separate effort from time, consider vacation and availability (e.g. a person is actually productive 4 days a week).
- Calibrate with the team. Review the final estimate with the people who will do the work. A prediction becomes reliable when the person who makes it owns it.
Attention: Ask the AI "how many days will this work take?" Asking and writing the resulting number directly into the chart is the most dangerous form of prediction. This number does not include your team's data; it is just a general pattern and may be a hallucination.
three mini cases
Case 1 — Honest plan with PERT. A software team asked the AI for O/M/P predictions and assumptions for a 12-task module. AI gave draft; The team corrected each mission with its own historical data. The total expected time was 38 days, the pessimistic total was 52 days. PM gave the customer a 45-day deadline and left a 7-day margin. The project was finished on the 44th day; If there was no share, there would be a delay.
Case 2 — Effort/time confusion. A coordinator asks AI, "How many days will it take to complete 80 man-day jobs?" he said. The AI answered "80 days" because it confused effort with duration. Knowing that the work could be done by 4 people in parallel, the coordinator corrected the time to ~25 days (not 20 because it was not divisible). Lesson: Humans should control the AI's effort/time split.
Case 3 — Unsubstantiated generalization. One PM wrote on the chart that AI responded "an API integration typically takes 5 days." The team's past experience with this API provider was poor and the documentation was poor; The work took 13 days. If team data were asked, this would be seen from the beginning. Lesson: generalization is no substitute for team reality.
Weak prompt / Strong prompt
Weak prompt:
How many days does it take to set up an e-commerce site?
This prompt forces the AI to generate a made-up number without knowing your project's data. There is no team, no scope, no uncertainty.
Powerful prompt:
Your role: an estimation specialist. Context: Below are 8 work packages and similar past assignments of our team. Task: Generate a THREE-POINT (optimistic O / most likely M / pessimistic P) duration estimate for each package and calculate the PERT expected duration with (O+4M+P)/6. Rules:- Just reference the historical data I have given; adding general market assumption.- Write the assumption underlying each estimate in a single line.- Show effort (man-days) and duration (calendar days) in a separate column.- Mark the package you are not sure about with "[team confirmation required]". Output: table (Package | O | M | P | PERT | Effort | Assumption). Historical data:- Payment integration (history: 9,11,8 days)- Product catalog (history: 6.5 days)- ... (other packages)
This claim is strong: method (PERT), data source limit, assumption requirement, and effort/time distinction are clear.
Additional templates:
# Assumptions hunterList all the assumptions you have silently accepted as true so that this prediction turns out to be true. For each assumption, estimate how much it will affect the time if it turns out to be wrong.
# Sensitivity analysisIf the most probable duration of the following task is M; How does the total project duration change if M increases by 20%?Which 3 tasks are most sensitive to total duration? (These tasks must be watched first.)
# Contingency recommender Based on the PERT deviations below, suggest a reasonable time allowance percentage for the entire project and explain what you base this on. This is a suggestion; The final share belongs to the manager.
Common mistakes
- Sticking to a single number: Saying "10 days" hides uncertainty; O/M/P range is honest.
- Confusing effort with time: 40 man-days is not 40 calendar days.
- Assuming 100% availability: No one can devote 5 days a week to this task alone; There are meetings, leaves, interruptions.
- Mistaking generalization for team data: AI's "typical time" is not your team's speed.
- Not writing down assumptions: An unseen assumption is a guess that explodes.
- No margin: If there is uncertainty, a reasonable margin of time/budget is mandatory.
Tip: When you're done guessing, ask yourself: "Is there a 90 percent chance of this number coming true or a 50 percent chance of getting it?" Most one-point bets are actually 50 percent bets; Know this when making a promise to the customer.
In summary
Estimation is the riskiest job of the project. Analogical, parametric, bottom-up, and three-point (PERT) methods offer different speeds and accuracy. AI is fast at draft forecasting and scenario generation, but it generalizes without knowing your team's data and can confuse effort/time. Safe guess; Putting team data first, working in three points, making assumptions visible, separating effort and time, and calibrating the final number with the team. The committed date and budget always undergo human verification.
Application task
Select 5 tasks from your project. Ask the AI for the O/M/P estimate and PERT expected time for each (give your team's historical data anonymously). Then write out the assumption behind each task and review the estimate with a team member who will do the work. Correct at least two estimates based on team data and determine a reasonable time margin for the total.
checklist
- [ ] I based the estimate on my team's historical data, not on generalization.
- [ ] I practiced O/M/P (three dots) instead of odd numbers.
- [ ] I wrote the assumption underlying each prediction.
- [ ] I separated the effort (man-day) and the time (calendar).
- [ ] I took into account real constraints such as availability/permission.
- [ ] I left a reasonable margin for uncertainty and calibrated the estimate with the team.