AI Cost Forecasting for Projects: A Practical Guide

Forecasting the final cost of a project is one of the hardest jobs in project controls. Formula-based methods such as BAC ÷ CPI are quick but crude; bottom-up re-estimates are accurate but slow; expert judgement is valuable but inconsistent. AI cost forecasting promises something in between: forecasts that learn from data, update continuously and flag problems early.

Some of that promise is real. Some is marketing. This guide explains what AI forecasting can realistically do on projects, what data it needs, how to combine it with established earned value methods and how to keep it accountable.

Start from the established methods

AI does not replace the basics; it builds on them. The core forecasting methods — estimate at completion by CPI, by CPI × SPI, or by bottom-up estimate to complete — are explained in our guide to estimate at completion formulas. Every AI forecast should be compared with them. If an AI model says the project will finish 5% under budget while CPI has been 0.85 for six months, someone needs to understand why before the number goes to the sponsor.

What AI can add

Capability	What it does	Value
Pattern-based forecasting	Learns how cost performance evolves on similar past projects and projects the current one forward	Better than single-formula EACs when history is relevant
Anomaly detection	Flags unusual cost postings, accrual gaps or progress claims inconsistent with spend	Catches data problems before they distort forecasts
Driver analysis	Identifies which factors (contractor, region, package type, design maturity) correlate with overruns	Focuses management attention
Range forecasting	Produces probability distributions rather than single numbers	Supports contingency decisions
Narrative drafting	Summarises movements in the forecast and their causes	Saves time in reporting

The data you need

AI forecasting is a data problem before it is a modelling problem. You need:

Consistent historical projects — final costs, baselines and monthly performance, coded to a comparable structure.
Clean current data — actual costs with accruals, commitments, progress measured by agreed rules of credit and an up-to-date schedule.
Context variables — project type, size, contract form, location, design maturity at baseline, estimate class.

If historical projects are coded differently, or if actual costs lag by months because accruals are missing, the model will learn noise. Many organisations find that the biggest benefit of an AI forecasting initiative is the data clean-up it forces.

Methods, from simple to sophisticated

Statistical baselines

Simple regression on historical data — for example, how final cost growth relates to CPI at 30% complete — can outperform raw formulas and is easy to explain. Start here.

Machine learning models

Tree-based models and similar techniques can capture non-linear relationships between many variables. They need more data and more care: test them on projects they were not trained on, and check that their predictions make engineering sense.

Time-series and probabilistic methods

These produce forecast ranges that update monthly. A range ("P50 of 52.4 million, P80 of 54.1 million") supports better decisions than a single point, especially when sizing contingency. Our guide to contingency vs management reserve explains how ranges feed reserves.

Language models

LLMs are not forecasting engines, but they are useful around the forecast: summarising cost reports, extracting information from change logs and meeting notes, and drafting variance narratives. Keep them away from calculating the numbers themselves unless their output is checked against the source data.

A worked example

A portfolio office has data from 40 completed building projects. At 30% complete, a new project shows:

BAC = 20.0m   EV = 6.0m   AC = 6.6m   CPI = 0.91   SPI = 0.96

Formula forecasts:

EAC (CPI)        = 20.0 / 0.91                       = 22.0m
EAC (CPI x SPI)  = 6.6 + (20.0 - 6.0) / (0.91 x 0.96)
                 = 6.6 + 14.0 / 0.874                = 22.6m

A model trained on the 40 projects notes that, in this organisation's history, projects with a CPI near 0.91 at 30% complete and early-stage mechanical packages still open tended to deteriorate further, because mechanical and electrical packages were the main source of overruns. It forecasts a P50 of 23.1 million and a P80 of 24.0 million, and identifies the two open mechanical packages as the largest drivers.

The controls lead does not simply adopt 23.1 million. She asks the package managers for bottom-up estimates to complete for the mechanical packages. They reveal a design change that has not yet been priced. The final reported EAC is 23.0 million, with a range, and the narrative explains the evidence. The model's real contribution was pointing at the right packages early.

Controls that keep AI forecasts honest

Always show the formula forecasts alongside. Large differences must be explained.
Back-test. Run the model on completed projects as if they were live and compare its forecasts with what happened.
Explain the drivers. If a model cannot indicate why it forecasts what it does, use it as a flag, not as the number.
Version and label. Record which model version produced each forecast, and label AI-assisted figures in reports.
Human sign-off. The forecast in the report is owned by a named person, not by a tool.
Watch for drift. Models trained on old projects may not reflect new contract forms, markets or methods.

These principles, and how to apply them across schedule, cost and reporting, are set out in our printable AI in project controls playbook (PDF). Our earned value management cheat sheet (PDF) puts the formula methods on one page for comparison.

Questions to ask before buying an AI forecasting tool

Vendors increasingly offer AI forecasting as part of cost management platforms. Before committing, ask:

What data does the model train on? Your organisation's history, the vendor's pooled data from other clients, or both? Pooled data may not reflect your contracts, markets or methods.
Can it show back-test results on our projects? A credible vendor will run the model on a sample of your completed projects and show how its forecasts at 20%, 40% and 60% complete compared with final outcomes.
How does it explain a forecast? Look for driver-level explanations — which packages, cost types or variables moved the forecast — rather than a single number.
How are ranges produced? A P50 and P80 should come from a stated method, not from a fixed percentage applied to the point forecast.
What happens when data is missing or late? Models that silently fill gaps can produce confident forecasts from incomplete inputs.
Where is the data processed and stored? Cost data is commercially sensitive; check against your information security and AI policies.
Can we export the forecasts and inputs? Auditability requires that you can reproduce what the tool showed at the time.

A tool that answers these clearly is easier to govern and easier to trust. One that cannot is better used as an experimental second opinion than as the source of reported numbers.

Getting started without a data science team

You can begin with tools most controls teams already have:

build a clean table of historical projects in a spreadsheet;
calculate how final cost related to CPI and percent complete at standard points;
use that relationship as a sanity check on current EACs;
add anomaly checks on monthly cost data — postings that are unusually large, negative or coded to closed accounts;
use an approved AI assistant to draft variance narratives from the numbers you have verified.

Optimize All's free course Project Controls with AI includes forecasting exercises in Excel and Python, and the course AI for Data Analysis and Decision Making covers the data skills behind them.

Building the skill

Forecasting with governed AI is a theme throughout the body of knowledge for the PCI AI (opens in a new tab) PCL-AI Project Controls Leader credential. See our PCI AI partner page for an overview of PCI AI's three certifications.

Frequently asked questions

Is AI cost forecasting more accurate than earned value formulas?

It can be, when trained on relevant, clean historical data and checked against the formulas. Without good data, it can be worse — and harder to challenge.

How many past projects do I need?

There is no fixed number. Simple statistical relationships can be useful with a few dozen comparable projects; more complex models need more. Comparability matters more than volume.

Should the AI forecast replace the project manager's forecast?

No. It should inform it. The reported forecast should be owned by an accountable person who can explain its basis.

What is the quickest win?

Anomaly detection on cost and progress data. Cleaner inputs improve every forecast, whatever method you use.

Optimize All is the official marketing partner of PCI AI and Certuvo.

All articles