It's 9:40 on Monday morning, monthly board review. On screen sits an alert produced by the AI tool: "cost overrun risk: 12% at project completion." The CFO asks: "Where does this number come from, which EVM data supports it?" The PMO leader scrolls through the slides but finds no CPI, no SPI — only a risk score with no traceability. The silence lasts three seconds too long, and the board's trust starts to crack.
Many think AI can calculate end-of-project cost projections in place of EVM. In reality, AI should never replace deterministic calculation, because only verifiable CPI and SPI hold up in front of a board: a predictive model remains a hypothesis to explain, not a data point to defend.
The difference between a calculated number and a forecasted number
A CPI of 0.87 comes from a fixed formula: budgeted cost of work performed divided by actual cost of work performed. Anyone working from the same data gets the same result. A 12% overrun forecast, by contrast, comes from a model that weighs historical variances, delay patterns and risk thresholds: it changes if the model changes, not only if the project data changes. This distinction, trivial on paper, is exactly what collapses the moment a PMO has to answer a direct question in a review.
The paper published in IJECS (vol. 15, no. 07, July 19, 2026) describes an XAI layer that maps delay-risk probability, calculated with XGBoost, onto the EVM performance quadrant, without replacing the EVM calculation itself. In other words, the model adds a risk label next to the CPI, not an alternative CPI. The publication on ResearchGate (February 19, 2026) points in the same direction: it describes EVM as a deterministic framework whose nature limits responsiveness in complex multinational contexts, but it does not propose replacing it with AI. Neither source supports the claim that AI "calculates" EVM in place of the PM.
Why an Opaque AI Layer Is Less Reliable Than a Well-Kept Manual EVM
Here's the point that flips the common intuition: a system that hides the EVM calculation behind an AI forecast presented as 'smarter' is less reliable than a well-kept manual EVM. With a traditional spreadsheet, however clunky, the PM can always trace back to every cell and show the formula. With an opaque predictive layer, that backward path often doesn't exist, and the board's question is left without a traceable answer.
The problem isn't the sophistication of the model, but its position in the decision flow. If the forecast visually replaces the CPI on the review slide, the PM ends up defending an output he didn't generate and can't break down. If instead the forecast sits alongside the CPI as an explanatory note, the PM regains control of the narrative: he can say 'CPI dropped from 0.91 to 0.87, and the AI flags the likely cause as vendor X's delay,' instead of just citing a risk percentage with no roots.
Declared simulation — Simulated scenario
An R&D project at the midpoint: CPI 0.87, SPI 0.91, both calculated by the EVM engine from actual cost and progress data. The AI layer, trained on the variance history of similar projects, flags a 12% overrun risk by project end. In the review, the PMO cites CPI and SPI as verifiable figures and shows the underlying formula; the 12% forecast isn't defended as a data point, but used to explain why the CPI is deteriorating, for instance by linking it to a delivery delay logged in the project records. The board accepts the explanation because the core number stays traceable, not because the forecast is more precise than the CPI.
Trade-off
- Benefit: A manually tracked EVM lets the PM trace back to every single data point and formula, making the board review defensible line by line.
- Cost: It requires constant update time and discipline in collecting cost and effort data at every reporting cycle.
- Risk: Manual entry exposes the process to human transcription errors or update delays that temporarily distort CPI and SPI.
- Prerequisite: A consistent, up-to-date project data source is needed, otherwise even the deterministic calculation produces fragile numbers.
- Limit: Across portfolios with dozens of parallel projects, manual EVM alone isn't enough to flag in time which variances deserve attention, and that's where an interpretive layer can help, without replacing the calculation.
This is what sets this article apart from other pieces in the same content cluster. The piece on real-time budget control covers continuous cost monitoring during execution, not the validation of an end-of-project forecast in front of the board. The cost variance analysis article explains how to read the deviations, but here the focus is on how to defend those numbers in a review, not how to compute them. The article on AI hallucinations in reporting deals with errors the model generates inside summary text, while the risk here is conceptual: mistaking a forecast for a verifiable calculation, even when the AI hasn't 'hallucinated' anything. The portfolio-level analysis piece compares multiple projects, while this article stays on a single project and its point-in-time EVM review.
The author's point of view
In ControlRoom the EVM engine stays deterministic: CPI, SPI and EAC are calculated from project data, with no AI model intervening in the math. AI comes in only afterward, to link a CPI shift to the evidence that explains it, such as emails, delivery logs or logged decisions. This choice has a cost: it doesn't produce a 'smarter' risk score ready to display to the board. In exchange, every number cited in a review stays traceable back to the underlying formula, and that traceability, not the model's elegance, is what holds up when people ask direct questions.
- Before presenting an AI alert to the board, check whether there's an actual calculated CPI or SPI that directly supports it.
- Ask the tool which deterministic formula sits behind the number: if there's no reproducible formula, treat it as a hypothesis, not as data.
- Build the review slide with CPI/SPI front and center, and the AI forecast as an explanatory footnote, never as the headline.
- Document the source of the EVM inputs (costs, hours, progress) used in the calculation, so you can answer questions about where the numbers come from.
- If the AI layer changes its forecast between reviews without the underlying EVM data changing, flag the inconsistency before reporting it to management.
Frequently asked questions
Can AI calculate CPI and SPI in place of the standard EVM formula?
No. The cited sources (IJECS, published July 19, 2026; ResearchGate, published February 19, 2026) describe layers that add an interpretation or a risk forecast on top of the EVM calculation, not a substitute for the deterministic CPI and SPI formula.
How do I present an AI overrun alert to the board without losing credibility?
Always cite the verifiable CPI or SPI as the headline number, then use the AI forecast only to explain the probable cause behind its deterioration, not as a figure to defend on its own.
Is a manually maintained EVM still valid if I have access to more advanced AI tools?
Yes, it stays valid and in some cases more defensible, because every figure is traceable by hand; the limit shows up on large portfolios, where the volume of data makes an added interpretive layer useful, without replacing the calculation itself.