AI-enhanced cost estimation for dyke construction with RCF-AI

Estimated reading time: 5 minutes

A practical framework for more reliable project costs’ forecasts

One of the most notorious challenges in infrastructure projects is figuring out if the preliminary cost estimation is in fact realistic. In our work with public sector clients and infrastructure stakeholders, we have seen how traditional estimation approaches often rely heavily on expert judgment, fragmented historical data and limited transparency around assumptions. To address this, we introduce RCF-AI, our AI-enhanced reference class forecasting tool.

For this case we apply the RCF-AI tool specifically to dyke reinforcement projects. It combines predictive modelling, explainability, inflation risk analysis and benchmarking against similar projects—all within a single, intuitive dashboard.

This article walks you through the tool and shows how it helps to answer three critical questions:

1. What drives the predicted project cost? 
2. How should inflation risk be measured and incorporated? 
3. What can we learn from similar historical projects? 

From data to insight: the Management Dashboard

The starting point is a centralized dashboard containing information on both historical and ongoing dyke reinforcement projects.

Here, users can quickly search or select a project and immediately access its defining characteristics.

For example, selecting a river dyke project reveals key indicators such as:

• Project length
• Completion year
• Population density
• Main failure mechanism group

The project is also visualized spatially via a polygon, providing geographic context within the estimation workflow.

Predicting cost with context

At the core of the tool is a predictive model that estimates cost per kilometer based on the selected project characteristics.

Behind the scenes, the model incorporates 50+ project features, including:

• Geospatial data (soil type, residential area overlap)
• Environmental constraints (Natura 2000 zones)
• Technical characteristics (failure mechanisms, design complexity)

For completed projects, users can compare:

• Actual cost (inflation-adjusted)
• Model-predicted cost

This enables immediate validation of model performance and builds trust in its outputs.

For ongoing projects, the model serves as a forward-looking estimate—grounded in empirical data rather than intuition alone.

Opening the ‘Black Box’: cost driver explainability

Predictions alone are not enough. Decision-makers need to understand why a project is expected to cost what it does.

That’s why explainability is a central feature of the tool.

The dashboard includes a waterfall chart that breaks down how each project characteristic influences the predicted cost relative to a baseline.

For example:

• Lower overlap with protected environmental areas may increase costs
• Lower risk reserves may decrease expected costs
• Certain soil types or urban density levels may significantly shift the estimate

A supporting table highlights the top contributing factors quantitatively, ensuring transparency and auditability.

This transforms the model from a “black box” into a decision-support system, allowing experts to:

• Challenge assumptions
• Identify cost sensitivities
• Communicate findings clearly to stakeholders

Measuring what matters: Inflation risk

Inflation is one of the most underestimated risks in infrastructure cost estimation.

Rather than relying on generic indices, the tool introduces a dyke-specific inflation framework.

The approach works as follows:

1. Decompose project cost structure using detailed cost breakdowns
2. Identify key cost components such as:
   • Labor (~33%)
   • Steel (~17%)
   • Asphalt (~15%)
3. Construct a custom inflation index weighted by these components

This allows users to:

• Track historical dyke-specific inflation trends
• Estimate forward-looking inflation risk tailored to the project
• Integrate inflation uncertainty directly into cost forecasts

The result is a far more realistic view of future materials costs and associated risks.

Learning from the past: Reference Class Forecasting

Even the best models benefit from grounding in reality. This is where the second core component of the tool comes in: the reference class framework.

Instead of only asking:

“What do we think this project will cost?”

We could also ask:

“What did similar projects actually cost?”

The system automatically identifies comparable historical projects based on multidimensional similarity across:

• Technical characteristics
• Geographic features
• Environmental constraints

For these projects, users can view:

• Key attributes
• Actual cost outcomes

From this, the tool generates a cost distribution of comparable projects, showing where the current estimate falls relative to historical reality.

This provides a powerful sanity check:

• Is the estimate aligned with past outcomes?
• Is it overly optimistic or conservative?
• Should we revisit assumptions?

Final thoughts

The strength of this tool lies in combining three complementary perspectives:

1. Model-based prediction — data-driven and scalable
2. Explainability — transparent and interpretable
3. Reference class benchmarking — grounded in real-world outcomes

Together, they provide a robust evidence base for decision-making.

The goal of this solution is not to replace expert judgment—but to strengthen it.

By integrating AI with domain knowledge and historical data, the tool helps project teams:

• Increase confidence in early-stage estimates
• Identify risks before they materialize
• Make more informed, defensible decisions

In a field where uncertainty is inevitable, better insight is the most valuable asset.

Read more about the HWBP pilot results, the methodology, validation and the coverage:

HWBP Pilot results

RCF-AI System card