AI and Digital Twins in Renovation Planning: Hype vs Useful

Renovation planning has attracted more technology promises than almost any other part of asset management. Digital twins, AI-driven condition assessment, generative scenario design — each arrives with a strong pitch and a demo that looks convincing.

Some of it genuinely improves decisions. Some of it adds cost, delay and a maintenance burden without changing what gets renovated or when. The useful test is not how advanced the technology is; it is whether it changes a decision you were otherwise going to get wrong.

What a digital twin is actually good for

“Digital twin” covers a wide range, from a geometric BIM model through to a live simulation fed by sensors. That range is the source of most confusion, because the versions have different costs and different payoffs.

Where it earns its cost: complex individual assets with high consequence and frequent intervention — a hospital’s technical plant, a tunnel, a treatment works. Here, a maintained model of systems and their interdependencies genuinely helps: you can test whether taking one system out of service is feasible, sequence work in a live environment, and avoid clashes that would be expensive to discover on site.

Where it usually does not: portfolio-wide renovation prioritisation across hundreds of similar buildings. To decide which twenty of two hundred schools to renovate first, you need comparable condition, cost and carbon data per building. You do not need a geometric model of each one. Building two hundred twins to answer that question is an expensive way to obtain data you could have captured directly — and the twins then need maintaining, or they quietly become fiction.

The failure mode is scope inversion: building high-fidelity models of everything before establishing which decisions are actually poorly informed.

Where AI genuinely helps

Three uses hold up well in renovation planning:

Filling and checking data. Portfolio records are incomplete and inconsistent. Models trained on comparable assets can estimate missing attributes, flag values that contradict the rest of the record, and prioritise which gaps are worth surveying. This is unglamorous and high-value, because bad data is the binding constraint in most portfolios.

Degradation prediction. Predicting how a component will deteriorate, given its type, age, use and environment, is a well-suited statistical problem — and it is the input that turns a condition snapshot into a schedule.

Searching the scenario space. The number of possible renovation programmes across a portfolio is enormous. Optimisation and machine learning are good at finding candidate programmes that a human planner would not have constructed, which can then be evaluated properly.

Notice what these have in common: each produces an input to a decision made by an accountable person. The AI narrows the field and quantifies consequences; it does not choose.

Where AI is oversold

“AI will tell you what to renovate.” Prioritisation embeds political, service and equity judgements — which school, which neighbourhood, which tenant group — that are not in the data and should not be delegated. A model can rank by cost-effectiveness or carbon; it cannot weigh those against each other on your behalf.

“AI removes the need for surveys.” Estimates from comparable assets are useful for planning at portfolio level and inadequate for committing to a specific intervention. Someone still has to look at the building before the works contract is signed.

“Real-time data is the answer.” For slow-degrading building fabric, annual or biennial condition data supports the decision perfectly well. Continuous monitoring is valuable for fast-failing technical systems and largely wasted on roofs.

Unexplainable recommendations. A recommendation that cannot be traced to its assumptions will not survive audit or a change of personnel. In public and regulated portfolios this alone disqualifies opaque models.

A practical filter

Before funding either, ask:

1. Which decision improves, and how would we know? If the answer is “better visibility”, it is not yet a business case.
2. What is the cheapest data that would improve that decision? Often it is a consistent condition survey, not a model.
3. Who maintains this in three years? An unmaintained twin is worse than no twin, because people still trust it.
4. Can the output be explained to an auditor? If not, it cannot carry a capital decision.

The version that works

In practice the effective setup is undramatic: consistent condition data across the portfolio, AI-assisted degradation models per asset, scenario comparison on cost, risk and carbon together, and detailed twins reserved for the handful of complex assets that justify them.

That is deliberately how Oxand Simeo™ is built — analysis assisted by AI, with the reasoning attached to every recommendation, so the plan can be defended rather than merely trusted. The 10,000+ predictive models built with asset owners are valuable because they are explainable, not because they are opaque.

The distinction between hype and useful is rarely about the algorithm. It is about whether anyone specified the decision first.

To discuss what would actually improve your renovation plan, talk to an Oxand expert or read about implementation best practices.