Sustainability Atlas
Climate Tech & Data·15 min read·

Case study: AI for energy and emissions optimization - A sector comparison with benchmark KPIs

Where the value pools are and who captures them. A sector comparison with benchmark KPIs for AI-driven energy optimization.

Case study: AI for energy and emissions optimization - A sector comparison with benchmark KPIs

Energy and carbon are now operational variables, not just "ESG reporting." In the EU, electricity and gas price volatility, EU ETS exposure (directly or via suppliers), and CSRD/ESRS reporting pressure are pulling energy, production, and finance into the same room. In that room, optimization beats reporting-only--but only if you pick the right sector wedge and the right KPI.

This case study maps the value pools for AI-driven energy and emissions optimization, compares sectors using benchmark KPIs, and explains who captures the value (operators, OEMs, integrators, or software).

TL;DR

  • The biggest near-term value pool is fuel + electricity savings in energy-intensive processes where setpoints can be safely adjusted (cement kilns, furnaces, steam systems, large HVAC, data center cooling).
  • The most defensible startups tie models to control decisions (closed-loop or operator-in-the-loop), not just dashboards.
  • Benchmarking starts with intensity KPIs (energy/unit output; tCO2/unit output) plus reliability KPIs (uptime, variability, quality) because energy is usually constrained by quality and throughput.
  • In heavy industry, small percentage improvements (2-4%) can mean six-figure to seven-figure annual savings per site--and fund deployment without big capex.
  • Compliance (CSRD/ESRS in the EU; evolving climate disclosure requirements elsewhere) increases demand for better data, but the winning products monetize operational savings first and treat reporting as a byproduct.

Who this is for

  • EU-focused founders building climate software, industrial AI, or "energy intelligence" products.
  • Teams deciding between wedges: data centers, cement, steel, chemicals/ammonia, commercial buildings, and biomaterials manufacturing.
  • Founders selling into asset-heavy customers with long capex cycles and strict operational constraints.

The opportunity map: where the value pools are (and who captures them)

Energy and emissions optimization creates value in five buckets. The key founder question is: Which bucket pays you, and who is structurally positioned to capture it?

  1. Direct energy savings (fuel + electricity)

    • Value driver: reduce kWh/GJ for the same output and quality.
    • Who captures: mostly the asset operator (P&L owner).
    • Who can charge: software vendors with credible M&V + integration partners.

  2. Carbon-cost exposure reduction

    • Value driver: lower Scope 1/2 emissions intensity, reduce ETS exposure, improve procurement outcomes.
    • Who captures: operator + sometimes customers (green premiums).
    • Who can charge: vendors that can prove emissions impacts and align with standards.

  3. Throughput/quality uplift (the hidden value pool)

    • Value driver: stabilizing a process reduces downtime, rework, variability; sometimes worth more than energy.
    • Who captures: operator (operations + sales).
    • Who can charge: vendors embedded into production workflows (APC + ML).

  4. Capex deferral / better capex allocation

    • Value driver: using AI to debottleneck can delay expensive upgrades; better evidence for what to upgrade first.
    • Who captures: operator and sometimes EPC/OEM ecosystem.
    • Who can charge: vendors that can connect ops data to investment decisions (finance-grade narratives).

  5. Compliance + auditability

    • Value driver: CSRD/ESRS reporting readiness, traceability, consistent "single source of truth."
    • Who captures: operator (risk reduction) and auditors/consultancies (service revenue).
    • Who can charge: usually modest unless bundled with real operational wins.

Founder takeaway: If your product only improves compliance, you will compete with consultants and ERP modules. If it improves energy + reliability and makes compliance easier as a side effect, you can price on ROI and grow inside the account.

Sector comparison with benchmark KPIs

Below is a founder-oriented benchmark comparison: what to measure, what "good" looks like, and where AI realistically helps.

Notes on benchmarks:

  • "Typical" values are global/industry references; site-specific baselines vary by technology, feedstock, electricity mix, and utilization.
  • Use benchmarks to size ROI and prioritize wedges--not as a substitute for plant-specific baselining and measurement and verification (M&V).

Cement (kiln + cooler)

  • Benchmark KPIs: Kiln thermal energy intensity (GJ/t clinker), clinker-to-cement ratio, fuel-derived CO2 intensity, quality variability
  • Reference benchmark: Kiln thermal energy intensity around 3.6 GJ/t clinker (2022 milestone); clinker-to-cement ratio around 0.71 (2022 milestone)
  • Where AI wins first: Setpoint optimization, alternative fuel management, variability reduction
  • Typical AI impact: 1-4% fuel/energy improvement when tied to control + operator workflow
  • Buyer: Plant manager + energy manager + process engineering

Steel (rolling / furnaces / EAF)

  • Benchmark KPIs: CO2 intensity (tCO2/t crude steel), energy intensity (GJ/t), furnace gas consumption per tonne, yield/quality variance
  • Reference benchmark: Global CO2 intensity around 1.92 tCO2/t crude steel (industry indicator)
  • Where AI wins first: Reheating furnace optimization, scheduling + energy peaks, anomaly detection
  • Typical AI impact: 1-3% energy improvement on large furnaces; bigger wins from uptime/quality
  • Buyer: Site ops + energy team; sometimes corporate energy

Ammonia / chemicals

  • Benchmark KPIs: Direct CO2 intensity (tCO2/t), energy intensity (GJ/t), hydrogen/steam balance KPIs
  • Reference benchmark: Ammonia around 2.4 tCO2/t production (direct CO2 basis, reference figure)
  • Where AI wins first: Advanced control, heat integration optimization, constraint-aware scheduling
  • Typical AI impact: 1-3% on energy-intensive steps; higher if constraints were unmanaged
  • Buyer: Plant ops + process engineering

Data centers

  • Benchmark KPIs: PUE, cooling kWh, percent time in economization/free cooling, thermal compliance
  • Reference benchmark: Industry average annual PUE around 1.56 (2024 survey)
  • Where AI wins first: Cooling optimization + dynamic controls, predictive maintenance
  • Typical AI impact: 5-15% of cooling energy; total facility impact depends on baseline
  • Buyer: Facilities ops; colocation provider sustainability leads

Commercial buildings

  • Benchmark KPIs: Energy Use Intensity (kWh/m2-year), HVAC runtime vs occupancy, peak demand (kW), comfort complaints
  • Reference benchmark: Benchmarks are building-type-specific; focus on baseline + "waste" percent
  • Where AI wins first: HVAC optimization, demand response, fault detection and diagnostics
  • Typical AI impact: 5-20% building energy in well-instrumented sites
  • Buyer: Property ops; ESCOs; FM providers

Biomaterials / fermentation-based manufacturing

  • Benchmark KPIs: kWh/kg product, steam per batch, yield percent, cycle time, CIP/SIP energy, downstream separation energy
  • Reference benchmark: Highly process-specific; benchmark internally across lines and batches
  • Where AI wins first: Batch optimization, utility system optimization (steam/chilled water), yield-energy co-optimization
  • Typical AI impact: 2-10% utility savings; often bigger gains from yield and cycle time
  • Buyer: Operations + process development; COGS owner

What these benchmarks imply for wedge selection

  • High-temperature continuous processes (cement, glass, steel furnaces): Small improvements compound because systems run near-constantly and energy is huge. Trust and safety constraints are the main barriers.
  • Discrete/batch processes (biomaterials): Energy is intertwined with yield and cycle time. The best wedge is not "energy dashboard," it is "optimize COGS with energy as a constraint."
  • Data centers and buildings: The KPI is standardized (PUE/EUI), integration is easier, sales cycles can be shorter--but competition is intense. Differentiation comes from control integration, not analytics.

Why It Matters

1) Energy is a top-line competitiveness lever in Europe

In energy-intensive sectors, energy can be one of the largest controllable operating costs. When prices spike, "nice-to-have efficiency projects" become survival projects. AI that can reliably deliver even 2-4% savings (without disrupting quality or uptime) can be finance-grade.

Concrete examples of why small percentages matter:

  • A steel site example (rolling furnace context) shows energy bills in the EUR 5M/year range for a single furnace system; a few percent becomes meaningful cash savings and fast payback.
  • In cement, thermal energy intensity is a core KPI because kiln fuel drives both cost and emissions; industry roadmaps reference around 3.6 GJ/t clinker as a recent benchmark milestone.

2) Reporting pressure is increasing demand for operational-grade data

EU sustainability reporting requirements are pushing companies to formalize data flows and controls. CSRD (Directive (EU) 2022/2464) expands and strengthens sustainability reporting obligations, and the ESRS provide the detailed reporting standards used by in-scope companies.

This matters to AI products because:

  • Better reporting typically requires better metering, data lineage, and governance--the same prerequisites for optimization.
  • Founders can "ride" the compliance wave, but should monetize through operations ROI first.

3) Standards are becoming the language of trust

Industrial buyers and auditors will ask: "How do you measure savings? How do you attribute emissions reductions?" Products win faster when they align with recognized standards:

  • ISO 50001 for energy management systems.
  • GHG Protocol (Scopes 1/2/3) for emissions accounting, especially when optimization claims affect reporting.

4) The SEC climate rule is a cautionary tale for regulatory volatility

Even if you are EU-focused, multinational customers (and their investors) operate across regimes. The SEC finalized climate disclosure rules in 2024, and subsequent legal/political developments have shaped its trajectory. For founders, the lesson is to build compliance support as a feature--not the entire business model--because requirements can change.

Key Concepts

Energy optimization vs emissions optimization

  • Energy optimization reduces kWh/GJ for a given service/output.
  • Emissions optimization depends on the energy mix and process chemistry:
    • Cement has large process emissions (calcination), so energy savings help but do not solve everything.
    • Steel route matters (BF-BOF vs EAF) and electricity carbon intensity changes the emissions outcome.
    • Ammonia is highly emissions intensive on a direct basis; route (SMR vs coal gasification) is decisive.

The KPI stack: from "plant truth" to "board truth"

A practical stack founders can implement:

  1. Primary intensity KPI (the "north star")

    • Example: GJ/t clinker, tCO2/t steel, PUE, kWh/kg biomaterial.

  2. Constraint KPIs (what prevents naive optimization)

    • Quality metrics, throughput, downtime, safety envelopes, environmental permit limits.

  3. Control KPIs (leading indicators)

    • Variability (e.g., standard deviation of key process variables), setpoint adherence, alarm rates.

  4. Finance KPIs

    • EUR saved, payback period, marginal cost of abatement (EUR/tCO2), capex deferral narrative.

"AI" in this category usually means one of three things

  • Predictive models (forecast consumption, quality, failures).
  • Optimization (compute recommended setpoints or schedules under constraints).
  • Control integration (deliver changes through existing control systems or operator workflows).

The best products combine all three, but ship in phases.

What's Working and What Isn't

What's Working

1) A tight wedge tied to a controllable knob

Winning products pick one system where you can change outcomes:

  • kiln fuel/air setpoints,
  • reheating furnace loading modes,
  • chiller sequencing,
  • steam header pressure,
  • batch utility scheduling.

2) Operator-in-the-loop trust before autonomy

In heavy industry, credibility is built by:

  • giving recommendations with constraints,
  • showing why now (drivers),
  • measuring savings conservatively,
  • allowing operators to override safely.

3) M&V that finance believes

Simple rule: if the CFO cannot defend it in an audit, it will not scale.

  • Baselines, counterfactuals, and drift detection are product features, not consulting add-ons.

4) Integration that respects plant reality

Plants already have historians, SCADA/PLC, APC, and maintenance systems. The best AI products augment existing systems, then gradually increase automation.

What Isn't Working

1) "Carbon dashboards" without operational levers

Dashboards can be necessary, but alone they rarely create budget authority.

2) Models that ignore constraints (quality, safety, permits)

If you cannot encode constraints, your recommendations get ignored.

3) Overselling savings

Founders lose trust by quoting blanket "20-30%" claims without context. In mature plants, 2-4% can be excellent--if it is real, repeatable, and safe.

4) Underestimating the capex reality

Even "software" deployments often require:

  • meter upgrades,
  • sensor calibration,
  • networking and cybersecurity approvals,
  • sometimes variable frequency drives or actuators.

Plan for a world where "minimal capex" is still some capex.

Examples

Example 1: Data centers - DeepMind cooling optimization at Google

Google/DeepMind reported that a machine learning system achieved a 40% reduction in energy used for cooling, translating into 15% reduction in overall PUE overhead (site-specific, after accounting for other losses). The key lesson is not "use Deep RL"--it is that cooling is a controllable subsystem with fast feedback loops and rich sensor data.

Founder lessons:

  • The KPI is standardized (PUE), so benchmarking is easier.
  • The integration point is clear (cooling controls), so you can close the loop.
  • As compute density increases, the value of better thermal control rises.

Example 2: Steel - ArcelorMittal Belval + AI energy performance (Energiency)

A published success story describes how ArcelorMittal's Belval site tested an AI/digital tool to improve energy performance in a reheating furnace context. It references:

  • large-scale data processing (1600+ data streams),
  • a high-reliability model,
  • and reported realized savings: more than 3% on the energy bill, represented as EUR 150,000 and 9 GWh annually in that context.

Founder lessons:

  • Start with one furnace, prove savings, then expand.
  • Energy optimization can also surface operational issues (e.g., deposits) and reduce shutdown risk--often the "hidden ROI."

Example 3: Cement - Heidelberg Materials + Carbon Re AI on top of ABB Expert Optimizer

A cement case study reports integrating AI "on top of" existing advanced control (ABB Ability Expert Optimizer) at Heidelberg Materials' Mokra plant, with key reported outcomes including:

  • 2.2% reduction in specific heat consumption,
  • 4% reduction in a fuel cost index,
  • 2% reduction in fuel-derived carbon emissions,
  • and reduced clinker quality variability (C3S variability reduction reported).

Founder lessons:

  • The winning wedge is often "augment the existing APC," not replace it.
  • Variability reduction is a strong selling point because it links to quality and throughput.
  • Cement is a prime EU wedge because fuel + ETS exposure + industrial decarbonization pressure converge.

Founder playbook: how to choose a sector wedge and win

Step 1: Choose a KPI with budget authority

A good wedge KPI is:

  • high-cost (big EUR impact),
  • frequent (daily/weekly decisions),
  • controllable (there is a knob),
  • auditable (M&V feasible).

Step 2: Pick a deployment motion that matches the sector

  • Cement / steel / ammonia: start with "recommendations + M&V," graduate to closed-loop control once trust is earned.
  • Buildings / data centers: start with controls integration, but differentiate on reliability + lifecycle management.

Step 3: Build standards into the product, not as PDFs

Treat standards as product requirements:

  • ISO 50001 alignment: energy review, EnPIs (energy performance indicators), continuous improvement loop.
  • GHG Protocol mapping: define system boundaries and how optimization affects Scope 1/2 metrics (and sometimes Scope 3 for product intensity).

Step 4: Price on value capture mechanics

Common models:

  • Subscription + services (simple, but may undercapture).
  • Share of savings (great alignment, requires strong M&V and customer trust).
  • Performance tiers (base fee + bonus if KPIs hit).

Step 5: Do not ignore CSRD/ESRS--bundle it

In the EU, many buyers will ask for CSRD-ready data flows even if they buy for energy savings. Make it easy:

  • provide exportable KPI definitions,
  • data lineage,
  • audit trails,
  • and documentation that fits sustainability reporting processes.

Action Checklist

  • Pick 1-2 sectors and define the one KPI you will sell first (e.g., GJ/t clinker, PUE, kWh/kg product).
  • Design a baseline + M&V plan (counterfactual, drift detection, seasonality, production mix normalization).
  • Map the control surface: which setpoints/schedules you influence, and what constraints must be respected.
  • Build the "trust layer": explainability, operator workflows, override, alerting, and change logs.
  • Package compliance as a byproduct: align KPI definitions with ISO 50001 and emissions reporting with GHG Protocol, and support EU CSRD/ESRS exports.

FAQ

Q: What KPIs should I benchmark first across sectors?

A: Start with energy intensity (kWh or GJ per unit output/service) and emissions intensity (tCO2 per unit output/service) at the system boundary you can influence. Then add constraint KPIs (quality, throughput, downtime). For data centers, start with PUE; for cement, kiln thermal energy intensity (GJ/t clinker) and clinker-to-cement ratio; for steel, tCO2/t and furnace fuel per tonne; for biomaterials, kWh/kg plus yield and cycle time.

Q: How does CSRD/ESRS change what customers will demand?

A: It increases demand for traceable, governed, auditable sustainability and energy data--often across multiple sites and suppliers. Practically, customers will want consistent KPI definitions, clear boundaries, and evidence. If you can show operational savings while making reporting easier, you will be pulled into both the operations and sustainability budgets.

Q: Does the SEC climate rule matter if I am EU-focused?

A: Indirectly, yes. Many EU industrials sell into global capital markets or supply U.S.-listed customers, and internal reporting often harmonizes across regimes. But regulatory requirements can evolve; founders should avoid betting the company purely on disclosure compliance. Build optimization-first products that can adapt to changing disclosure rules.

Sources

Topics

biomaterialsCSRDcapexstandardsSEC climate ruleenergy optimizationemissions reduction