AI for Direct Marketing of Renewables and PPA Optimization: How Mittelstand Direktvermarkter Defend Margin in 2026

At 11:47 on a Wednesday in March 2026, the head of trading at a German Mittelstand direct marketer watches the intraday price for the 14:00 quarter-hour drop from 62 euros per MWh to minus 18 euros in under twenty minutes. Three wind farms in the portfolio are over-producing against the day-ahead schedule. The intraday desk sells 28 MWh into the falling market and books a loss against the imbalance position. By 14:15 the price has rebounded to 41 euros. Total damage for the hour: roughly 4,200 euros that should not have been lost. Multiplied across 8,760 hours per year, across a 600 MW portfolio, on negative-price days alone, the same pattern explains why three direct marketers in this segment have already exited the market in 2025 and 2026.

This is the operational reality of Direktvermarktung in 2026. EPEX day-ahead moved to 15-minute products on 30 September 2025^1,2, quadrupling bid complexity overnight. Intraday volatility is at record levels with tens-of-euros-per-MWh swings inside an hour now routine². Negative pricing on sunny windy days is normal, not an exception. Corporate PPA volumes in Germany crashed 84 percent year-on-year as cannibalisation broke the simple pricing models⁴. The decade-old playbook of hourly forecasts, rule-based ETRMs, and one senior trader per shift no longer fits the market.

This article is a practical guide to the AI agent layer that actually fits how direct marketing works in 2026. Seven high-ROI use cases, the honest build-vs-buy decision against ETRM-embedded AI and specialist trading vendors, the cost comparison, the architecture that respects REMIT, MAR, and EU AI Act, and a 90-day plan to take the first agent live.

Why Direct Marketing Is Now AI-or-Die

Four structural changes hit Direktvermarktung in the 2024 to 2026 window. Each one would be manageable in isolation. Stacked, they break the previous decade’s operating model.

Change 1: The 15-minute day-ahead market

• What happened - EPEX SPOT transitioned the day-ahead auction from 60-minute to 15-minute products on 30 September 2025^1,2. The PV bell curve and wind ramp can now be bid directly into day-ahead instead of being smoothed across hourly products.
• What it does to operations - Bid complexity multiplied by four. 96 quarter-hourly products per day per portfolio instead of 24 hourly ones. The work is no longer feasible by hand for any portfolio with meaningful volume.
• What it changes structurally - Forecast accuracy at quarter-hour granularity matters in ways it never did before. Imbalance positions that used to net out within an hour now stand exposed for individual quarters.

Change 2: Record intraday volatility

• What happened - Intraday price swings of tens of euros per MWh within a few hours have become routine in early 2026². The intraday market is now where forecast errors meet reality.
• What it does to operations - The trading desk cannot manually monitor and respond to every product cycle. The agent layer becomes operationally necessary, not optional.
• What it changes structurally - The competitive gap between direct marketers with algorithmic intraday capability and those without is now 5 to 15 percent of marketing revenue per year.

Change 3: Cannibalisation and negative prices

• What happened - When wind and solar all produce simultaneously, prices fall to zero or below. Price cannibalisation in Germany has worsened drastically⁴.
• What it does to operations - Curtailment decisions, negative-price hedging, and contract clauses for asset owners now require active management on every high-production day.
• What it changes structurally - Corporate PPA volumes in Germany dropped 84 percent year-on-year as the old simple pricing models broke⁴. The market is being redrawn around producers and off-takers that can quantify and hedge negative-price risk explicitly.

Change 4: AI demand reshaping the demand side

• What happened - AI-driven electricity demand made 2025 the second-highest year on record for corporate clean-energy PPAs globally, even as German volumes fell sharply³. Hyperscaler PPAs reshape the European market.
• What it does to operations - Mittelstand direct marketers compete against utilities with full quant desks for the same renewable output. Without an algorithmic layer, the Mittelstand asset value erodes.
• What it changes structurally - “Utilities hold an advantage because they have the expertise in the intraday trading and ancillary markets to optimise the value of a PPA”^3,5. The agent layer is how Mittelstand direct marketers close that gap.

Where the Margin Actually Lives

Direct-marketing margin in 2026 is built or destroyed in four operational decisions. Understanding where the margin actually lives is the prerequisite for putting AI in the right place.

1. Forecast accuracy

Every percentage point of forecast error on day-ahead translates directly into imbalance cost on delivery day. A 100 MW wind portfolio with 8 percent mean absolute forecast error pays 150,000 to 400,000 euros per year more in imbalance than the same portfolio at 7 percent. Forecast improvement compounds: better forecasts also mean more confident intraday positions, fewer hedging penalties, better PPA pricing.

2. Intraday capture

The intraday market exists because the day-ahead schedule never matches actual production exactly. Closing the gap costs money when you sell low into a falling market or buy high in a spike. Algorithmic intraday capture turns this cost into a partial profit centre - sometimes by anticipating ramp errors before the market prices them, sometimes by holding flexibility for the right minute.

3. Imbalance (Ausgleichsenergie) cost

Whatever does not match between the day-ahead schedule and actual delivery becomes imbalance, settled at the TSO’s imbalance price. Imbalance is asymmetric: when the system is short and you are short, the penalty is severe. Forecasts, intraday trades, and BESS dispatch all aim at the same goal of minimising imbalance.

4. Multi-market and BESS stacking

A given megawatt of flexibility can earn on day-ahead, intraday, balancing energy (Regelreserve), capacity, and via grid-fee optimisation. Most Mittelstand operators monetise one or two of those. The agent layer makes stacking operationally feasible, often lifting BESS NPV 20 to 50 percent compared to single-revenue operation.

Seven High-ROI AI Use Cases

The use cases below are ranked by typical Mittelstand direct-marketing ROI within the first 12 months. Each integrates with existing ETRM, SCADA, weather provider, and market data feeds - none requires replacing core systems.

Use case 1: Production forecasting (wind, solar, hybrid)

• What the agent does - Ensembles weather provider data (DWD, ECMWF, MetOffice, commercial nowcasting), SCADA history, site-specific microclimate adjustments, and seasonal patterns. Outputs probabilistic 15-minute forecasts up to 14 days ahead with explicit uncertainty bands.
• Where it sits - Between weather data and ETRM. Feeds forecasts into the bidding agent and into PPA risk models.
• What it removes - Reliance on a single forecast provider that misses local conditions. Forecast errors that propagate into both day-ahead bids and intraday corrections.
• Typical ROI - 0.5 to 2 percentage points forecast accuracy improvement vs single-vendor baseline. On 100 MW wind, 150,000 to 800,000 euros per year in imbalance savings.
• Time to ROI - 3 to 6 months.

Use case 2: Day-ahead 15-minute bidding

• What the agent does - Constructs the 96 quarter-hourly bid curves daily, balancing forecast confidence, market view, portfolio constraints, BESS flexibility, and PPA obligations. Submits bids into the EPEX SPOT day-ahead auction before gate closure.
• Where it sits - Above ETRM. Reads forecasts and contract obligations, proposes bid curves, the trader approves or auto-submits within limits.
• What it removes - The manual construction of 96 bid points per portfolio per day. The conservative bias that under-monetises tight forecast hours.
• Typical ROI - 2 to 5 percent uplift on day-ahead revenue capture vs rule-based bidding.
• Time to ROI - 6 to 9 months.

Use case 3: Intraday continuous trading agent

• What the agent does - Monitors intraday continuous order book on EPEX SPOT 24/7 within defined risk limits. Detects mispricing, adjusts the position as production forecast updates intraday, takes profit on volatility spikes when permissible, closes imbalance exposure ahead of delivery.
• Where it sits - Between ETRM and the EPEX SPOT API. Operates within explicit risk limits (max position, max single-order MW, daily P&L drawdown, frequency caps).
• What it removes - The 24/7 manual monitoring problem. The latency between a production update and a market action.
• Typical ROI - 3 to 8 percent uplift on intraday revenue capture. Imbalance reduction of 20 to 40 percent.
• Time to ROI - 6 to 12 months.

Use case 4: Imbalance (Ausgleichsenergie) minimisation

• What the agent does - Looks ahead to delivery, reads updated forecasts and intraday positions, decides whether to close imbalance in intraday or hold against the imbalance price. Optimises across the portfolio rather than per asset.
• Where it sits - On top of intraday agent and BESS dispatcher. Consumes the imbalance-price forecast model.
• What it removes - Reflexive intraday closing of positions that would have netted out favourably against imbalance. The single biggest variable cost in direct marketing.
• Typical ROI - 15 to 35 percent reduction in net imbalance cost.
• Time to ROI - 6 to 12 months.

Use case 5: BESS multi-revenue dispatch

• What the agent does - For each MWh of BESS capacity, optimises across energy arbitrage (day-ahead, intraday), balancing energy (PRL, SRL, MRL), capacity, and grid-fee optimisation (§14a EnWG, atypische Netznutzung). Reasons about state-of-charge constraints, cycle limits, warranty implications.
• Where it sits - Above BESS EMS and ETRM. Reads market data, contracts, technical limits.
• What it removes - Single-strategy operation that leaves 20 to 50 percent of NPV on the table.
• Typical ROI - 20 to 50 percent NPV uplift on new co-located BESS projects. Often the difference between a financeable and non-financeable case.
• Time to ROI - 9 to 18 months (rises with deployment scale).

Use case 6: PPA pricing, risk modelling, negotiation support

• What the agent does - Reads historical price data, forward curves, weather scenarios, plant generation profile, contract clauses, off-taker credit. Runs Monte Carlo on negative price risk, profile risk, balancing risk, credit risk. Outputs a defendable price range with explicit assumptions.
• Where it sits - Above ETRM, weather, market data, contract database. Feeds origination, structuring, and pricing decisions.
• What it removes - The guessing on three of four PPA risk dimensions that crashes deals when one breaks. The reason German corporate PPA volumes fell 84 percent in 2025⁴.
• Typical ROI - 50 to 200 basis points on PPA pricing accuracy. Higher origination volume because more deals get priced credibly.
• Time to ROI - 12 to 18 months.

Use case 7: Multi-market and ancillary services stacking

• What the agent does - Coordinates portfolio participation across day-ahead, intraday, balancing energy markets (PRL, SRL, MRL), and capacity products. Reasons about cross-market opportunity cost and time-locking.
• Where it sits - Top-of-stack coordinator above the individual bidding agents.
• What it removes - Sub-optimal market allocation when reserves are sold cheap and the energy market spike is missed, or vice versa.
• Typical ROI - 3 to 7 percent total revenue uplift across the stack.
• Time to ROI - 12 to 18 months.

Build vs Buy: ETRM, Specialist, Custom

Every Mittelstand direct marketer chooses between three paths. The right answer depends on the use case mix, in-house quant capability, and how differentiating the portfolio strategy is.

Path 1: ETRM-embedded AI (ProCom, eZ-Trade, in-house ETRMs)

• What you get - Algorithmic bidding and forecast integration inside the ETRM. Workflow remains in one system. Vendor follows the regulatory and market changes.
• Where it fits - Direct marketers happy with a single vendor stack, with workflows largely inside the ETRM. Smaller portfolios where the specialist edge does not justify a separate platform.
• Where it does not fit - Multi-revenue BESS stacking. PPA risk modelling that crosses to weather and credit data. Portfolio strategies that are part of the competitive edge. ETRMs follow the market, they rarely lead.
• Typical cost - Bundled with ETRM licence plus AI module fee, usage-based market data costs separate.

Path 2: Specialist trading-AI vendors (Suena, Entrix, Inavitas, enercast, ProCom-AI)

• What you get - Best-in-class capability in a narrow domain. Forecasting (enercast, energy & meteo systems, EWC). Intraday algos (Suena, Entrix). BESS optimisation (Inavitas, kiwigrid).
• Where it fits - When you have one specific use case (typically forecasting or intraday) and the workflow is largely self-contained.
• Where it does not fit - Cross-use-case coordination (forecast plus intraday plus BESS plus PPA). Differentiating logic the vendor will not customise.
• Typical cost - 50,000 to 500,000 euros per year per specialist platform, plus implementation, plus integration back to ETRM and SCADA.

Path 3: Custom AI agents on top of your stack

• What you get - An agent layer purpose-built for your portfolio, sitting above ETRM, SCADA, weather, market data, BESS EMS. Cross-use-case, cross-system, portable across ETRM vendors, and the agent itself is yours rather than rented.
• Where it fits - When use cases cross forecasting plus bidding plus BESS plus PPA. When portfolio strategy is part of your edge. When you run mixed assets and contract structures.
• Where it does not fit - When the single use case is genuinely contained and a specialist vendor already does it well. When the portfolio is too small to justify a custom build.
• Typical cost - 80,000 to 250,000 euros per use case for the build, plus 5,000 to 15,000 euros per month per active agent, plus market data and LLM inference at cents per task.

The Honest 3-Year Cost Comparison

Take a Mittelstand direct marketer with a 400 MW mixed portfolio (220 MW wind, 150 MW solar, 30 MW co-located BESS), 1.4 TWh annual production, REMIT-reporting obligation, and an ambition to add corporate PPAs. Three years on three paths.

What is not in the table

• Asset value uplift - Better-managed assets are worth more in secondary-market transactions. Hard to attribute precisely, real in M&A pricing.
• Counterparty risk - PPAs priced with risk-aware models reduce off-taker default exposure. Showed up sharply when several offtakers wobbled in 2024 and 2025.
• Trader retention - Senior traders prefer working with agent support to grinding manual bid construction in 96 quarter-hourly products.
• Strategic optionality - With the agent layer in place, adding ancillary markets or new asset classes is a 2-month project, not a year-long initiative.

SCADA + ETRM + Agent Architecture

The architecture that survives the first 18 months in a direct-marketing operation is intentional and audit-aligned. The ETRM stays as the system of record for positions, schedules, settlements, and regulatory reporting. The agent reads SCADA, weather, market data, contracts, and writes orders into the ETRM through proper APIs. The agent never bypasses the ETRM for the regulatory trail.

The four-layer trading stack

• Layer 1: External data - Weather providers (DWD, ECMWF, MetOffice, commercial nowcasting), EPEX SPOT market data, TenneT/Amprion/50Hertz/TransnetBW grid data, BNetzA REMIT publications, news feeds. The volatile edge of trading.
• Layer 2: AI agents - Production forecasting, day-ahead bidding, intraday continuous trading, imbalance optimisation, BESS dispatch, PPA pricing. The new reasoning layer.
• Layer 3: ETRM as system of record - ProCom, eZ-Trade, in-house systems. Positions, schedules, settlement, REMIT reporting, MaBiS, P&L. The official record.
• Layer 4: Adjacent systems - SCADA (asset-level production, availability, curtailment), BESS EMS, billing, finance, contract database (PPAs, EEG, capacity contracts).

Where the agent reads and writes

BNetzA, REMIT, MAR, and EU AI Act

Direct-marketing AI sits squarely inside three overlapping regulatory regimes. Designed deliberately, the agent layer makes all three easier, not harder.

REMIT (Regulation on wholesale Energy Market Integrity and Transparency)

• Trade reporting - All wholesale energy trades must be reported to ACER through the ETRM. The agent generates trades that the ETRM reports - the regulatory trail is intact.
• Inside information - The agent must not trade on undisclosed inside information (planned curtailment, asset outage, regulatory pre-knowledge). The agent platform documents what data was available at the time of every decision - cleaner than human trader records.
• Market manipulation prevention - The agent must not engage in spoofing, layering, wash trades, or cross-product manipulation. Risk limits, order frequency caps, and audit logs prevent the patterns and document compliance.

MAR (Market Abuse Regulation)

• Algorithm documentation - The agent’s strategy, parameters, and risk limits are documented and version-controlled. A regulator inspecting the desk gets a clearer answer than from a human trader with notebook annotations.
• Surveillance and kill switch - The agent monitors its own behaviour and stops when limits trigger. A kill switch above the agent is mandatory and documented.
• Order audit - Every order is logged with the model inputs that produced it. Reconstructing why any specific trade happened takes seconds, not days.

EU AI Act

• Risk classification - Most direct-marketing AI sits in limited-risk territory. Article 12 logging duties are met natively by the agent platform.
• Documentation - Risk classification, data sources, decision logic, human override path. Standard deliverable.
• Human oversight - The agent operates within explicit human-set limits. A trader can override at any time. Required by both EU AI Act and MAR.

BNetzA, EEG, MaBiS

• Marktprämie - Direct marketing under §20 EEG remains a regulated revenue stream. The agent must respect Marktprämie eligibility rules per asset.
• MaBiS imbalance - Reporting to the TSOs for imbalance settlement runs through the ETRM. The agent feeds correct positions; the ETRM reports.
• BNetzA reporting - Aggregator and direct-marketer reporting obligations remain in the ETRM. The agent does not replace this, it feeds cleaner data into it.

How Superkind Fits

Superkind builds custom AI agents that sit on top of existing direct-marketing stacks - ProCom, eZ-Trade, in-house ETRMs - and the SCADA, BESS EMS, weather, market data, and contract systems alongside them. We do not replace the ETRM. We build the reasoning layer that does what the ETRM was never built to do across the full direct-marketing workflow.

Core capabilities for direct-marketing environments

• ETRM integration - ProCom, eZ-Trade, BelVis Trading, in-house systems. Stable interfaces and database connectors for positions, scheduling, settlement.
• EPEX SPOT integration - Day-ahead bidding API, intraday continuous order book, M7 trading interface.
• SCADA and asset data - Wind, PV, BESS, biogas, hydro assets. Real-time production, availability, curtailment status.
• Weather provider ensembling - DWD, ECMWF, MetOffice, commercial vendors (enercast, energy & meteo systems, EWC, ConWX). Multi-vendor probabilistic forecasts.
• BESS EMS integration - Tesla, Sungrow, BYD, Fluence, Kraftblock, Wärtsilä. State of charge, cycle limits, warranty constraints.
• Forecasting agents - 15-minute probabilistic forecasts for wind, solar, hybrid, with explicit uncertainty bands and scenario fan charts.
• Day-ahead bidding agents - 96 quarter-hourly bid curve construction with portfolio constraints, BESS flexibility, and PPA obligations.
• Intraday continuous agents - 24/7 algorithmic trading with hard risk limits and human kill-switch.
• Imbalance optimisation - Cross-portfolio imbalance management with imbalance-price forecast model.
• BESS multi-revenue dispatch - Energy + balancing + capacity + grid-fee optimisation in one reasoning model.
• PPA pricing and risk - Monte Carlo on negative price risk, profile risk, balancing risk, credit risk.
• REMIT, MAR, EU AI Act audit trail - Every decision logged with inputs and model state. One log, three regulatory purposes.
• EU deployment and DSGVO alignment - Agents run on EU cloud or your own infrastructure. Data does not leave your defined perimeter.
• 8 to 12 weeks to first production deployment - From process assessment to live operation on a focused first use case.

The 90-Day Plan

This plan covers selecting the right first use case, validating data, deploying the agent in limited scope, and reaching first measurable value. Use it to align trading leadership, asset management, IT, and risk.

Weeks 1 to 3: Use case selection and data audit

• Quantify the three biggest revenue leaks - Forecast MAE per asset class, imbalance cost per MWh, intraday capture vs benchmark, BESS revenue per MW, PPA price gap vs reference. Numbers, not opinions.
• Pick three candidate use cases from the seven - Score each on revenue impact, deployment complexity, data readiness, regulatory readiness.
• Pick one use case for the 90-day pilot - Bias toward production forecasting (foundational) or intraday continuous trading (largest single revenue lever after forecasts).
• Audit the data the use case needs - SCADA history (minimum 12 months), weather vendor data, market data history, contract terms. Identify gaps.
• Confirm API access - ETRM, SCADA, weather, EPEX SPOT, BESS EMS. Document integration plan.
• Brief risk and compliance - REMIT, MAR, BNetzA implications mapped before code is written.

Weeks 4 to 8: Build and test

• Detailed process map - Inputs, outputs, decision points, risk limits, escalation triggers, kill-switch logic.
• Agent build against the process map - Model architecture, ETRM integration, risk limits, human-in-the-loop checkpoints, audit logging.
• Backtest against historical market data - 12 to 24 months of actual market data and plant generation. Compare to actual trader desk decisions. Document gaps and edge cases.
• Validate exception handling - What does the agent do when a weather provider fails, when an asset is curtailed, when intraday is illiquid, when an order is rejected.
• Confirm REMIT and MAR audit logging - Every decision traceable to inputs and model state.
• Train the desk - Traders learn the agent’s strategy, limits, and override paths.

Weeks 9 to 12: Production and learning

• Deploy to limited scope - One zone, one asset class, or a fraction of portfolio volume. Parallel running with the existing desk.
• Daily review cadence - Every decision the agent made, every escalation, every limit hit. What worked, what did not.
• Measure against the baseline - Imbalance per MWh, intraday capture, BESS revenue, forecast MAE. If the numbers do not move, diagnose before scaling.
• Expand once metrics validate - Two to three weeks of stable operation at limited scope before scaling to full volume.
• Document lessons for the next use case - The second agent will be twice as fast to deploy.

Related Articles

• ERP or AI Agent: Where the Boundary Runs Through Mittelstand Operations in 2026
• Predictive Maintenance for Hidden Champions: From Sensor Data to an Autonomous Maintenance Agent
• AI Agent ROI: The KPI Framework That Convinces CFOs in 90 Days
• Sovereign AI for the Mittelstand: Why German SMEs Are Choosing Custom Agents Over US-Hosted Tools
• AI Agent Security for the Mittelstand: How German SMEs Lock Down Autonomous AI
• Multi-Agent Systems for the Mittelstand: When One Agent Is Not Enough
• Standard Software or an AI Agent: How the Mittelstand Should Choose Where Software Budget Goes in 2026

Frequently Asked Questions

Henri Jung

Co-founder at Superkind

Co-founder of Superkind, where he helps SMEs and enterprises deploy custom AI agents that actually fit how their teams work. Henri is passionate about closing the gap between what AI can do and the value it creates in real companies. He believes the Mittelstand has everything it needs to lead in AI - it just needs the right approach.