The task: shift consumption without disrupting work

AI POWER ApS is a Danish microbusiness classified by the European Digital Innovation Hubs Network as a firm with 1–9 employees. Its IntelliCharge.AI product helps schedule battery and electric-vehicle charging using electricity-price and consumption data. The platform operates through API integrations and does not require additional hardware at every customer's site. An important question for the company was the quality of consumption forecasts, particularly for a new user with no measurement history.

The case is useful not because every business urgently needs AI for charging. It demonstrates the order of decisions: first identify which demand can actually move in time, the business constraints and available data; next test a simple schedule; only then assess the value of a more complex forecast. If a service van must leave at six in the morning, a plan to charge it after breakfast is not a saving.

The company worked with Denmark's GC EDIH and Alexandra Institute under a test-before-invest approach. According to the published account, the partners compared potential AI/ML approaches to consumption forecasting, assessed benefits and limitations, and prepared an ideas catalogue for next steps. The case says AI POWER improved its consumption scheduling and its approach to onboarding users with no history. It does not publish a percentage reduction in bills or forecast error; assigning such figures to the company would be wrong.

What needs a model and what is ordinary automation

The task has two distinct parts. A forecast estimates how much energy a site will need and when. A scheduler decides which permitted intervals should be used to charge a battery or vehicle, considering price, the deadline, charging power and equipment constraints. The first component may use statistics or machine learning. The second is often better expressed as explicit rules and constraints so that an operator can explain the decision.

A minimal chain for a small fleet or workshop is:

  • Obtain the current tariff or other price that actually applies to the company's contract, rather than an abstract wholesale quote.
  • Capture hourly or more frequent consumption, state of charge, available power and vehicle-use schedules.
  • Calculate eligible intervals with a margin for departure deadlines and the site's power limit.
  • Compare the schedule with a simple baseline, such as fixed overnight charging or the operator's existing routine.
  • Propose the plan to a person, or send it to a controller only after testing emergency and manual-override scenarios.

A large language model need not control a relay here. Time, kWh and charging limits are structured data; time-series forecasting and optimisation are appropriate tools. An LLM might explain a proposed interval to an operator, but it should not be the only component enforcing maximum power, a departure deadline or battery protection. Local execution may help where data boundaries or availability require it, but the AI POWER case does not establish that the company used a local model.

Data to collect before a pilot

The primary prerequisite is genuine flexibility in the load. If all vehicles arrive late and depart early, there is little room to shift charging. If the contract does not differentiate price by time, a model might still help with peak demand or self-generated power, but no one should promise savings from "cheap hours" automatically. The business needs both meter data and the terms of its electricity supply.

Check five data groups: consumption history and completeness; applicable interval prices; battery state and charger constraints; a calendar of departures or production shifts; and external factors that actually affect demand, such as weather or on-site solar generation. For a new customer without history, a conservative baseline profile and manual power limits are useful. This cold-start problem is one AI POWER investigated; missing observations should not be replaced with a confident-looking forecast.

Integrations depend on the site: a meter or building-management system, charger APIs, the fleet calendar and possibly an ERP order schedule. Keep each decision with the tariff version, original forecast, constraints, actual schedule and reason for any manual change. Only then can the team tell whether a model helped or whether a different result came from weather and operating patterns. If an API fails or a price is stale, a safe baseline schedule should take over rather than replaying an old command indefinitely.

What the case confirms, and what it does not

The EDIH publication confirms the microbusiness's participation in testing, clarification of AI/ML options, and improvements to consumption scheduling and the approach to new users. It does not provide battery capacity, tariff structure, integration cost, test-set size or a monetary saving. This is a case about selecting an architectural direction, not a finished business plan with proven payback. The company's website confirms that it offers charging-management solutions, but the benefits described by the EDIH remain the project participants' claims.

Independent context comes from the International Energy Agency, which treats shifting or adjusting demand as a distinct resource in its analysis of demand flexibility. That supports the practical relevance of the problem category. It does not prove AI POWER's specific outcome or transfer European electricity-market pricing to a Russian company's contract.

Economics: measure the difference, not forecast elegance

For an investment decision, "money per controllable kWh after all costs" is more useful than model error alone. Consider an illustrative calculation with explicit assumptions: 6,000 kWh per month can genuinely be shifted; the average difference between applicable expensive and cheap intervals is RUB 1.2/kWh; charging constraints permit the shift without making vehicles unavailable. The maximum gross difference would then be RUB 7,200 a month. This is neither an AI POWER estimate nor a market tariff. Battery losses, connectivity, maintenance, changes in peak charges and the cost of scheduling errors still have to be deducted.

If integration and monitoring cost RUB 2,000 each month, the hypothetical remainder is RUB 5,200. A one-off RUB 150,000 project would have a simple payback of about 29 months, before risk and contract changes. Halve the flexible load and the conclusion changes sharply. These numbers only test the scale of the opportunity: a small site may find that a simple fixed schedule should come first, with a more advanced forecast justified only by variable demand and enough controllable energy.

A small next step

Choose one site and one controllable group, such as service vehicles. Collect actual charging, readiness deadlines and applicable prices for two to four weeks. Calculate an alternative schedule every day in shadow mode without controlling equipment. Compare it with the current routine and a simple rule under identical constraints. Count missed readiness deadlines, stale inputs and manual corrections separately. If a baseline rule captures almost all the benefit, there is no reason to deploy AI for its own sake. If a material gap persists, test a forecast and controller next.

The illustration was created by the editorial team using AI and schematically depicts charging; it is not a photograph of AI POWER.