Planning a rooftop PV system or heat pump used to need a site visit, manual roof measuring and an hour or more in design software before a customer saw a price. Computer vision on aerial imagery and 3D building data now detects roof planes and obstacles, lays out modules, simulates yield and sizes battery and heat pump automatically, so installers and sales teams can send a first binding-looking offer within minutes of receiving an address, and reserve site visits for qualified leads.
Problem & value
- Today: installers (Handwerk) and energy sales teams quote from phone calls, photos and manual drawing in tools. Site visits for every lead are expensive and most leads never convert. Heat pump quotes also need heating-load and system-temperature estimates, which are often rough.
- AI value: faster first response (a key conversion driver), fewer pre-sales site visits, consistent layouts, better-qualified leads, scalable quoting in a market with a skilled-labour shortage. The sales-to-engineering handover becomes a structured design instead of a PDF and a phone call.
- Evidence: vendors such as Reonic, Aurora Solar, OpenSolar and Google's Solar API publish product claims about minute-level design and quoting [1][2][3][4]; independent, peer-reviewed conversion-rate gains are scarce, so measure on your own funnel.
- Demand drivers (EU/DE): the recast Energy Performance of Buildings Directive requires member states to phase in solar installations on buildings [6]; in Germany, BEG heat-pump subsidies via KfW programme 458 (base grant plus bonuses; conditions were reportedly tightened in July 2026, so check current rules) shape the offer and must be checked per date [5].
Variants
- Installer toolchain (SMB): roof-first quoting inside a CRM/design suite, then site-visit verification.
- Lead-gen/web configurator: homeowner enters address, sees roof overlay, yield and price range (private-use funnel; also useful for self-assessment by homeowners and housing cooperatives).
- Energy retailer/utility/Stadtwerke: campaign-scale roof screening of customer addresses and municipal solar cadastres (Solarpotenzialkataster).
- Heat-pump extension: building-age/envelope estimate from imagery plus public data, heat load estimate, radiator and system-temperature check, subsidy calculation.
- Commercial and multi-family roofs: flat-roof east-west layouts, ballast limits, tenant-electricity (Mieterstrom) models.
- Drone/3D variant: drone or phone/LiDAR scans for complex roofs; overlaps VIS-040 (scan-to-BIM) and VIS-003 (drone inspection).
How it works
- address
- geocode + imagery/3D fetch (orthophoto, DSM, LoD2, building footprint)
- roof-plane segmentation (pitch, azimuth, area) + obstacle detection (chimney, dormer, skylight, existing PV)
- shading model (DSM horizon, trees, neighbours)
- module layout optimiser (fire-safety setbacks, module dims, string/inverter constraints)
- yield simulation (irradiance data + PV model) + load profile
- self-consumption
- battery & heat-pump sizing (heat load, COP, tariff) + optimisation of component mix
- pricing from catalogue + subsidy and ROI
- quote document
- human review (installer checks roof, structure, electrical, meter cabinet)
- binding offer
- Vision: instance segmentation of roof faces and obstacles from orthophotos, often fused with DSM/point-cloud height to obtain pitch; Segment-Anything-style promptable models speed up annotation and fallback.
- Simulation/optimisation: PV models (pvlib, PVGIS irradiance data [7][8]) and a solver/MILP or heuristic for layout, battery size and heat-pump choice.
- LLM/VLM role: extract consumption from uploaded bills, draft the offer text, answer customer questions, explain the plan; numbers always come from the deterministic simulation, never from the LLM.
- Human-in-the-loop: installer confirms roof condition, statics, electrical installation and heating system on site or by customer photos before the binding quote; planner can drag and correct the auto-layout.
Data preparation
The data is mostly external and clean: aerial imagery, 3D building data (DSM, LoD2), irradiance data. The internal part is the product catalogue and price lists (ERP or wholesaler feeds, often Excel), subsidy rules, and the customer's consumption data (bills as PDF or photo, smart-meter data, CRM records).
Typical gaps: outdated imagery or missing coverage in some regions, catalogue items without technical attributes (module dimensions, inverter limits), price lists that live in several spreadsheets, bills that are scans, and CRM addresses that are incomplete or duplicated.
Preparation steps: 1. Consolidate the catalogue into one table with technical attributes and current prices (see PREP-011). 2. Connect the ERP or wholesaler feed instead of manual updates (PREP-005). 3. Clean CRM addresses and remove duplicates before geocoding. 4. Collect 20-50 completed projects with the real roof layout as a test set to compare against the automatic design (PREP-018). 5. Agree on a standard for bill uploads (photo or PDF) so extraction works.
Effort share: about 10-15 percent of the total effort. With path A it is mostly the catalogue import. Related: DOC-014 for bill extraction.
Implementation paths
Which path for whom
- smb (A): mature EU-hostable installer suites such as Reonic cover design, quoting and CRM links at per-seat prices; building your own rarely pays back.
- enterprise (B): large installers and energy retailers need their own pricing logic, catalogue and CRM integration, which a Solar API plus own configurator delivers.
- developer (B): combine Solar API or open LoD2/DSM data with PVGIS and pvlib to build a branded configurator.
- private (A): homeowners use free or low-cost online roof checks and installer tools; no own build is needed.
Path A — Buy (recommended)
- Installer suites: Reonic (DE, EU), Aurora Solar, OpenSolar, SMA Sunny Design and manufacturer configurators; EU-based ones: Reonic, 1KOMMA5 platform tools, Enpal in-house tooling (not generally for sale). Heat-pump planning: Viessmann/Vaillant/Stiebel Eltron partner tools and Heat-pump sizing modules in the above suites.
- Check: EU hosting, coverage and imagery freshness for your region, CRM/ERP and wholesaler catalogue connectors, subsidy and grid-registration templates, white-label web widget.
- Choose A when: you are an installer or retailer wanting quotes now; the build effort rarely pays back.
Path B — Build on cloud APIs
- Roof data: Google Maps Platform Solar API (building insights, roof segments, irradiance; verify coverage for your countries) [4]; commercial aerial providers (e.g. Vexcel, Bluesky Geospatial) or national/state open data (ALKIS footprints, LoD2, DOP orthophotos, DSM).
- Yield: PVGIS API (EU JRC, free) [7] and pvlib-based services.
- LLM/VLM: Claude, Gemini, GPT or Mistral (EU provider) for bill extraction, offer text and photo checks of roof/meter cabinet with structured outputs.
- Orchestration: n8n/Make or custom service, CRM integration (HubSpot, Salesforce, Pipedrive).
- Choose B when: you need a branded configurator or a very custom pricing logic.
Path C — Self-host (open source)
- Vision: SAM/SAM 2 or Mask2Former/YOLO segmentation models fine-tuned on roof datasets; OpenDroneMap for drone photogrammetry [9].
- Simulation: pvlib-python [8], PVGIS data, NREL SAM; optimiser via OR-Tools/HiGHS.
- Stack: PostGIS + QGIS/GDAL, open geodata (OpenStreetMap, state geoportals), self-hosted VLM (Qwen-VL class via vLLM) for document extraction.
- Hardware: HW1-HW2; segmentation runs on a single 16 GB GPU, batch screening of thousands of roofs per night.
Path D — Custom training
Worthwhile for a differentiated vendor or large retailer: train roof-plane/obstacle models on your own corrected layouts (every planner correction is a labelled sample), region-specific roof types and local imagery quality. Hundreds to thousands of corrected roofs are typical starting data.
State of the art (2026)
- Roof-plane detection from orthophotos plus DSM is productised; layouts and yield are generated in seconds. Foundation models (SAM family) reduced annotation effort and improved zero-shot obstacle handling.
- Integrated "PV + battery + heat pump + wallbox" quoting with tariff-aware optimisation is the competitive frontier (see OPT-032).
- Limits and failure modes: outdated imagery (new dormers, roof renewals, trees), low pitch accuracy on flat or complex roofs, unknown roof condition and structural capacity, shading underestimated by DSM resolution, heat load guessed without building data, subsidy rules changing. Auto-quotes should show ranges and a clear "subject to site check" note. Do not promise yields beyond the simulation uncertainty.
Shared resources & combinations
- Same stack: VIS-021 (satellite/aerial analysis), VIS-040 (scan-to-BIM), VIS-003 (drone inspection) share segmentation, geodata and photogrammetry pipelines.
- Downstream/adjacent: OPT-032 (home energy management once built), OPT-007 (building HVAC), PRED-007 (generation/load forecasting), PRED-031 (smart-meter disaggregation as load input), PRED-048 (renovation roadmaps, EPC/iSFP).
- Shared GPU/VLM server and geodata layer across these use cases.
Effort, cost & KPIs
- Path A: days to weeks, licence per seat or per quote.
- Path B: 2-4 months for a configurator with CRM and catalogue integration; imagery/API fees per request are a cost driver.
- Path C/D: 4-9 months, ML plus GIS plus energy-engineering skills.
- KPIs: time to first quote, lead-to-quote and quote-to-order conversion, site visits per order, deviation of simulated vs. measured yield, share of layouts accepted without manual change, roof-detection error rate, quote correction rate after site visit.
People & change
Involve management (budget, quote policy) and the sales lead early; they own the process. The installers and planners who review the auto-layout are the key end users: let them test two or three real projects in a pilot group and collect their corrections. IT connects CRM and catalogue; the data protection officer clears address and consumption data (GDPR applies, AVV with the platform vendor, a DPIA is usually not needed for plain quoting).
Works council: depends. The tool is mostly customer-facing, but if it logs which employee designed or changed which quote and how fast, it can monitor performance and §87(1) no. 6 BetrVG applies. Inform the council in good time under §90(1) no. 3 when work processes change.
AI literacy: support staff with a short training on what the simulation does, what it cannot see (roof condition, statics) and why the numbers come from the deterministic model, not the LLM (GUIDE-006).
Concerns: "the tool replaces planners" - position it as removing manual drawing; "wrong quotes" - the installer always confirms on site before a binding offer. Communicate openly, name a champion, and measure quote time and error rate (GUIDE-008, GUIDE-009, GUIDE-005).
Risks, regulation & compliance
- EU AI Act: minimal risk; no Annex III use. AI literacy (Art. 4) applies. Chatbot/offer assistants must disclose AI interaction (Art. 50) [10].
- GDPR: address, roof imagery and consumption data are personal data of the homeowner; need legal basis, privacy notice, DPA with API/imagery vendors, EU hosting and retention limits. Avoid drawing conclusions about persons from imagery; no scoring of residents.
- Consumer and sector rules: quotes must be clear about assumptions (UWG/consumer law), safety/structural responsibility stays with the installer, grid registration (VDE-AR-N 4105, grid-operator rules) and Marktstammdatenregister registration apply; subsidy rules (BEG/KfW 458, GEG) change, so version them in the pricing logic [5].
- Liability: over-optimistic yields and wrong subsidy statements are main exposure; log inputs, imagery date and simulation version per quote.
Getting started
- Take 50 past projects and compare auto-generated roofs/layouts with the installed designs.
- Test 2-3 Path A tools on your region's imagery quality and catalogue integration.
- Connect the load input (bill upload) and keep a fixed, versioned subsidy and price table.
- Define the human check before a binding offer (photo checklist or site visit).
- Track quote-to-order and correction rates monthly; feed planner corrections back.
Sources
- Reonic - AI-based planning and offers for PV, storage and heat pumps - https://www.reonic.com
- Aurora Solar - solar design and sales software - https://aurorasolar.com
- OpenSolar - design and proposal platform - https://www.opensolar.com
- Google Maps Platform - Solar API documentation - https://developers.google.com/maps/documentation/solar
- KfW - Heizungsfoerderung fuer Privatpersonen (458) - https://www.kfw.de/458
- Directive (EU) 2024/1275 (EPBD recast) - https://eur-lex.europa.eu/eli/dir/2024/1275/oj
- European Commission JRC - PVGIS - https://re.jrc.ec.europa.eu/pvg_tools/en/
- pvlib-python - https://github.com/pvlib/pvlib-python
- OpenDroneMap - https://github.com/OpenDroneMap/ODM
- Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj