Customer subsidy application agent (BEG / BAFA / KfW)

For every heat pump, solar-thermal system or insulation job, German installers are expected to tell the customer whether and how much funding is possible, and to deliver the confirmations the funder requires before and after the work. The rules change every few months, so this is done by hand with PDFs and experience. The agent checks a measure against a maintained rule base, pre-fills the data for the confirmation and application, and builds the proof package after installation. The customer or the expert still files and signs. It works today where the rules are kept in a versioned rule base; a model answering from memory is not enough.

Problem & value

• Today: an SHK business (heating, plumbing, sanitary) answers "what do I get?" from memory or from a funding calculator, then chases the customer for documents, types data into portals and, after installation, collects invoices, protocols and photos for the proof. Mistakes are expensive: an application filed after work started, a contract without funding condition, or a missing hydraulic balancing record can cost the customer the grant and the installer the relationship.
• AI value: (1) a fast, reproducible eligibility check against the rules in force on the contract date, (2) pre-filled application data from quote, heat-load calculation and data sheets, (3) a proof package that matches the funder's checklist, (4) deadline reminders.
• Who needs it (from the Mittelstand process analysis of 12 business types, 1 of which needs it): shk-handwerk: funding eligibility (~3 h), application data (~6 h), proofs after installation (~4 h), together about 13 h per month. The saving per firm is modest; the value is fewer rejected grants, faster sales conversations and a stronger offer when the customer sees the funded price.
• Why existing use cases do not cover it: KNOW-017 only finds funding programmes and LANG-006 writes funding proposals. Neither checks a concrete measure or builds the confirmation and proof data.
• Rules move quickly: in 2026 the KfW heating funding (458/459) was cut on 21 July 2026 (speed bonus 20 % to 16 %, efficiency bonus and emission surcharge removed, eligible costs for the first unit 28,000 EUR instead of 30,000 EUR) with further step-downs planned from February 2027 [2][3]. That is exactly why a dated rule base, not a model's memory, is the core of this use case.

Variants

• Heating replacement (KfW 458 / 459): owner-occupiers apply themselves as natural persons (458); companies, GbRs and other legal entities use 459. The agent prepares the BzA data and the customer's checklist; the installer or expert issues the BzA and later the BnD [1][2].
• Individual measures BAFA (BEG EM): building envelope, system technology, heating optimisation. With an energy-efficiency expert (TPB-ID at application, TPN-ID at proof of use), without one for heating optimisation [4][5].
• Business customers and landlords: more cases (several units, caps per unit, aid-law checks).
• Adjacent programmes: municipal and regional grants, feed-in or bonus programmes. Only as additional rule-base modules, with a lower confidence level.
• Private use: an owner-side assistant is a different product (see PERS family); here the installer prepares the case for the customer.

How it works

1. customer enquiry / quote (ERP, PDF) + building data
2. rule base lookup (programme, version valid on contract date, technical minimum requirements)
3. eligibility check per measure (building age, owner type, old system, product on list, hydraulic balancing)
4. funding estimate (base rate, bonuses, cost cap) + open questions for the customer
5. contract wording check (funding condition, application before start of work)
6. pre-fill BzA / TPB data from heat-load calculation and data sheets
7. expert or installer reviews and signs
8. customer files application (Meine KfW / BAFA portal) with checklist from the agent
9. after installation: collect invoices, commissioning protocol, balancing record, photos
10. draft BnD / TPN data and proof package
11. human release
12. customer uploads proof
13. deadline tracking (completion window, proof deadline, payout)

• Rule base: a versioned table or small knowledge base (programme, valid-from, rates, caps, documents, deadlines) fed from the funder's pages and guidelines. Changes are reviewed by a named person before they go live. Rates and caps are never taken from the model's own knowledge.
• LLM role: read quotes, data sheets and heat-load PDFs, map them to rule fields, explain the result in plain German and list what is missing. The calculation itself is deterministic code.
• Human-in-the-loop: the energy-efficiency expert or listed specialist company is legally the one who confirms (BzA/BnD/TPB/TPN). The agent only prepares. The customer files and bears the declaration to the funder.

Data preparation

Which data: quotes and orders, heat-load and hydraulic balancing calculations, product data sheets with the technical key figures, customer and building master data, and the current funding guidelines and fact sheets.

Where it lives: craft-business software or ERP (e.g. weclapp, Business Central, Hero, Craftnote, Lexware), a DMS or file share for data sheets and protocols, the site team's phones for photos, and the funders' websites for the rules.

Typical gaps: quotes without the key figures the funder needs (efficiency, refrigerant, sound level), data sheets as scans, protocols on paper, photos without a customer reference, and rules stored as a bookmark list nobody maintains.

Preparation steps: 1. Build the rule base for the two or three programmes you actually sell (KfW 458/459, BAFA BEG EM), with valid-from dates and source links (PREP-016 for written rules). 2. Standardise the quote template so product, key figures and hydraulic balancing appear as fields. 3. Collect 30 to 50 closed past cases with the result of the funder as a test set (PREP-018). 4. Fix a naming scheme for photos and protocols per order (PREP-006, PREP-023 for access).

Tools: spreadsheet or small database for the rule base, document parser for data sheets, the existing ERP for order data.

Rough share of effort: about 30 percent of the project, mostly the rule base and quote template. See GUIDE-003.

Implementation paths

Which path for whom

• SMB - Path B: no mature off-the-shelf product covers the whole chain for a small installer, so a light build on a cloud API with a self-maintained rule base is the pragmatic route. Path A (funding calculators, heating-planner and quoting software with funding modules) covers parts and is the first thing to check.
• Developer - Path C: vendors of trade software can embed the rule base and open models for document reading on their own hardware.

Path A - Buy (off-the-shelf)

• Funding calculators and advisor tools: web calculators and funding-database search exist (e.g. the federal funding database, KfW and BAFA calculators); they estimate, but do not read your quote or build the proof.
• Trade software: some craft-business and quoting solutions include funding hints or exports. Check whether they use a dated rule base and who updates it. Treat marketing claims with care.
• Choose A when: you only need a rate estimate for sales conversations.

Path B - Build on cloud APIs (recommended for SMBs)

• Models: a general LLM with structured output for document reading and plain-language explanations, EU region or EU provider preferred (customer data is personal data).
• Orchestration: n8n or Power Automate reading quotes from the ERP, calling a rule-base service, writing results back to the order and to a checklist for the customer.
• Rule engine: deterministic code or a decision table; the LLM only fills fields and explains.
• Choose B when: you have a standard quote process and a person who can own the rule base.

Path C - Self-host (open source)

• Document parsing: Docling or PaddleOCR for data sheets and protocols; a small open VLM (e.g. Qwen-VL class) served via vLLM or Ollama.
• Hardware: HW2 is enough for a few hundred documents a month.
• Choose C when: the vendor serves many installers, or data sovereignty is a hard requirement.

Path D - Custom training

Not worth it. The task is rule lookup and field mapping; fine-tuning does not replace a maintained rule base.

State of the art (2026)

• Maturity: early-adopter. LLM-based document reading and rule-based checks are solid; there is no widely adopted product that covers eligibility, application data and proof for installers end to end. Most real deployments are internal tools or add-ons to trade software.
• Official interfaces: funding is filed by the applicant in KfW's customer portal and in the BAFA portal. The agent cannot and should not file in the customer's name; it prepares and checks. Browser agents in these portals are fragile (virtual waiting room, identity checks) and conflict with the funders' identification rules.
• Rule drift is the main failure mode: rates, caps, bonuses and eligibility change by guideline and by date (e.g. new BEG EM guideline dated 17 July 2026, KfW changes from 21 July 2026), and secondary sources contradict each other [2][3][4]. Always pin the version valid on the contract or application date.
• Limits: LLMs misquote rates and mix up programmes (458 vs. 459 vs. BAFA); funding is subject to budget availability with no legal entitlement; the agent cannot judge building-specific questions (listed buildings, mixed use) or issue the expert confirmation.

Shared resources & combinations

• Same stack (llm-general + ocr-docparse + workflow-automation): AGENT-035 (job costing and invoice assembly) already collects invoices and protocols per order, and the proof package reuses that data. AGENT-004 and AGENT-027 cover order and quote intake in the same ERP.
• Upstream: KNOW-017 (finding programmes), LANG-006 (written proposals for non-standard programmes), CONV-001 (customer questions on the website: "what funding do I get?").
• Document checks: DOC-016 (rule checks on documents), DOC-003 (form processing), DOC-010 (contract wording, funding condition).
• Deadlines: AGENT-034 (deadline tracker) can carry the 36-month completion and proof deadlines.
• Workflow engine shared with AGENT-002.
• Biggest combined effect: one quote-intake pipeline that feeds quote, funding estimate and later proof package. Build the rule base once and use it in sales, planning and handover.

Effort, cost & KPIs

• Path A: days; subscription for a calculator or trade-software module.
• Path B: 2 to 4 months for a first version covering KfW 458/459 and BAFA BEG EM; the rule base and its maintenance are the main cost, API usage is small.
• Path C: 3 to 5 months incl. hosting; mainly useful for software vendors.
• Ongoing cost driver: rule-base maintenance after each guideline change (plan a few hours per quarter and a named owner).
• KPIs: share of applications approved without queries, time from quote to complete application data, time to a complete proof package, number of cases lost through formal errors (target: zero), share of quotes with a funding estimate.

People & change

Involve early: management (scope, liability), the sales and project leads as process owners, the energy-efficiency expert or the person who issues the BzA/BnD, and IT for the ERP interface. The DPO checks customer data, income proofs and processor contracts (AVV, Art. 28 GDPR); see GUIDE-005 and GUIDE-010.

Works council: no in a typical small craft business. If one exists, the workflow logs who prepared and released which case, so §87(1) no. 6 BetrVG may apply; a short agreement is enough (GUIDE-004).

AI literacy: Art. 4 AI Act training for sales and back office: what the agent checks, what it cannot (GUIDE-006).

Communication and pilot: pilot with one product line (heat pumps) and 10 to 20 real cases in parallel to the normal process; name one owner for the rule base (GUIDE-008, GUIDE-009).

Typical concerns: liability for wrong funding promises (solved by wording: estimate, not commitment), fear that customers handle the portal alone, extra work for the expert.

Humans in control: expert or specialist company signs the confirmations; the customer files and signs; the installer never promises a grant amount.

Risks, regulation & compliance

• Funding law: the grant is a subsidy from the federal budget with no legal entitlement; BEG EM is governed by its guideline, KfW 458/459 by the programme conditions. Application before start of work, contract with funding condition, minimum use period and proof deadlines are mandatory [1][2][4].
• Liability: a wrong funding statement can create claims against the installer. Use "estimate" wording, show the date of the rule version and store the check result with the quote.
• GEG and technical requirements: the building must comply with the Building Energy Act after the measure; technical minimum requirements must be confirmed by the qualified person [4].
• GDPR: customer name, address, building data and sometimes income proofs; legal basis contract (Art. 6(1)(b)) or consent for income documents; AVV with processors; data minimisation for income proofs (the funder needs them, the installer does not store them longer than necessary).
• EU AI Act: minimal risk; no Annex III use. AI literacy duty applies (Art. 4).
• Fraud and false declarations: the agent must not fill declarations that only the expert or the applicant may make. Keep an audit log of what was suggested and what was signed.
• Mitigations: versioned rule base with review, dated check results, human release on every confirmation, no automatic portal submission.

Getting started

1. List the programmes you actually sell and write each as a one-page rule sheet with valid-from date and a link to the funder's page.
2. Take 20 closed cases and replay the eligibility check by hand and with the prototype; compare with the funder's decision.
3. Add the key figures the funder needs to your quote template.
4. Build the proof checklist per programme and link it to the order folder.
5. Run the pilot for one product line; review rule changes monthly and after each guideline change.

Sources

1. KfW - Heizungsförderung für Privatpersonen Wohngebäude (458) - https://www.kfw.de/inlandsfoerderung/Privatpersonen/Bestehende-Immobilie/F%c3%b6rderprodukte/Heizungsf%c3%b6rderung-f%c3%bcr-Privatpersonen-Wohngeb%c3%a4ude-(458)/
2. KfW - Heizungsförderung für Unternehmen Wohngebäude (459) - https://www.kfw.de/inlandsfoerderung/Unternehmen/Wohnwirtschaft/F%C3%B6rderprodukte/Heizungsf%C3%B6rderung-f%C3%BCr-Unternehmen-Wohngeb%C3%A4ude-(459)/
3. reduco - KfW 458 Heizungsförderung 2026: new rates from 21 July (secondary source, not official) - https://reduco.ai/blog/kfw-458-heizungstausch
4. Förderdatenbank des Bundes - Bundesförderung für effiziente Gebäude, Einzelmaßnahmen (BEG EM) - https://www.foerderdatenbank.de/FDB/Content/DE/Foerderprogramm/Bund/BMWi/bundesfoerderung-effiziente-gebaeude-em.html
5. co2online - BAFA-Förderung: application steps, TPB/TPN (secondary source) - https://www.co2online.de/foerdermittel/bafa-foerderung/