AI for Solar Energy Companies: Proposals and After-Sales Support

AI for Solar Energy Companies: Proposals and After-Sales Support

A solar energy company uses AI most safely when it applies it in three places: qualifying the leads that arrive on WhatsApp, organizing the data that feeds the proposal, and following up with the customer after installation. System sizing, equipment selection and technical responsibility stay with the engineer or the company's design software. AI for solar energy companies reduces repetitive work and lost information, but it does not decide what only a qualified professional can sign.

Key takeaways

• AI helps most around the technical project: lead triage, document collection, proposal summaries and after-sales follow-up.
• System sizing, capacity, estimated generation and financial return must come from design software and the responsible engineer, never from model-generated text.
• Utility bills, addresses and personal documents require a clear purpose, restricted access and attention to Brazil's data protection law (LGPD).
• The best first step is a pilot on a single workflow, with every decision logged and an easy handoff to a person.
• Post-installation is usually the most neglected area, and it is where organized follow-up does the most to protect the company's reputation.

Where does AI help, and where should it stay out?

A solar integrator's process has stages that are very different from one another. Some involve conversation and organizing information. Others are engineering. Generative AI works well in the first group and should stay far from deciding the second.

Where it usually helps:

• Talking to the lead and collecting the basic information the sales team needs.
• Reading a utility bill sent as a photo or PDF and turning its fields into structured data, with human review.
• Summarizing the customer's history for the salesperson before a call or site visit.
• Building the list of pending documents for the utility approval stage and reminding the customer to send them.
• Drafting the proposal text from numbers already calculated by the design software, without changing those numbers.
• Triaging questions and tickets after installation and routing anything technical to the right person.

Where it should not decide on its own:

• Sizing the system, choosing modules and inverters, or estimating generation.
• Stating a payback period, bill savings or financing terms.
• Interpreting utility or regulatory rules as if they were technical advice.
• Diagnosing an inverter or installation fault without a technician.

This separation is the foundation for the rest of the article. Companies that hold to it avoid the most expensive mistake in the sector: an attractive number in a proposal that nobody can support later.

How do you qualify solar leads without turning the chat into a form?

Solar leads usually arrive with little information and high expectations. They ask whether it is worth it, how much it costs and how long it takes to pay off. The AI agent should not answer those questions with promises. It should explain what the company does, collect what is needed for a serious quote and set the next step with a person.

A reasonable qualification conversation collects the following, in a natural tone and a little at a time:

• Property type: home, business, industrial site or farm.
• Whether the person owns or rents the property, and whether someone else is involved in the decision.
• City and electricity utility, when the customer knows them.
• A photo or PDF of the most recent utility bill, with a clear notice about how the document will be used.
• Interest in paying upfront, financing, or not yet decided.
• Urgency: ready to decide now, still researching, or comparing proposals.

The agent's wording should avoid an interrogation tone. Instead of dumping every question at once, it alternates a short explanation with a single question. For a deeper look at this design, see lead qualification on WhatsApp with AI, without the form.

The expected result is not a magic lead score. It is an organized record the salesperson can read in a few seconds: what the customer said, what is missing and the suggested next action. If the lead asks about price or savings, the correct answer is that it depends on an analysis of the bill and the site, done by the technical team.

How do you organize proposal data without inventing the sizing?

For many companies the bottleneck is not calculation but gathering scattered information: a bill photo on WhatsApp, measurements in a voice note, an address typed in three places, a salesperson's notes in a notebook. AI is useful for consolidating this material into a single record.

A safe workflow has four layers:

• Extraction: the model reads the utility bill and fills in fields such as account holder, utility, consumption class, consumption history and site address.
• Validation: a person checks the extracted fields, especially those that feed the calculation. A doubtful field is flagged, never filled in by guesswork.
• Calculation: the design software, or the engineer, produces the sizing, estimated generation and other technical numbers.
• Drafting: the model receives the finished numbers and assembles the proposal text, the executive summary and the list of assumptions, without recalculating anything.

The golden rule is that the model is never the source of a technical number. It carries and organizes. To reinforce this, give the agent read-only tools that query the company's systems and forbid it from writing values that did not come from there. Output validation techniques, like those described in guardrails for AI agents, help block a proposal that cites a value missing from the official calculation.

Quick checklist before sending any proposal drafted with AI support:

• Do all technical numbers match the design software report?
• Are the assumptions (shading, orientation, tariff used) written out?
• Is there any savings or return promise that is not documented?
• Has a responsible engineer reviewed the document?
• Does the customer's personal data appear only where it is needed?

If the company's CRM already stores the leads, the summary and the validated fields can be written back to it. The article on integrating AI with your CRM via API, webhooks or MCP compares the integration paths and the risks of each.

What can AI do after installation?

Once the system is running, many companies treat the customer as a closed sale. This is the period when questions turn into complaints: the first bill that did not look the way the customer expected, the inverter app showing an alert, a question about cleaning the modules, a request for a warranty document.

A tightly limited after-sales agent can:

• Answer common questions based on company-approved material, such as the user manual, warranty policy and maintenance guidance.
• Explain in plain language how to read the manufacturer's monitoring app, without interpreting specific readings as a diagnosis.
• Open a ticket with the right data: address, equipment model, problem description, photo and the time the customer noticed it.
• Classify the ticket by type (usage question, suspected fault, billing issue, document request) and route it to the correct team.
• Send agreed reminders, such as a preventive check or cleaning, as long as the customer has authorized that contact.
• Log everything in the customer's history so the technician arrives knowing what has already been said.

What the agent does not do: say the system is producing less than expected, promise a repair, approve a warranty claim, or tell the customer to touch a breaker or inverter. When there is an electrical risk, the response should be direct: stop using whatever is safe to stop, call the technical team, and hand off to a person immediately.

For outbound contact such as maintenance reminders, you need to understand consent and message template rules. The article on WhatsApp outbound messages, templates and opt-in covers that point. Companies that also work on construction projects can draw on AI for construction and real estate developers: buyer support, which tackles a similar problem: a customer who is happy with the delivery but faces a long stretch of questions and adjustments.

What does this look like in practice? An illustrative scenario

Illustrative scenario: a small integrator receives a lead on WhatsApp. The person says they want to lower the household electricity bill and sends a photo of the bill.

The agent greets them, explains that the company runs a technical analysis before any proposal, and asks whether they own the property. It reads the bill, extracts the fields and confirms with the customer only what was illegible. It asks how the customer is thinking of paying and whether anyone else is involved in the decision. It mentions neither savings nor payback time.

The salesperson receives a record with the extracted data, the fields that need checking and a suggestion to schedule a conversation. They check the bill, bring in the responsible engineer, and the design software produces the sizing. The proposal text is drafted with the model's help from those results, and the engineer reviews it before it goes out.

After installation, the customer writes that the app shows an alert. The agent asks for a screenshot, records the equipment model, opens the ticket as a suspected fault and says a technician will follow up. It does not try to explain the alert. The technician receives the full ticket and decides on a visit.

In this scenario, AI handled the conversation, the extraction and the record keeping. Every technical and commercial decision stayed with people.

What data, LGPD and security precautions apply here?

The solar business handles personal data frequently: name, tax ID (CPF), address, consumption history, photos of the property and sometimes financing documents. That calls for explicit decisions before any agent goes live.

• Purpose: define why each piece of data is collected. A utility bill serves the quote, not another use without a proper legal basis.
• Minimization: ask only for what the stage requires. Financing documents can stay out of the agent and follow their own channel.
• Access: the agent reads what it needs and writes only to defined fields. Avoid broad permissions to the CRM or the project system.
• Retention: decide how long photos and bills are kept and who can delete them.
• Contract: check the AI vendor's position as a data processor and what it does with the data.
• Untrusted input: the content of a PDF or a message may contain malicious instructions. The agent should treat third-party text as data, not as an order.

For the broader picture, the article on LGPD and artificial intelligence for companies organizes the most common obligations. For risks specific to applications built on language models, such as prompt injection and excessive permissions, the OWASP list is a useful primary reference, and the NIST risk management framework helps structure the review.

How to start: map, limit, pilot and hand over

The method we apply in projects like this follows four moves, without assuming any particular result.

• Map: describe the real path from lead to after-sales, including who does what, in which tool and where information gets lost. Record a simple baseline, such as the time between first contact and the salesperson's reply.
• Limit: choose one workflow, define which data the agent sees, which actions it can take and when it hands off to a person. Anything involving technical calculation, price or warranty stays out.
• Pilot: run that workflow on real cases, with every conversation and decision logged, and a success metric agreed before you start.
• Hand over: deliver documentation, access and a review routine, then decide with the team whether to expand, adjust or stop.

A good first workflow for an integrator is usually lead triage with utility bill reading, because it has a clear input and an output that is easy to check. Post-installation comes next, once there is a reliable history of customers and equipment.

If the company has not yet mapped its process, start there with how to map business processes before applying AI. Automating a confused workflow only speeds up the confusion.

Frequently asked questions

Can AI size the solar system for the customer? It should not. Sizing depends on site data, technical assumptions and professional responsibility, and it must come from the design software and the responsible engineer. AI can organize the input information and draft the text from results that have already been validated.

Can I use a customer's utility bill in an AI agent? Yes, as long as there is a clear purpose, a notice to the customer, restricted access and a retention policy. The bill carries a name, an address and a consumption pattern, which are personal data. Also check in the contract how the model vendor handles that data.

Can the AI agent promise savings or a payback period? No. Those numbers depend on an individual analysis and on assumptions the team needs to document. The agent should explain that the estimate comes after the technical assessment and hand the lead to a person.

Where should a small company start? Start with a single workflow, such as lead triage with bill reading and a record for the salesperson, with human review and logged conversations. Define a simple metric beforehand and a criterion for deciding whether the pilot continues. Only then move on to proposals and after-sales.

Sources

• NIST: AI Risk Management Framework
• OWASP: Top 10 for LLM Applications
• ANPD: Autoridade Nacional de Proteção de Dados
• Anthropic: Building effective agents