Superintelligence for Solar Energy Companies: Proposals to Service

Superintelligence for Solar Energy Companies: Proposals to Service

Superintelligence for solar energy companies, today, is not a system that knows everything about photovoltaic projects. It is a well-designed AI layer that connects proposal, site survey, installation and after-sales, with organized data, clear limits and human review, and that gets better as the models improve. What lasts is the structure of your company, not the model of the moment.

Key takeaways

• The term superintelligence describes a direction, not a product. What a solar integrator can run today are agents and workflows that support decisions, not ones that decide on their own.
• The layer must be separated from the model: data, rules, tools and logs stay with the company, and the model can be swapped.
• Proposal, site survey, installation and after-sales produce data that connects. When each stage lives in its own corner, AI only automates the rework.
• System sizing, technical reports and engineering responsibility stay with the qualified professional. AI organizes, summarizes and drafts.
• The safe path is to map, limit, pilot one real workflow and only then expand, with a record of decisions and a metric agreed in advance.

What does superintelligence for solar energy companies mean in practice?

In public debate, superintelligence is a system that would outperform humans at nearly every cognitive task. That does not exist in any solar integrator, and anyone selling it that way is selling expectation. To understand the gap between the concept and what a company actually operates, read the difference between AGI, ASI and AI agents and the superintelligence myths that get in the way of decisions.

The useful way to use the term, for a solar company, is architectural. You build a layer of intelligence around your business: it knows your products, your commercial rules, your installation history and your procedures. It uses the best available model to read electricity bills, write proposals, organize site survey photos and answer customers. When a better model appears, the layer takes advantage of it without being rebuilt.

This changes the decision question. Instead of asking which AI to buy, ask which parts of the process must stay under your control so that any model can be used safely. The general architecture behind this idea is covered in the company superintelligence layer and its five building blocks.

How do you structure a solar proposal with AI without promising generation?

The proposal is where the gain is largest and the risk is highest. The gain is turning a bill, a few photos and a WhatsApp conversation into an organized draft. The risk is the AI stating savings, timelines or energy output that the company cannot stand behind.

A division that works is to separate what is calculation, what is text and what is commitment:

• Calculation: sizing, generation estimates and payback come from your design tool or your engineer, not from a language model inventing numbers.
• Text: the AI writes the explanation in plain language, adapts it to the customer profile and keeps the company's visual style and tone.
• Commitment: deadlines, commercial terms, warranties and any promise go through a fixed rule or human approval.

Input checklist for the AI to prepare a proposal draft:

• A legible electricity bill, with connection type and utility information.
• Roof type and condition, or installation site, with photos.
• Consumption profile and plans for change, such as new equipment.
• The integrator's current catalog of equipment and services.
• Approved commercial rules, with no exceptions improvised in the conversation.

If any item is missing, the correct behavior for the agent is to ask for the data or hand off to a salesperson, not to fill the gap with a guess. For the commercial side of the conversation, the article on AI for solar energy companies goes deeper into proposals and after-sales support.

Site survey: what AI organizes and what the technician decides

Site surveys tend to lose information. The technician takes photos, jots a note on the phone, sends a voice message, and later someone has to reconstruct the picture for the design. Here AI is good at organizing and weak as a judge.

What it can do safely:

• Receive photos and audio and build a standardized report with fixed fields, such as roof, structure, electrical panel and service entrance.
• Flag what is missing, for example a panel photo with no legible labeling.
• Highlight inconsistencies between what the customer said and what was recorded.
• Transcribe and summarize the technician's audio, respecting notice and legal basis when there is a recording (see recording meetings with AI and LGPD).

What must stay with the qualified professional: assessing structural and electrical feasibility, deciding what is safe, and signing off on anything that requires technical responsibility. The AI report is an input, and the technician confirms or corrects it. Correcting should be easy, because every correction becomes data to improve the workflow.

Installation: how can AI help coordinate without getting in the crew's way?

On the installation side, the value is in cutting the back and forth of information between office, field crew and customer. Nobody wants an AI agent messaging the installer while he is on the roof. What you want is a clear record of what was done.

A simple structure is a checklist per stage, filled in on the phone by the crew, with mandatory photos at agreed points. The AI reads these records, flags open items before closing the job and drafts the summary for the customer. It also helps keep the customer informed about progress over WhatsApp, always using approved text for sensitive situations such as delays or scope changes.

For messages the company starts, you need to respect the channel rules. Review WhatsApp outbound messages, templates and opt-in before automating installation notices, because sending without permission raises the risk of the number being blocked.

Another concern is personal data. Installation photos show the customer's home, sometimes people and documents. Define who has access, how long the files are kept and where the data goes. Brazil's data protection authority, the ANPD, is the reference for guidance on data protection.

After-sales: where the layer learns the most

After-sales is where an integrator earns its reputation and where AI has plenty of material: inverter alerts, questions about the electricity bill, cleaning requests, warranty claims, referrals. It is also where a poorly limited agent does the most damage, because the customer has already paid and expects a correct answer.

Workflows that usually make sense:

• Ticket triage: the agent identifies whether it is a usage question, a system alert, a billing issue or a suspected defect, and opens the right record.
• Reading monitoring alerts from the manufacturer, when the integrator has authorized access, with a summary for the technical team.
• Explaining the post-installation bill in plain language, with human handoff when there is a discrepancy.
• Preventive maintenance reminders and review requests, within opt-in rules.

What should not be automated: deciding warranty coverage, promising a technical visit without checking the schedule, or stating the cause of a failure without a diagnosis. In those cases the agent collects the information, explains the next step and hands the conversation to a person with full context. When the conversation depends on the channel, the service window rule matters; see how the agent should act in the WhatsApp API 24 hour window.

If a customer challenges a decision made with the help of automation, they need a path to human review. The article on LGPD automated decision-making and the right to review helps you think this through.

How does the layer improve with new models without being rebuilt?

Models change often. If your operation depends on the specific behavior of one model, every update becomes a risk. The way out is to separate what is stable from what is replaceable.

What should be yours and stable:

• Knowledge base: catalog, policies, site survey and installation procedures, approved texts.
• Tools with minimum permissions: check the schedule, read installation records, create a ticket. Each tool with well-defined inputs and outputs.
• Decision log: what the agent read, what it answered, what it escalated.
• Test set: real, anonymized cases with the expected answer.

What can be swapped: the model that reads, reasons and writes. With the test set, you switch the model, run the cases and compare before going to production. Anthropic publishes guidance on evaluating AI agents and on building effective agents that favors simple, composable workflows. For risk, the NIST AI Risk Management Framework offers a useful vocabulary for governance.

Illustrative scenario: a small integrator uses an agent to read the bill and draft the proposal. When a new model becomes available, it does not switch straight away. It runs the set of past bills, including hard cases like a blurry bill and an unusual connection type, and checks whether the new model makes fewer mistakes, asks for data when it should and respects the catalog. Only then does it change the configuration, keeping the old model as a fallback for a while. The layer got better without anyone rewriting the operation.

If the company has data restrictions and wants to evaluate a local model for some tasks, running Ollama local models in a company shows how to do it without improvising. If data is sent to providers abroad, check international data transfer under LGPD when using AI providers.

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

The method we apply to this kind of project follows four moves, with no promise of results before measuring.

• Map: describe the current process from first contact to after-sales, identify where there is rework and where information gets lost, and set a baseline for time and errors, measured by the company itself.
• Limit: define which data the AI can access, which actions it can perform, what requires human approval and what it never does, such as promising a deadline, promising energy output or deciding a warranty claim.
• Pilot: choose a single real workflow, for example reading bills and drafting proposals, or triaging after-sales tickets. Run it with full logging and a success metric agreed beforehand.
• Hand over: give the team the documentation, tests and logs, and decide together whether the pilot will be expanded, adjusted or closed.

This order avoids the most common mistake, which is starting with the most ambitious agent and discovering later that the underlying data was disorganized. Stopping is also a valid outcome: if the workflow does not improve the process, it is better to know early.

Ongoing operation, with monitoring and maintenance of the agents, can be handled by your own team or by partners such as Be Human. What matters is that someone is responsible for reviewing the logs, updating the knowledge base and running the tests when the model changes.

Frequently asked questions

Does superintelligence for solar energy companies exist today? Not in the sense of a system that outperforms humans at everything. What exists are AI models that are capable at specific tasks, such as reading documents, drafting text and organizing information, which can be arranged into a support layer for proposals, site surveys and after-sales. Calling that superintelligence is a way to name the direction, and it is wise not to promise more than the system delivers.

Can AI size the photovoltaic system on its own? It should not. Sizing and technical feasibility need a proper design tool and a qualified professional who takes responsibility. The AI can prepare the data, draft the explanation and point out what is missing, but numbers and technical reports go through human validation.

Where should a small integrator start? With a single workflow that is low risk and uses data you already have, such as organizing site survey reports or triaging after-sales tickets. Measure the process first, run a pilot with everything logged and decide based on what the team observed. Only then connect other stages.

What should you do when the AI model changes or is updated? Run the test set with real, anonymized cases and compare the new model's results with the current one before switching. Keep the knowledge base, tools and logs under the company's control, so that changing the model is a configuration change and not a new project.

Sources

• NIST: AI Risk Management Framework
• Anthropic: Building effective agents
• Anthropic: Demystifying evals for AI agents
• ANPD: Autoridade Nacional de Proteção de Dados