Building Solar/ Energy Products: AI Agents, VPPs, and the Future Grid

I have had the same conversation few times this month: Once with a client that I am consulting asking about AI agents and Energy-as-a-Service through QuantiEdge, next with solar and battery operators sizing Virtual Power Plant bid, and with my team building LeadLink - our own agentic lead qualification platform for Solar/ Energy and home improvement field sales.

The question sounded different each time. The pattern was the same:

We are drowning in AI demos. What is actually working in solar/ energy/IoT and home improvement industry?

This is my answer - from a product leader’s seat, not from hype

1. Don't start with AI. Start with a broken workflow.

While working with solar and home improvement companies, I watched field sales burn hours on unqualified doors while homeowners had no clear communication about their qualification, appointments, and next steps. Installers lost money to bad leads and missed follow-ups.

Later, through QuantiEdge, I founded LeadLink to flip that funnel: QR -> AI-assisted answers -> qualified pipeline.

AI only sticks when it sits on a painful, repeatable workflow: not on the premise of AI transforming the complete process or industry.

I even evaluated third-party NL analytics tools for LeadLink, and then still built our own “Ask Your Data” stack (PostgreSQL + pgvector + ontology + Claude) because external tools still forced us to own vectorization, context/ ontology creation, and workflow integration. In my experience retaining the domain ownership beats the speed for an external vendor solution.

2. Energy AI agents create value on assets and delivery platforms create value on shipping.

When people say “AI agents in energy,” they mix two layers:

Layer	What it does	Examples
Asset / grid agents	Dispatch batteries, predict failures, optimize trading	VPP/DR, O&M triage, DERMS
Software delivery agents	Build, test, deploy, secure apps safely	Harness-style CI/CD agents

Software delievery agents are CI/ CD pipeline-native agents with RBAC, audit trails, and deploy workflows that keep execution inside your environment.

Energy companies are becoming software companies. Regulated ones need both layers.

3. EaaS O&M priorities

Solar/Battery operators own a 20–25 year residential fleet. They and not only selling panels and assets but owning outcomes: kWh, uptime, trust, and increasingly grid services.

My O&M priority stack:

1. Can we see the asset? (comms / telemetry)

2. Is it producing revenue? (performance vs guarantee)

3. Can we fix it without a truck roll?

4. Does the customer still trust us?

5. Can we protect the asset legally? (default, tamper, recovery)

Remote shutdown gets a lot of attenton. In practice, “shutdown” is three different things — safety, VPP curtailment, and contract default, and each has a different workflow.

The underrated pain? State management after upgrades, outages, or default: branding, production history, billing clocks, portal access. Spec those explicitly or your product will become a customer service nightmare.

4. With utilities, safety is table stakes. Audit is the deal.

Working with utilities on behalf of solar companies, I have seen two distinct workflows:

Customer workflow: inform before curtailment, shutdown, DR events, or service visits.

Utility / compliance workflow to prove following:

• Which metering path was used (legacy AMI vs AMI 2.0 vs OEM cloud vs gateway)

• What telemetry was collected

• At what frequency

• Who accessed it and when

• Whether the call was a read or a command

Every API touch belongs in an audit log. Utilities are buying controlled access with audit/ proof points

EaaS O&M at scale is half asset ops, half compliance ops.

5. Credible VPP capacity is not hardware capacity

When aggregators bid into programs like ELRP, DSGS, or DERP, they project capacity on basis of asset capacity. In my experience, the dominant factors are:

1. Online / availability rate (usuallly #1)

2. cCstomer backup reserve

3. AC / export / panel limits (not DC battery mentioned limit)

4. Temperature throttling (worst on peak heat days)

5. Historical event compliance

6. Enrollment filters

7. Geographic correlation (diversity credit or heat-storm penalty)

I tell bidding teams: bid on reachable, deliverable kW, not installed kW.

The hardes part of residential fleets: temperature and customer behavior often get worse exactly when the grid needs capacity most.

6. AIDC is a power problem dressed as a compute boom

U.S. AI data center growth is real, concentrated, and faster than interconnection can absorb. That’s creating new energy business models:

• Behind-the-meter / co-located generation (“speed-to-power”)

• Flexible AI factories that throttle non-latency work for grid value

• Large-load tariffs with collateral and minimum take

• Distributed inference at the edge (Span XFRA–class models)

• Grid capacity–finding software, not just more wires

Hyperscale still wins for training. Distributed / edge complements for inference.

7. “Electricity -> Compute” from rooftop solar: my feasibility test

The pitch: turn surplus green electricity into compute revenue.

My evaluation:

Dimension	Verdict
Inference with battery + orchestration	Feasible
Training on rooftop alone	Not feasible
Single-home DIY economics	Weak
Builder-led, third-party-owned fleet	Stronger
“24/7 green AI” marketing	Weak without matching + storage audit

Solar alone does not make a data center. Solar + battery + grid + orchestration can make a curtailable inference node.

The real business value today is speed-to-compute and monetizing underused electrical headroom - not a pure carbon narrative.

8. Plug-in balcony solar for renters: early, important, modest

Most of the U.S. still treats tiny plug-in PV like “real interconnection.” Utah flipped that first with statewide HB 340: portable devices up to 1.2 kW AC, no utility agreement if certified. Maine and Virginia are following. Dozens of states have bills.

For renters and LMI households, this is an access wedge: portable, low CapEx, no roof needed.

Limitations are honest: landlord/HOA rules still matter, savings are modest, fully certified retail kits are catching up to the law. It’s entry-level DER, not a replacement for rooftop TPO or community solar; and very different from XFRA-scale home compute.

The 6 prerequisites before “deep AI” in energy: Before I recommend agentic automation, I score the company on:

1. Data foundation: asset IDs, telemetry, lineage

2. Process standardization: SOPs agents can follow

3. Fault tolerance: reversible actions, kill switches

4. Clear ROI: one metric per workflow

5. Integration + governance: APIs, RBAC, audit

6. Org readiness: P&L owner, not just an innovation lab

   
   

Energy score today: ready for narrow, governed agents; not ready for autonomous utility everywhere.

Model capacity vs edge hardware is a permanent catch-up cycle, and yes, the saying is real: Models evolve in months. Edge hardware sits for years. That’s why edge AI is an orchestration business. Hyperscale can refresh GPUs. Homes cannot, hence distributed computing should design for inference, modularity, and workload routing, not “install once and hope the model stops improving.”

About the author : Sanjay Bhatia

I have built products inside the largest residential solar company and IoT companies in the U.S. I founded QuantiEdge to combine industry consulting with custom AI. I am shipping LeadLink because field sales still needs the lead qualification layer.

The next decade in energy won’t be won by teams with the flashiest model.It will be won by teams that can orchestrate energy from distributed sources, meet customers needs, build software to solve pain points, and provide audit trails.

I Advises on solar, VPP/DR, TPO/ABS, utility data, and agentic product systems via expert networks and direct consulting. I am open to consulting / fractional CPO conversations for solar/ energy, battery storage, VPP, and edge compute companies.

Reach out at sanjay.bhatia@quantiedge.com