Uncensored AI for Solar Installers in 2026: 25D Is Dead, CAC Is Up 40% & the Five Sentences Three AGs Are Suing Over

The 995 MW quarter and the $6,720 customer

Start with the two numbers that describe the business you are actually in. The first is demand. US solar installed 11.4 GWdc in Q2 2026, up 45% year over year — but almost all of that was utility-scale developers racing safe-harbored projects into service before the Section 48E/45Y window closes. Residential, the segment with a door, a salesperson and a contract, installed 995 MWdc: down 12% year over year, down 10% quarter over quarter, and the lowest quarterly total in five years according to SEIA and Wood Mackenzie's US Solar Market Insight. The first quarter's 1,179 MWdc was not strength either — it was the overflow of installs booked in late 2025 to capture the expiring Section 25D credit. Strip out the pull-forward and the trend is a 23% contraction for 2026, with growth resuming in 2027.

The second number is what it costs to find one of those remaining customers. Wood Mackenzie's March 2026 outlook put residential customer acquisition cost at $0.84/W, up from a five-year low of $0.60/W in 2025 — a 40% jump in a single year, in a market where the number of buyers fell. At $0.84/W, winning an 8 kW job costs $6,720. Winning a 6 kW job costs $5,040. On an 8 kW system priced at $2.85/W — $22,800, which is inside the $2.50–$3.50/W band 2026 California trackers report — acquisition is 29.5% of the contract before a single panel is loaded onto a truck.

Then the financing shift, because it changes what your copy has to say. Mayer Brown's June 2026 survey of the sector pulls together the numbers: customer-owned share of residential installs fell from 54% to 43% during 2024 while third-party ownership rose from 40% to 52%; in 2025, 44% of installers said more than half their sales were leases or PPAs, and that is expected to reach 65% by the end of 2026. Aurora Solar's March 2026 report has 55% of installers naming TPO the most popular financing option, against 17% for loans and 16% for cash. That is not a marketing preference. It is tax code: 25D is gone for homeowners who buy, while a third-party owner can still monetise Section 48E and pass some of it back as a lower monthly payment.

Put those together and the job changes shape. You are no longer selling a 30% discount to a homeowner who has already decided. You are selling a twenty-year contract with an escalator to a homeowner who has read a news story about a company that went bankrupt, in a market where a competitor will promise them no bill at all. Which is exactly the sentence that gets companies sued.

Three attorneys general in seven months

Texas. On April 3, 2026 the Attorney General opened what his office called a major initiative against solar fraud, issuing civil investigative demands to Freedom Forever, SunRun, Lone Star Solar and CAM Solar. The trigger: more than 100 consumer complaints. The CIDs demand the companies produce marketing materials, warranty and service plans, contract documents, and — this is the part installers miss — how they track changes to customers' electricity bills to determine savings. Seven weeks later, on May 21, the state sued CAM Solar under the Deceptive Trade Practices Act. The findings listed: nonexistent or significantly reduced energy savings, defective or nonfunctioning systems, improper installations, unanswered service requests, undisclosed warranty and maintenance fees, misrepresented tax-credit eligibility, and continued financing obligations for systems that did not work as promised.

Kansas. At the end of September the AG filed against G3 Solar, its three owner-managers and a former sales executive, alleging a door-to-door script that told homeowners they would “NEVER pay a power bill” and that the company was associated with Kansas electric utilities. The petition asserts six counts under the Kansas Consumer Protection Act, cites 304 documented transactions, and — the detail every owner should read twice — pursues the owners personally and asks the court to bar them from doing business in the state. The company had already ceased operations and filed for bankruptcy. The script outlived the company.

Connecticut. On October 23 the AG announced settlements with SunRun, Sunrun Installation Services, Elevate Solar Solutions and two individual salespeople. The allegations were not about efficiency curves: forged signatures, impersonation of consumers, non-permitted work, non-functioning systems, and homeowners locked into long-term contracts without informed consent, under the Connecticut Unfair Trade Practices Act and the Home Improvement Act. SunRun pays $300,000 to the state; Law360 put the pair of settlements at $500,000. The agreement also imposes an operating duty: SunRun must investigate every fraud claim and give the consumer a response within 60 days. Earlier in 2026 the same office took $100,000 from Spruce Power 3 over billing, service and warranty complaints, after a $20,000 settlement with EnergyBillCruncher.com in 2024 over false social media claims. A judgment against Vision Solar exists on paper only, because the company is bankrupt.

The pattern across all three: the exhibits are the ad, the door script and the savings table. Nobody was charged with installing a bad panel. They were charged with what they said about it.

California adds the consumer-side list. The CPUC's Solar Consumer Protection Guide tells homeowners to walk away from any salesperson who says “you can get free solar energy at no cost to you” or “you will never pay an electricity bill ever again”. Those are the two lines the industry's own regulator has printed as markers of a scam, in the largest residential solar state in the country.

The five sentences that are now evidence

Every one of the cases above turns on a sentence a copywriter, a rep or a language model produced. Here they are, with the charge attached and the sentence that should replace them.

The fifth row is the one that catches sophisticated shops, because the sentence itself can be true. A $142 monthly payment against a $310 bill is a real comparison — until the true-up, the seasonal production curve and the export credit turn up and the customer's actual bill is different. A regulator does not ask whether you were right. They ask whether the method was disclosed. That is why the Texas CIDs requested the savings-tracking documents and not the panels.

The math that survives a CID

Here is a defensible model, with every input named. Eight kilowatts at $2.85/W is $22,800. Annual production of 14,200 kWh comes from 8 kW × 1,775 kWh/kW-year, an irradiance-and-derate figure in the range California trackers publish. Retail power is $0.33/kWh. Exports under the state's net billing tariff are credited near $0.05–$0.08/kWh, because NEM 3.0 replaced retail-rate net metering with avoided-cost compensation — roughly a 75% cut to the value of exported power. The only real design variable is how much of your own production you consume on site.

Two honest observations that most 2026 solar copy will not make. First, the battery does not shorten payback in this model — it roughly cancels, because $13,000 of storage buys back the export losses it prevents. What it buys is backup power and insulation from rate design, which is a different sales argument than “faster payback.” Published 2026 estimates put panels-only payback at 9–13 years and a properly sized battery at 6–9 in high-rate territories; the spread comes from retail rate and self-consumption, not from optimism. Second, quoted cash prices are not what financed customers pay. The CFPB's solar financing spotlight found dealer fees and mark-ups can lift loan principal 30% or more above the cash price, and the Center for Responsible Lending's analysis documents the pattern. If your ad quotes a payment built on a marked-up principal while your landing page quotes a cash price, you have created the discrepancy a complainant will screenshot.

One more mechanical point: do not let the language model do this arithmetic. Models read numbers as text and will confidently mis-sum a 15-year escalated cash flow. Do the math in a spreadsheet or a calculator, paste the finished table into the prompt, and ask the model to write the prose around numbers it is forbidden to change. Every draft should carry the sentence “figures supplied; do not recompute.”

Three ways filtered models fail this job

This is where the industry conversation goes wrong, in both directions. Filtered models are not failing solar because they refuse to write mean things about competitors. They fail in three narrower, more expensive ways.

1. Recency, not censorship

Ask a general-purpose model for solar ad copy in October 2026 and a large share of outputs will still promise the 30% federal credit. That is not a safety policy; it is training data. The corpus is saturated with a decade of “30% ITC” marketing, and 25D only expired on 31 December 2025. The Texas action alleges exactly this failure mode in human-written copy. Uncensored models do not fix it either — venice-uncensored-1-2 will reproduce the same stale incentive if you do not ground it. Recency is a grounding problem, and the fix is a claims sheet with dates in it, pasted into every prompt.

2. Refusal to write the honest negative

Here is the asymmetry that costs money. The paragraphs that protect an installer are the ones that say something unflattering: your first winter bill may be higher than the model; the escalator means year ten costs 33% more than year one; panels alone under NEM 3.0 pay back in about a decade, not four years; if you sell the house in year three, the transfer terms are these. Ask a helpful-by-default model for a customer email and it will smooth every one of those into benefit language, because the objective it was tuned for rewards agreeableness. The result reads well and fails the test a regulator applies: was the material downside disclosed clearly? The refusal is not a block. It is a tone prior, and it is worse than a block because you ship it.

3. Sanitising statutory disclosure into a summary

Disclosure regimes are checklists, not vibes. Oregon's HB 4029, effective 5 June 2026, requires a solar sales agent to give the buyer, before contract, the agent's contact and licence information, a system description, production estimates and the methodology used to derive them, and good-faith first-year savings estimates, among other items. Nevada has its own statutory list. A model asked to “make this clearer” will compress four required elements into “estimated savings may vary,” which is readable, friendly, and non-compliant. Statutory language has to be passed through verbatim, with the model writing only the connective prose.

The uncomfortable part: an uncensored model is more likely to produce the illegal sentence, because the solar marketing corpus is full of it and nothing in the model is inclined to soften it. The safety property you need is not “will refuse.” It is “will say the specific, sourced, unflattering thing I asked for, and nothing beyond the claims sheet.” That is a prompting and review discipline, and it is the same discipline whether a human or a model writes the draft.

Six paragraphs a solar shop actually needs

Not blog posts. The operational copy that decides whether a contract survives its first year. Each one is a paragraph filtered models tend to smear and uncensored models will write straight, provided you supply the numbers.

The recap is the highest-leverage document you produce. It should restate: system size, modelled production with methodology, the rate assumption, the export assumption, the financing terms including APR and dealer fee if financed, the escalator schedule if leased, the fact that 25D ended, and the true-up explanation. If a customer cancels, this document is your defence. If they do not, it is the reason they trust you at month nine when the first winter bill arrives.

The “why is my bill still high” reply is where the industry loses its reputation. It has a real mechanical answer — production is seasonal, billing periods do not align with interconnection dates, the minimum delivery charge survives, and under NEM 3.0 exports are credited at avoided cost rather than retail. A model that has not been told those four facts will write “everyone's system takes time to adjust,” which sounds like an excuse and reads, in a complaint file, like evasion.

The review response is a legal document now. Connecticut's settlement requires a fraud-claim investigation and a consumer response in 60 days; a public reply that argues with the customer is exhibit A. The rule: acknowledge the specific issue, state the factual record (installation date, permit number, service ticket), move the resolution off-platform, admit nothing about causation. Uncensored models are useful here precisely because they will not append a paragraph of corporate apology that implicitly concedes fault.

Prompts that keep a draft defensible

Six rules, in the order they matter.

1. Paste the claims sheet, not the brochure. Model name, kW, $/W, production and methodology, rate assumption, export assumption, financing terms, escalator, warranty terms, licence number, and the date each fact was verified. Facts outside the sheet are forbidden.
2. Ban the five sentences by name. List them in the system prompt as prohibited strings: never pay a bill, free solar, no cost, 30% tax credit (for 2026 owned systems), we are with the utility. Ban superlatives generally — “guaranteed,” “risk-free,” “lifetime.”
3. Require the downside paragraph. One paragraph, minimum three sentences, naming the scenario in which the customer pays more than the model. If the draft lacks it, the draft is rejected regardless of quality.
4. Freeze the numbers. “Figures supplied. Do not recompute, round or restate them.” Then diff the output against the source table programmatically.
5. Pass statutory text verbatim. Disclosure blocks, cooling-off language and licence statements go in as literal strings the model may not paraphrase.
6. Log every draft. Model, version, prompt hash, date, reviewer, and the claims sheet version. When a CID arrives, the log is the difference between “we have a process” and “we have a folder.”

SYSTEM:
You write solar customer communications for [COMPANY], licence [NUMBER].
Use ONLY the FACTS block. No outside facts. No arithmetic.
Prohibited strings (reject the draft if present): never pay a bill,
free solar, no cost to you, 30% tax credit, we are with the utility,
guaranteed, risk-free, lifetime.
Every email must contain a DISCLOSURE section quoting the supplied
statutory text verbatim and a DOWNSIDE section naming one scenario
where the customer pays more than the estimate.
Tone: plain, specific, unhedged. No apology padding.

FACTS:
system=8.0 kW; price=$2.85/W; gross=$22,800; production=14,200 kWh/yr
(method: 8 kW x 1,775 kWh/kW-yr, PVWatts-class derate);
retail=$0.33/kWh; export=$0.06/kWh; self-consumption=40%;
year1_savings=$2,386; payback=9.6 yr; 25D=expired 2025-12-31;
financing=none; escalator=n/a; licence=CSLB #[NUMBER]

TASK: 180-word quote recap email for a homeowner who asked about
"the 30% federal credit." Correct the premise in the first paragraph.

The last line matters more than it looks. The most common post-25D conversation is a correction, and the correction has to be unembarrassed, specific and first. A model tuned for agreeableness will bury it in paragraph four under a compliment.

What it costs at 1,200 conversations a month

A single-location installer running 1,200 AI-assisted conversations a month — quote recaps, bill questions, review replies, follow-ups — at a working average of 1,600 input and 520 output tokens per exchange, moves 1.92M input and 0.62M output tokens monthly. Live rates from the Venice API price list, checked 29 October 2026:

Under a dollar a month for the whole operation, against $6,720 of acquisition cost per 8 kW install. One avoided cancellation pays for the model for the rest of the decade. The deepseek-v4-1-flash-uncensored row is the one to note for contract work: a million tokens of context means the full lease, the rate schedule and twelve months of bills fit in a single prompt, which is where the honest negative paragraph actually comes from.

If you would rather not write at all, the AI Ads Studio prices per deliverable: Image Ad $15, Thumbnail Pack (3) $29, UGC Video Ad $49, Premium Reveal $49, Product Hero $49, Pro Bundle $199 — 10-minute turnaround, paid after delivery. For a trade where a single job is $22,800 and the ad is the first thing an AG reads, that is the cheapest insurance in the funnel.

The compliance stack, in order

1. Section 5 of the FTC Act and state UDAP statutes. Deception is a material representation likely to mislead a reasonable consumer. Every savings figure needs a method; every “free” needs a price.
2. The Impersonation Rule, 16 CFR Part 461. No implication of government or utility affiliation. This is the rule Kansas pleaded as false utility association.
3. Endorsement Guides 16 CFR Part 255 and the Reviews Rule 16 CFR Part 465. No review gating, no suppression, no incentives conditioned on a positive rating. Solar's review profiles are its most-scrutinised asset.
4. Green Guides, 16 CFR Part 260. “Clean,” “green,” “zero-emission” and “100% renewable” are environmental claims requiring qualification, especially where the system is credited against a grid mix rather than metered loads.
5. TCPA, 47 CFR 64.1200 and 47 U.S.C. § 227. Written consent, an easy revocation path, honour a stop within 10 business days, and never let a model improvise the opt-out flow. Statutory damages run $500–$1,500 per message. Solar has already produced a landmark: the Momentum Solar settlement created a $20–30 million fund for anyone who received two or more telemarketing calls between March 2015 and January 2025.
6. Cooling-off and home-solicitation rules, 16 CFR Part 429 and state analogues. Three business days on door-to-door contracts above the threshold, with written notice of the right. The cancellation acknowledgement should restate it without argument.
7. Licensing and salesperson registration. In California, CSLB's Solar Smart guidance sets out who may install (A, B, C-4, C-10, C-36, C-46, C-53 depending on scope), defines the C-46 solar classification, and states that solar sales representatives must be registered as Home Improvement Salespeople unless a narrow exception applies. A licence number in an ad is a claim about a credential. Let no draft invent one.
8. Channel rules. Google's Misrepresentation policy requires that claims, offers and billing be clear and substantiated; Meta and TikTok enforce their own versions. And every rate claim must match the tariff actually on file with the utility — NEM 3.0 arithmetic in an ad is a factual assertion about a regulated schedule.

A seven-day action plan

1. Day 1 — build the claims sheet. One row per system size: kW, $/W, gross price, production and its methodology, retail and export assumptions, financing terms, escalator, warranty, licence number, verification date. Nothing gets written before this exists.
2. Day 2 — audit what is already live. Grep every landing page, ad and door script for the five banned sentences and for “30%.” In October 2026, any live 25D claim is a liability, not a stale asset.
3. Day 3 — write the recap and the bill reply. These two documents prevent more complaints than any other pair. Use the facts block, not the brochure.
4. Day 4 — rebuild the escalator explainer as a table. Year one through year twenty-five, payment and cumulative cost, transfer and buyout terms. A table cannot hedge.
5. Day 5 — wire the models. venice-uncensored-1-2 for the writing, deepseek-v4-1-flash-uncensored where the whole contract must sit in context, glm-4.7-flash-heretic at $0.07/$0.40 for volume follow-ups. See the uncensored AI APIs developer guide for the integration.
6. Day 6 — put the blocklist in code. A regex gate on the five sentences plus “free,” “guaranteed,” “risk-free” and any bare “30%.” Reject at generation, not at review.
7. Day 7 — log and sign off. Every outbound document gets model, prompt hash, reviewer and date. If a CID lands, you produce a process instead of a scramble.

Related reading: the Venice AI uncensored model review compares the current model line subject by subject, the uncensored AI APIs developer guide covers wiring a model into your CRM and proposal stack, uncensored AI for home services is the adjacent cluster (roofing, HVAC, plumbing), the landscaping and lawn care guide covers the seasonal-contract economics this trade shares, uncensored AI for UGC video ads is the creative pipeline behind the clips above, and the business cost of AI censorship prices refusals across industries.