Find my rebates →
Every federal, state, and utility rebate for your ZIP.
Policy · AI transparency

How we use AI.
And where we won't.

AI tools help us research policy, draft content, and run the Atlas concierge. Every published piece is human-reviewed. Rebate math is deterministic. Here's the full policy.

How does ElectrifyAtlas use AI?

AI assists with research drafts, featured-image generation, and the Atlas concierge chat. Every published post has human editorial review. Rebate math is deterministic (Rewiring America API + published caps), never AI inference. We don't train models on user data.

Where AI helps

Atlas concierge. The "Ask the atlas" chat runs on Amazon Bedrock (Anthropic Claude). It answers from a curated knowledge base we maintain in S3 — state HEEHRA status, federal policy, tool pointers, key facts. Atlas has explicit rules that forbid inventing dollar amounts and require citing program names. The full persona file is versioned and reviewed quarterly.

Content drafting. AI tools help produce first drafts of long-form journal posts. Drafts enter our standard 6-step editorial review: research → draft → fact-check → technical review → copy edit → datestamp. Nothing is published without a human editor signing off. See the editorial team page for the full process.

Featured image generation. Editorial-magazine images on blog posts, guides, and case studies are generated with Amazon Bedrock Stable Image Ultra, constrained by our locked aesthetic spec (Monocle/Kinfolk feel, architectural heat pump infrastructure, muted earth palette, no people). The hero + OpenGraph share image are also AI-generated. Product diagrams, schematics, and the brand logo are hand-authored.

Where we don't use AI

Rebate math. Live rebate amounts come from the Rewiring America Incentive API — a real database of state + utility programs, not a model. The stacking simulator and operating savings calculator use deterministic formulas with published HEEHRA caps, AMI thresholds, and standard COP values by climate zone. No language model estimates any dollar amount.

Program launch dates. State HEEHRA status changes in our rollout tracker are sourced from state energy office announcements, not AI summaries. Each event carries a source URL.

Installer routing. The three installers who quote your job come from a human-maintained, state-certified partner list. No AI scores or ranks installers. No installer pays to appear in routing.

Data + training policy

We do not send user ZIP codes, income data, or personal details to any AI training pipeline.

Atlas concierge conversations are sent to Anthropic via Amazon Bedrock for inference only. Per the Bedrock API terms, Anthropic does not train on Bedrock inference traffic. Your chat history is retained for the session only and is not stored beyond Bedrock's short-lived logs.

Contact info you submit to request installer quotes goes only to the specific installers you request — not to any AI, not to ad networks, not to any data broker.

Limits + corrections

AI-assisted drafts sometimes drift from primary sources. Atlas concierge sometimes misreads a question. The 6-step editorial review catches most errors, but not all.

If you spot an answer from Atlas, a blog post, or a case study that conflicts with a primary source, email [email protected] with the URL and the source showing the correct value. Verified corrections land within 48 hours with a timestamped note.

Full methodology + sourcing policy is on the methodology page. Editorial team + review process is on the editorial page.

AI transparency — frequently asked

Yes. The "Ask the atlas" chat runs on Amazon Bedrock (Anthropic Claude). It answers from a curated knowledge base we maintain in S3 — not from raw model knowledge. Atlas follows explicit rules: never invent dollar amounts, always cite program names, defer HVAC-install decisions to licensed installers, and confirm 25C repeal facts. See the persona file for the full rule set.
AI tools assist with research and first drafts on some content — never final publication. Every published piece goes through a 6-step human review: research, draft, fact-check against primary sources, technical review (HVAC engineering), copy edit, datestamp. Nothing ships without a human editorial pass. Every piece carries a named byline.
No. Live rebate math comes from the Rewiring America Incentive API, not a language model. The stacking simulator and operating savings calculator use deterministic formulas with published program caps and standard COP values, not AI inference.
Yes — the editorial-magazine featured images on blog posts and case studies are generated with Amazon Bedrock Stable Image Ultra, constrained by a locked aesthetic spec (Monocle/Kinfolk, architectural infrastructure, muted earth palette). Hero + OG images are AI-generated too. Product diagrams, schematics, and the logo are hand-authored.
Yes. The knowledge base is curated and the persona has strict fact-citation rules, but AI can still misread a question or produce an off-base answer. If Atlas says something that conflicts with the rebate finder result or a cited program URL, trust the primary source and flag the Atlas response to [email protected].
Yes — the concierge is fully optional. The rebate finder, calculator, stacking simulator, decision quiz, operating savings calculator, and equipment recommender all work without opening the chat. Your ZIP + income stays in your browser; none of it is sent to an AI model unless you chat.
No. We do not send user ZIP codes, income data, or personal details to any AI training pipeline. Atlas concierge conversations are sent to Anthropic via Bedrock for inference only; Anthropic does not train on Bedrock inference traffic per their API terms. Contact info you submit for installer quotes goes only to the specific installers you request.
See our process

Editorial + methodology.

Who writes, how we source, and where installers never touch the content.

Open methodology →Editorial team