How to automate energy sales outreach without sounding like a robot
Most "AI sales sequences" fail energy buyers in the first sentence: a hallucinated company description, the wrong supplier name, or a vague reference to "your industry" that proves no one actually read the account.
Done well, AI-written outreach beats a human SDR on volume and matches them on relevance. Here are the six rules we ship with DigitalBar by default — every one of them is the result of a thread that didn't land for a reason we could name.
1. Research before generation, always
Pull the supplier mix, contract end date, decision-maker title, and the last public PR moment before a single token of the email is generated. If the research step fails — no website, blocked scrape, no public signal — the email doesn't send. Better an empty queue than a wrong-name send.
In practice this looks like a two-step pipeline: the research agent runs first, populates a structured profile, and the generator only fires when that profile passes a minimum-completeness check. The cost of an extra research call is a fraction of the cost of a burned address.
2. Cap length at 100–150 words
Energy buyers don't read more. Anything longer reads as a template even when it isn't. We enforce this at the generator level — the model is instructed, the output is measured, and emails outside the band are regenerated. Three paragraphs maximum. One ask, not three.
3. Strict HTML, no Markdown
Markdown leaks into email clients as raw asterisks (**bold** displayed literally) or stripped formatting that flattens the message. HTML renders. This is a deliverability detail most teams discover only after a campaign goes out looking broken in Outlook. Generators should output HTML or refuse.
4. Block bracketed placeholders
"Hi [First Name]" in production email is the fastest way to lose a deal — and somehow it still ships. Generators should reject the send if a bracketed placeholder, an empty variable, or a {{merge_field}} token survives into the final body. We treat this as a hard block, not a warning.
5. Send from an authenticated mailbox
OAuth-based sending from the rep's own Google or Microsoft account outperforms any shared no-reply domain on both deliverability and reply rate, often by a factor of 3–5x. The reasons are stacked: SPF/DKIM/DMARC align naturally, the sender already has reputation with their existing contacts, replies thread cleanly into the rep's inbox, and recipients see a real human name in their list view.
A noreply@yourcompany.com address tells the recipient's mail client "this is a campaign, filter accordingly." A personal mailbox tells it "this is a person, deliver normally."
6. Reply-track at the thread level, not the pixel level
Tracking pixels and link wrapping hurt deliverability and most energy buyers' mail clients block them anyway. Worse, they signal to spam filters that this is a marketing email — which puts the legitimate ones in the same bucket as the bad ones. Thread-level reply detection (did they hit reply? did anyone in their org?) is the signal that matters and the only signal the recipient's mail provider treats as positive engagement.
What a well-run sequence looks like
Run all six rules and a single rep can comfortably manage a 1,500-prospect quarter without the outreach reading like spam. A typical cadence:
- Day 0: Researched, personalized opener. One paragraph of context that proves the account was read. One sentence of relevance. One soft ask.
- Day 4: Short follow-up with a specific resource (case study, RFP template) tied to what the research surfaced.
- Day 11: Re-ask, shorter still. Single question.
- Day 21: Breakup email. Clear, polite, ends the thread. Often pulls the reply that none of the previous three did.
The metric to watch
Reply rate, not open rate. Open rate is a vanity metric polluted by Apple Mail Privacy Protection and inflated by pre-fetchers. Replies — positive, negative, or polite no — are the only signal that the email reached a human and the human cared enough to respond.
A healthy reply rate for cold energy outbound, done well, is 8–14%. If you're below that, the problem is almost always research depth or length, not the model.
