Skip to main content
Academy · Module 9 of 9

Using AI without losing trust

AI makes drafting, research and triage fast and makes it cheap to send too much. What the evidence says about AI-written messages, where the rules touch, and how to keep a human in charge.

12 min read · Course overview

Objectives

In this module

  • Identify where AI helps in outreach and where it adds risk.
  • Explain the research on AI-written messages and perceived sincerity.
  • Know the regulatory touchpoints: the FTC, the EU AI Act and GDPR.
  • Recognize the six common AI failure modes.
  • Apply a review policy matched to risk.
9.1

Where AI helps

AI is now used throughout outreach: summarizing research on an account, drafting first emails and replies, varying copy, classifying replies, and in “AI SDR” products that run whole sequences. Gartner predicts AI agents will outnumber human sellers ten to one by 2028, yet fewer than 40% of sellers will say agents improved their productivity [1].

The receiving side uses AI too. Since January 2026 Gmail has offered AI summaries of email threads to all users [2]. Your email may well be read as a summary first.

9.2

AI-written messages and trust

A 2025 study of 1,100 professionals by Coman and Cardon found that only 40–52% saw a supervisor's message as sincere when it used heavy AI assistance, against 83% with light assistance; professionalism ratings stayed comparatively high [3]. A separate preregistered experiment found AI-involved messages were judged less trustworthy and authentic, and that disclosing AI use, while people said they wanted it, also lowered ratings [4].

Neither study tested cold sales email directly, but both point the same way. Recipients notice, and the cost falls on sincerity. Use AI for structure and research; keep the human parts (empathy, apology, judgment, relationship context) human.

9.3

Where the rules touch

The FTC has said plainly that “there is no AI exemption from the laws on the books” [5]. CAN-SPAM's honesty requirements apply to AI-generated subjects, headers and claims exactly as they do to yours [6].

In the EU, the AI Act's transparency obligations, including telling people they are interacting with an AI system, apply from August 2, 2026, with fines up to €15 million or 3% of turnover [7]. Whether an AI-drafted email a human reviews and sends is covered is unclear. An AI agent that converses with prospects on its own is the case to take to counsel. And feeding personal data into AI tools needs a lawful basis and proper processor terms under GDPR [8].

Anti-abuse groups are watching too: Spamhaus singles out AI-generated templates sent at scale [9], and M3AAWG targets messages that fake one-to-one correspondence [10].

9.4

Six failure modes

Each of these has the same fix: a human decides what goes out, and the system records why.

  • Invented facts: made-up customers, results, funding rounds or “I saw your post about…” references. Deceptive, and embarrassing when checked.
  • Uncanny personalization: shallow compliments that are obviously generated.
  • Volume creep: AI makes more email cheap, but complaint thresholds don't rise with it, and per-mailbox volume correlates with lower replies [11].
  • Unauthorized commitments: an autonomous reply that quotes a price, agrees terms or promises something no human approved.
  • Data protection: personal data sent to AI vendors without a lawful basis or processor terms.
  • Lost opt-outs: a “stop” in an AI-handled thread that never reaches global suppression.
9.5

A responsible-use policy

Write the policy down before you scale anything. A workable version:

  • Review by risk: first-email templates, and any reply that commits you to something, get human approval. Low-risk classification (out-of-office, unsubscribe) can be automated, with a log.
  • Ground every claim: each fact in a draft should trace to a stored source, such as a CRM field or URL. Reject drafts with claims you can't trace.
  • Cap volume at what your mailboxes have proven they can sustain, and keep spam rate under 0.1% [12].
  • Never fake a thread or a familiarity that doesn't exist.
  • Disclose AI where the law or the context calls for it, especially for conversational agents.
  • Log what was sent and why, and compare positive reply and complaint rates for AI-drafted and human-written emails, with the statistical caution from Module 8.
Key takeaways

Remember this

  • AI is good at research, structure and triage; humans own judgment and relationships.
  • Heavy AI assistance lowers perceived sincerity.
  • There's no AI exemption from CAN-SPAM, GDPR or the EU AI Act.
  • Ground every claim in a source; reject drafts that invent.
  • Review by risk, log everything, and keep volume within proven limits.
Exercise

Write your AI review policy

Half a page is enough.

  1. Step 1

    List every place AI touches your outreach today: research, drafting, replies, classification.

  2. Step 2

    Classify each as needs-approval or automate-with-log.

  3. Step 3

    Take three recent AI drafts and check every factual claim against a source. Note how many you couldn't trace.

  4. Step 4

    Confirm that an opt-out written inside an AI-handled thread reaches your global suppression list.

In AngelSend

Where this lives in the product

Review before sending

Turn on review before sending for a campaign and its AI-written drafts wait in the Review queue for a person to approve, edit or reject.

Assistant with your data

The AI assistant works from your own contacts, campaigns and conversations, so drafts start from stored facts rather than guesses.

Records of what went out

With the compliance archive on, a copy of every email sent from the project is kept before it leaves, so you can always see exactly what was sent.

Check yourself

Quiz

Pick an answer to see the explanation.

  1. Q1In Coman and Cardon's study, what happened when messages used heavy AI assistance?

  2. Q2An AI-drafted email claims the prospect “recently raised a Series B.” You can't find a source. What do you do?

  3. Q3Which AI use is most reasonable to automate without per-message approval?

Sources

  1. [1]Gartner Predicts AI Agents Will Outnumber Sellers 10 to 1 by 2028 — Gartner
  2. [2]Gmail is entering the Gemini era — Google
  3. [3]Is writing with AI at work undermining your credibility? — University of Florida
  4. [4]The AI penalty and disclosure paradox: trust, authenticity and knowledge — Macquarie University
  5. [5]FTC Targets Unfair or Deceptive AI Practices With Five New Enforcement Actions — Orrick
  6. [6]CAN-SPAM Act: A Compliance Guide for Business — U.S. FTC
  7. [7]EU AI Act: Transparency Obligations Take Effect 2 August 2026 — Cooley
  8. [8]ICO updates legitimate interests guidance after DUAA changes — DMA (UK)
  9. [9]Spamhaus' take on Cold Emailing… AKA spam — Spamhaus
  10. [10]M3AAWG Position on Cold Email — M3AAWG
  11. [11]Cold Email Benchmarks — Woodpecker
  12. [12]Email sender guidelines — Google

This course summarizes published rules, regulator guidance and third-party research as of October 2026. It is not legal advice; check with counsel before emailing into a jurisdiction. Vendor benchmark figures are each vendor's own data and definitions.

8. Measuring and testing honestlyFinish: course overview

Practice on
your own outreach.

Start with a $3.99 Starter Pack — no subscription required.

Get started freeBook a demo

Starter pack from $3.99 · CAN-SPAM compliant · Reputation protected