Deliverability Debt vs. Three AI Initiatives
Fernwood Supply is a B2B distributor of industrial safety equipment with 70,000 HubSpot contacts. Over four years, staff who have since left built nine active workflows that overlap: a new contact can be enrolled in four at once and receive three emails in a day. Bounce rate is 5.8%, spam complaints 0.14%, and 28% of contacts have not opened an email in 12 months. Marketing has three people and no developer. The new VP wants three AI initiatives this quarter: (1) an AI chatbot on the pricing and product pages, (2) AI-generated personalized first lines in outbound and nurture emails, (3) 150 AI-written blog posts a month to "win organic." Fernwood sells PPE and fall-protection gear, where incorrect product specifications can create liability, and its top pages are product spec pages for respirators and harnesses.
1. Rank the three AI initiatives and justify the ranking against the account's current state.
Read the worked answer
The ranking question is a trap: nothing AI should ship before the email foundation is fixed. A 5.8% bounce rate and 0.14% spam complaints mean sending reputation is already damaged, and every AI initiative that increases send volume multiplies the damage. Ranking the AI ideas while ignoring that is optimizing the wrong layer.
Order of operations:
Fix workflows and suppression first (not AI, and non-negotiable). Nine overlapping workflows causing four simultaneous enrollments and three emails a day is a design failure that automation will scale, not solve. Consolidate enrollment criteria, add suppression lists, and cap frequency at one marketing email every three days. Sunset the 28% twelve-month non-engagers through a two-email re-permission, then suppress.
AI email personalization - modify and pilot. Lowest technical risk, but only pointed at engaged segments after suppression. With three people and no developer, a contained pilot in one workflow is realistic.
AI blog scaling - last, and not at 150/month. Content volume does not fix a reputation or a spec-accuracy problem.
Chatbot - conditional, and the highest liability of the three. See next answer.
2. State what you would approve, modify, or reject for the chatbot and the 150-post plan, naming the specific failure modes.
Read the worked answer
Chatbot on pricing and product pages - modify heavily, or defer. Failure modes are concrete in this business:
- Specification hallucination. A confident wrong answer about respirator cartridge compatibility or harness load ratings is a safety and liability event, not a CX annoyance.
- Pricing improvisation. An AI that quotes or negotiates price creates quote discrepancies that sales then has to unwind.
- Silent failure. Without transcript review, nobody learns it gave bad advice.
Approve only as a retrieval-based assistant limited to indexed, approved product documentation, with: no price quoting, an explicit "confirm specifications with our team" disclaimer, one-click escalation to a human, full transcript logging, and a weekly audit of 20 conversations for the first two months. If it cannot be constrained that way with no developer, defer it.
150 AI-written posts per month - reject.
- Three people cannot review 150 posts; unreviewed content on a site whose top pages are spec pages is a compliance risk.
- Volume at that scale produces near-duplicate content that competes with existing pages (cannibalization) and reads as low-effort to both buyers and search engines.
- The site does not have a traffic problem that 150 posts solve; it has a trust and conversion problem. Counter-proposal: 12–20 posts a month drafted with AI, each reviewed by someone who can verify the technical claims, aimed at the questions sales hears in deals.
AI personalized first lines - approve with constraints. Test in one segment, draft-only with human send, and forbid invented facts about the prospect's company.
3. Describe the governance and measurement model you would put around any AI content or automation at Fernwood.
Read the worked answer
Governance - the point is that a named human owns every output.
- Source of truth. Product specs come from a maintained, approved data source, not from a model's memory. Any AI drafting tool references that source only.
- Named reviewer per content type. Spec pages and safety guidance require technical review before publication; blog and email require marketing review. "AI wrote it" is never an acceptable explanation to a customer.
- Claims checklist. A short, fixed list of prohibited claims (certification guarantees, compliance assurances, lead times) that every reviewer signs off against.
- Audit log. Every AI-assisted asset records who prompted, who reviewed, and when - necessary for both quality control and liability.
- Automation change control. With nine workflows already tangled, no new workflow goes live without documented enrollment criteria, exit criteria, and suppression rules. That one habit prevents the next mess.
Measurement - pick metrics that fail loudly.
- Email: bounce rate and spam complaint rate as guardrails (target under 2% and 0.05%), with reply rate and meeting rate as the upside metrics, compared against a holdout that gets the non-AI version.
- Blog: engaged sessions, scroll depth, and assisted conversions per post - not post count. If 150 posts a month produced traffic that bounces, the number that matters is zero.
- Chatbot: containment rate, escalation rate, and defect rate from transcript audits; any spec defect is a stop-the-line event.
- Keep a written rollback plan for each initiative. If a metric breaches the guardrail for two consecutive weeks, it goes off.