7 AI marketing projects: real numbers, real costs, one real failure
Most AI marketing case studies are vendor materials with better formatting. The wins are real. The costs are missing. The failures are nowhere. You read three of them and walk away thinking everyone is succeeding with AI overnight on a shoestring. That’s not the full picture.
What follows is a different kind of rundown: seven real AI marketing implementations from Lilach Bullock, a marketing consultant with 21 years of experience who went all in on AI in 2024. Names are anonymised where clients didn’t give explicit permission. Every number is real. The failure case is included because it’s the most instructive one in the set.
The Pattern Upfront
Before the case-by-case breakdown, here’s the pattern that holds across all seven: AI as a multiplier on existing judgment and operational tasks produces wins. AI as a substitute for editorial judgment or human-to-human communication produces failures. Keep that in mind as you read.
Case 1: B2B SaaS, Lead Enrichment
A sales team was spending 2.5 hours per day on lead research and triage. About 30% of qualified inbound leads were going cold before anyone reached them.
The fix: an AI lead enrichment workflow connecting HubSpot to Claude via the API, with structured prompts for company classification, ICP fit scoring, and recommended outreach approach. Built on n8n + Claude API + HubSpot custom properties. Six weeks from design to full production rollout, with a parallel manual review pilot in week four.
After 90 days:
- Time from inbound to first sales touch: 4.2 hours down to 38 minutes
- Lead qualification accuracy vs. manual baseline: 89% agreement
- Sales team time freed: roughly 1.5 hours per rep per day
One prompt engineering issue came up early: initial prompts over-categorised companies as “enterprise” because they had “Inc” in their name. Two rounds of iteration fixed it.
Case 2: Professional Services, Proposal Drafting
Senior consultants at a professional services firm were spending 6 to 12 hours per proposal, writing 3 to 4 per month. The proposal turnaround was slowing the sales cycle.
The solution used Granola to transcribe discovery calls, then fed transcripts into Claude with a structured prompt that produced a first draft matching the firm’s template, voice, and pricing tiers. Senior consultants reviewed, edited, and shipped within 90 minutes instead of a full day. Setup took three weeks, including voice extraction from eight past proposals and testing on five historical discovery calls.
After 60 days:
- Average proposal turnaround: 8.5 hours down to 2.8 hours (67% reduction)
- Proposal acceptance rate: 41% up to 47%
- Monthly proposals shipped: 3.2 up to 5.8
Getting the voice right took four prompt iterations. The final prompt included eight explicit voice patterns the firm uses. Early drafts sounded generic until those patterns were extracted and embedded.
Case 3: E-commerce, 450 SKU Product Descriptions
A UK-based DTC consumer brand with an 8-person team had 450 SKUs with thin product descriptions. The workflow pulled product attributes from Shopify, generated a 200-word description per SKU using a brand voice prompt, generated SEO metadata, queued outputs for human review, and published approved descriptions back to Shopify.
Every description was human-reviewed before going live. Technical specs came from the supplier database, not from the AI. Hero product copy was written manually. The rollout took eight weeks, with 100% quality review across all 450 SKUs.
After 90 days post-rollout:
- Non-brand organic traffic: +22%
- Product page average time on page: +18%
- Add-to-cart rate on previously thin-description pages: +14%
- Conversion rate on those pages: +9%
Case 4: IT Services, Sales Call Prep and Follow-Up
A mid-market US IT services firm with eight salespeople was losing 30 to 50 minutes per call on prep and 60 to 90 minutes per call on follow-up. Two paired workflows addressed both.
The prep workflow ran 30 minutes before any meeting with a new prospect: it pulled LinkedIn, company website, and news mentions, then generated a 1-page brief covering likely topics, recommended questions, and anticipated objections. The follow-up workflow used Granola to transcribe the call, then an LLM generated a structured follow-up email for the rep to review and send within 30 minutes.
After 90 days:
- Average prep time per call: 42 minutes down to 8 minutes
- Average follow-up time per call: 75 minutes down to 18 minutes
- Sales cycle length: 31 days down to 18 days (41% reduction)
- Win rate change: not statistically significant (held within ±2%)
First-version follow-up emails sounded too clinical. Three rounds of iteration brought the tone closer to each salesperson’s natural style.
Case 5: Coaching Business, Content Production
A coaching business was capped at one blog and three social posts per week. Scaling without hiring would have eroded margin.
The solution was a hybrid workflow with a clear human/AI split. The coach handled topic selection, outline, opening, contrarian take, and close. AI handled research, structure validation, supporting sections, social repurposing, SEO metadata, image generation, and scheduling. Setup took two weeks, with daily review for the first two weeks post-launch.
After 90 days:
- Blog publishing cadence: 1/week up to 3/week (no quality drop)
- Social posts: 3/week up to 12/week
- Email list growth: +28%
- Coach time on content: 12 hours/week down to 6 hours/week
The reason it worked: the coach is opinionated. AI couldn’t substitute for the voice, but it removed the operational friction around publishing. The human-heavy split is what held quality.
Case 6: lilachbullock.com, Self-Built SEO Recovery
This one is Bullock’s own business. The personal site had decayed: broken sitemap, low schema coverage, thin internal linking, organic traffic flat-lined. The consulting pipeline needed rebuilding.
The solution was built end-to-end in Claude Code over six weekends. Zero external spend. Roughly 60 hours of total build time. The components:
- WordPress publishing pipeline (.docx to SEO-ready draft)
- Schema markup generator (FAQPage, HowTo, BreadcrumbList)
- Sitemap diagnostic and re-save trick for stuck posts
- Internal linking analyser and deployer
- Data pulls from GSC, GA4, and Bing Webmaster Tools
After four months:
- Sitemap: 904 URLs up to 1,300+
- 280 stuck posts unstuck via the re-save trick, recovering roughly 83,000 monthly impressions
- 70+ posts now carry FAQPage schema; 17 carry HowTo schema
- 15 new cornerstone posts published in one month
- Organic traffic: recovered from a 35% mid-2025 drop (caused by AI content over-leaning) and finished up 12% on baseline by month four
- Time savings: roughly 400 hours per year on content publishing operations alone
Payback: six weeks. The self-built case consistently outperforms many paid consulting engagements on ROI because the labour cost is your own time, not a vendor rate.
⚠️ Case 7: The Failure — AI Cold Outreach at Scale
This is the case worth reading most carefully.
A US-based B2B agency with 18 people wanted to scale cold outreach from 80 emails per week (written by one BDR) to 1,000 per week using AI personalisation. The LLM read each prospect’s LinkedIn profile, generated a personalised opening line, and varied body copy by inferred persona. Connected to Smartlead for delivery. Two weeks to build, two weeks to rollout.
What happened:
- Human-written baseline response rate: 11.2%
- AI-generated response rate: 3.8%
- Replies labelling the emails as clearly AI-written: 47 in the first month
- Prospects who publicly called out the agency on LinkedIn: 2
One LinkedIn callout reached 30,000 impressions within 24 hours. The campaign was cancelled after six weeks. Brand recovery took four months of reduced outreach volume and personal apologies to affected prospects.
Three things went wrong:
- The base assumption was incorrect. Volume isn’t the constraint in cold outreach. Quality is. Adding 10x volume at lower quality is net negative.
- AI personalisation reads as AI personalisation. Mentioning a prospect’s LinkedIn post the way an AI would mention it still feels generic. Perceived human attention can’t be faked at scale.
- Brand damage propagated faster than any campaign benefit could accumulate. One public callout outpaced weeks of outreach gains.
“AI cannot scale anything that depends on perceived human attention. Cold outreach is exactly this category.”
The same lesson applies to networking, warm-prospect follow-up, and customer success communications. Use AI for prospect research in cold outreach. Don’t use it for the writing.
What the Seven Cases Tell You
The pattern holds across all seven engagements:
- Operations work pays back fast. Cases 1, 2, 4, 5, and 6 are all operations-focused. All paid back in under eight weeks. AI removes friction without changing quality when the task was previously routine human work.
- Content work pays back with discipline. Cases 3 and 5 both worked because humans stayed in the loop for quality decisions. Mass AI content publishing without review is a different strategy and a riskier one.
- Brand-touching work fails when AI replaces humans. Anywhere the work product is a direct human-to-human signal, AI substitution fails. Case 7 is the clearest example.
- Self-build is competitive. Case 6 delivered better ROI than several of the consulting engagements because the cost was time, not fee.
If you’re deciding whether AI fits your current problem, the question isn’t “can AI do this?” It’s “does this task depend on perceived human attention?” If yes, AI handles the research and prep, not the output itself.
I’m founder of BizStack at Cagri Sarigoz LLC and a passionate advocate for entrepreneurs.
With over 14 years in tech, marketing, and AI, including my role as Head of SEO at CitizenShipper and co-founder of TaleBot at Intale AI, I’m dedicated to sharing genuine, useful product insights and tips.
At BizStack, I aim to cut through the digital noise to provide clear, actionable advice.
And more than all else, I’m a father to a (always) little girl and a husband.
Contact me at [email protected] for assistance.
