Jonathon's AI Wiki

AI Transformation Case Studies from The Artificial Intelligence Show (HubSpot, Peapack, David's Bridal, Baptist Health)

15 min read

Source: The Artificial Intelligence Show (Marketing AI Institute / SmarterX), YouTube (yt-podcast), August–September 2026. The four case studies come from the “AI Transformations” series hosted by Mike Kaput and “presented by Google Cloud” (each episode carries a Gemini Enterprise ad read):

The leader Q&A is from the “AI Answers” series: Ep 236, raw/Ep._236_-_AI_Answers_-_No_Time_for_AI_AI_Budgets_Vendor_Terms_Data_Risk_AI_Disclosure.md (Paul Roetzer and Cathy McPhillips) · https://www.youtube.com/watch?v=NLFQkHiaCiQ.

Four first-hand accounts of company-wide AI adoption, each told by the executive who ran it, plus a Q&A episode of leader tactics from the show’s hosts. Every number below is the guest’s own claim, unverified and delivered on a sponsored series. The value is in the operating mechanics, which are specific enough to copy: how each company found use cases, organised builders, measured results and handled governance.

Key Takeaways

  • Measure the business metric, not the bot.

    • Peapack judged its “Penny” chatbot by the support team’s ticket volume (down 60%), not by questions answered.
    • HubSpot reports save rates (+7 points), meetings booked (10,000 in a quarter) and AEO conversions (almost 2,000% growth).

  • Every company put coaches inside the business:

    • HubSpot: open Slack sharing channels
    • Peapack: 29 AI champions embedded in business lines
    • David’s Bridal: the head of product “deployed” to each leader
    • Baptist Health: AI Sherpas

    In each case the coaches pull use cases out of the business rather than having IT push tools in.

  • Centrally built beats rep-built. HubSpot’s globally built agents, “tuned with the right context and the right evals”, outperform what individual reps build for themselves in Gems or Glean.

  • Governance speed is now the bottleneck for marketing teams in regulated firms. Baptist’s CMCO calls year two “quicksand” and wants “lanes that are appropriately slow and appropriately fast.” Roetzer’s fix is “horizontal approvals”: approve a connector once for everyone.

  • Prerequisites are sequential. Peapack’s CTO: the data warehouse (built about five years earlier), in-house developers, then agents — “these things all need to happen linearly.”

  • AI shifts build-vs-buy even for a 700-person firm. Peapack replaced a “$375,000 plus a year” vendor quote with a Codex-built internal tool. It was “the first year that the number of third party technologies our bank used actually decreased.”

HubSpot — Agents Across the Whole Customer Journey (Ep 234)

John Dick, Chief Customer Officer, leads HubSpot’s global sales and customer success. Vendor caveat: much of his stack is HubSpot’s own product.

  • The frame. Go-to-market problems haven’t changed (“how do I build demand? How do I win deals?… happy loyal customers”). AI is interesting because it breaks old constraints: the share of sales-rep time spent talking to customers “just never moves”, and one-to-one personalised customer-success emails were impossible at scale.

  • Fluency came first. It came from the top: the CEO shared a video of how she used AI in her own day. Staff got dedicated time and “a culture of… no fear sharing” through Slack channels for posting wins.

  • Order of deployment:

    • support first (“the first place where generative AI had super clear product market fit”)
    • then marketing content and engineering code
    • then the full journey

    The early lesson from support was to optimise for CSAT, not resolution rate: “if you build great CSAT your resolution also goes up.”

  • Agents and reported results:

    • an AEO strategy plus an AEO agent: AEO conversions grew “almost 2,000%” over “the last couple months”
    • a demand agent that sets fit criteria and finds companies with intent signals (the “best version of our understanding of our TAM”)
    • an AI SDR sales bot handling “over 80%” of website chats
    • a prospecting agent that “booked 10,000 meetings” last quarter
    • Guided Success, a custom rep assistant built on HubSpot’s Breeze: “the win rate is up a bunch”
    • a CS assistant that triggers agents: “a seven point increase in save rates”

  • Org design went through three phases:

    1. “Wild west” plus hackathons.
    2. About a year of cross-functional pods (subject-matter experts, engineers, data people) under one AI-focused marketing growth leader, Kieran Flanagan (captioned “Flanigan”). “We moved all the KPIs.”
    3. About three months before the interview, all go-to-market agentic engineers, systems, data and product people and SMEs were direct-lined to him, because pods still meant “competing priorities” and “coordination cost.”

    The trade-off Dick accepts: global priorities win over local experiments.

  • The goal is “from individual productivity to institutional productivity”, meaning moving P&L lines, not saving individuals minutes.

  • Advice:

    • “You are not behind.”
    • Pick one real problem.
    • Set a high quality bar: “I hate slop… human authenticity plus AI efficiency is like the secret sauce.”

Peapack Private Bank & Trust — Champions, Codex Builds and Digital Employees (Ep 238)

The CTO (first name John; surname garbled in captions as “Cowall”/“Coowalt”) speaks for a boutique private bank: about 700 employees, 13B.

  • 2023 start. They began when Microsoft Copilot was still “Bing Chat for enterprise”. Policies and governance came first, then extensive training. By the end of 2023 every employee had AI chat and AI meeting summaries.
  • 29 AI champions, the CEO’s idea, are embedded in business lines and moved AI “from being technology-driven to business-driven”:
    • after one year: about 100 projects completed and a pipeline of 80+
    • the champions meet monthly as a group and individually with the CTO
    • they present wins at a quarterly internal webinar
    • they hold a dotted-line report to the CTO
    • the ask to them is “bring us the challenge”, not the answer
  • The CEO asked the CTO to spend 70% of his time on AI. That triggered a reorganisation and a dedicated AI engineering team. Everyone in IT has a role; support teams also support the AI tools.
  • The stack is standardised primarily on OpenAI: ChatGPT Enterprise for staff, Codex for developers, enterprise APIs for agents. Project Atlas layers AI over the data warehouse schema plus a knowledge base of how the bank operates, turning analysts’ questions into queries. It logged 2,000+ interactions in a month.
  • Build instead of buy. A wealth-division account-review process lived in spreadsheets; a vendor wanted “$375,000 plus a year.” AI engineers built a Codex “mini platform” pulling from the warehouse, and each review takes 50% less time.
  • Digital employees are named agents that sit on top of other agents, automations, data and classic tools. They reach staff through Teams (urgent) and Outlook (routine), you can reply to them, and each has its own Mac mini where it needs a computer. They call deterministic scripts where AI isn’t needed.
    • Alex (IT) watches the ServiceNow queue: it suggests replies, routes tickets and monitors escalations. A few weeks after launch it was “closing roughly half the tickets of a full-time employee.” It is being given a skill to message staff on Teams when they near their compliance-training deadline.
    • MIA (“marketing intelligence agent”) runs a weekly review of all the bank’s websites for dead links and outdated content, and makes AEO/SEO/GEO suggestions.
    • Naming them as members of a department rather than after a task changes what people ask of them: “It’s not the IT support bot. It’s Alex.”
  • Measured outcomes:
    • Penny (financial-centre chatbot): support tickets down 60% in four months
    • an iPad proforma app for bankers: 100 proformas in the first month, made in about half the manual time
    • a redesign of the ~55-step loan-closing process with an outside AI partner (captioned “Verapath”)
  • Lessons:
    • “Start simple, learn, and the innovation will come.”
    • Finding use cases isn’t the hard part; being ready to implement them is.
    • When something fails, “try again in a few months.” Document Q&A failed in 2023 and is now core.
    • The key guardrail: “AI cannot replace a control. It can supplement controls.” Keep a formal inventory of everywhere AI is used in a process.

David’s Bridal — Pearl Planner and an In-House Harness (Ep 240)

Alina Vilk (President and Chief Business Officer) and Mike Ball (Head of Product). The company was founded in 1950; Vilk claims “90% of all brides” pass through its ecosystem.

  • The problem is stress. Typical wedding checklists have about 30 steps. The real job is hundreds of connected tasks (Kaput cites 300+); one checklist line, the engagement shoot, hides about 20 decisions.
  • Pearl Planner turns date, style, budget and preferences into a guided plan. It shows only what matters now (“if you’re getting married in 18 months… you’re going to see only what you need to see right this moment”) and matches vendors at the right point in the sequence.
  • The vision quiz uses expert knowledge as the eval rubric. Brides pick images instead of describing a style they can’t name. Ball sat with stylists and merchandisers (“why do we call it boho?”) across dresses, flowers and cakes, and uses that knowledge on the back end to judge what the AI infers from the images.
  • Budget by priority, not dollars. Couples mark categories as splurge or save. When they later need money, the assistant suggests cuts only from low-priority categories, using memory that persists across the platform.
  • The stack is their own orchestration harness: agent roles, permissions, triggers and guardrails, built in-house rather than using an out-of-the-box framework so it can be reused internally. They rotate models only after reviewing evals and logs, and let new releases “sit for a little bit” first. Several agents are in use, for example one that logs expenses from receipts, screenshots or emails.
  • Pearl Connect is the vendor side, “entirely powered by Google Cloud” (the series sponsor). Vendors see a bride’s vision before replying, and incoming messages are triaged by intent and sentiment.
  • Traction: “hundreds of thousands of users” before any marketing; free for couples.
  • Internal adoption:
    • no usage mandate; start from each team’s problem statement (merchandising and finance differ)
    • Ball’s rule: “30 minutes of my time will save you 10 to 15 hours a month”
    • a monthly survey where most staff self-report at least 20 extra hours of output a month
    • a maturity-level framework before anyone goes “full speed with Fable”
    • the PMO standardised Slack and Confluence structure, with org-level rules so agents know where each project’s information lives

Baptist Health South Florida — Marketing Leads the Enterprise (Ep 242)

The Chief Marketing and Communications Officer, Christine (surname garbled in captions), runs marcom at South Florida’s largest health system: 12 hospitals, 30,000+ employees, 4,500 physicians.

  • The trigger was a late-2024 talk by Roetzer (the host’s own CEO) to a room of CMOs: her “holy bleep moment.”
  • Year one: literacy with teeth.
    • A year-long AI curriculum built like a college syllabus: external modules, IT and HR lunch-and-learns, keynote watch parties, and dedicated work time to practise.
    • Progress was tied to performance conversations and evaluations: “This wasn’t optional.”
    • Baseline surveys at the start, repeated quarterly.
  • Supporting structure:
    • a department AI council linked to the enterprise council
    • marcom-specific usage guidelines
    • learning labs^[ambiguous] (see Open Questions)
    • office hours and a Slack channel
    • a learning hub holding resources, the team’s own use cases, and its successes and failures with the reasons
    • short pilots decide whether a tool is for everyone, a smaller group or no one (per Kaput’s intro)
  • AI Sherpas coach; they don’t do the work. Early on they tested tools and tutored. Now a team brings a problem, the Sherpas advise, the team builds, and comes back for more help: “iterative and supportive rather than hey, go do this for me.”
    • The standout Sherpa came from restaurant entrepreneurship, not tech. The traits that mattered were curiosity, resourcefulness and fearlessness.
    • The Sherpa role has become a formal job: “marketing technology transformation and enablement.”
  • Results:
    • Online listings: a manual job that “once took somebody about 10 hours every week now takes about two minutes”, as an automated red/yellow/green dashboard built in a couple of weeks. It is now being rolled out to the other teams that manage listings.
    • Creative self-service: designers built on-brand Adobe Express templates; account teams now produce routine jobs themselves, with a human approving at the end. That cut the creative backlog and reliance on external freelancers.
  • People lessons:
    • What looked like resistance was “pride of expertise”, which called for more empathy while keeping expectations firm.
    • Other departments (strategy, finance, HR) now join marcom’s sessions.
  • Year two is “quicksand”. Tools are built but integrations stall in security and governance review. Her case: split clinical work (slow lane) from corporate functions like marketing (fast lane), because “this talent won’t sit tight.”
  • Year three is workflow redesign. “Because we’re starting with traditional structures, it is very hard to build an AI native… team.” Whiteboard by function, not by people’s names, then decide where a human is actually needed.
  • Advice:
    • “You got to go first.”
    • Leaders “cannot outsource their own AI literacy.”
    • Give people “real tools, real rules, time to learn, visible support.”

Leader Q&A — Tactics From the Hosts (Ep 236)

Paul Roetzer (CEO, SmarterX) answering audience questions, with Cathy McPhillips (CMO):

  • For a time-starved team, run a forcing-function workshop. “We are all going to leave here in 90 minutes with one to two AI use cases… that’s going to save us at least three hours a week.”
    • About 20 minutes setting the stage, 25–30 minutes listing tasks over 3 hours a week and prioritising, then 30 minutes building or sharing.
    • Stretch it to two hours for a real build session, and repeat every other week or monthly.
    • McPhillips adds “jam sessions” where someone screen-shares while doing the work.
  • Use consultants as a short-term fix and build the skill in-house. SmarterX is creating an internal “Labs” unit that works like forward-deployed engineers. AI ops roles can come from technical marketers or project managers.
  • Horizontal approvals. “If we want to connect HubSpot to ChatGPT, let’s approve it once.” Read-only, for many uses, so nobody re-requests the same connection.
  • Human checkpoints in agents. Build approval gates or read-only access. Let only low-risk actions run unapproved (his example: a weekly marketing summary emailed internally).
  • Vendor terms triage. Paste a tool’s terms of use into ChatGPT and ask how it handles your data, or compare the contract against your AI policy and ask what to flag for your attorneys. Legal still decides. Re-review when terms change (he cites the Fable 5 data-retention change).
  • Budgeting AI.
    • Priority order: people to decide, then platforms and token budgets, then infrastructure (Mac minis, GPUs), then possibly your own fine-tuned open-weight models.
    • The only forecasting model he has seen is a monthly cap.
    • Plan for “intelligence redundancy” rather than betting on a single provider.
  • Disclosure default. “When authenticity matters you should be doing the work.” Otherwise disclose simply, e.g. “co-authored with Claude”, and always if an agent answers your email for you.
  • Entry-level talent. Roetzer’s working theory is apprenticeships supported by AI-guided learning. He says nobody has solved where future managers come from.

Try It

  1. Map your customer journey against AI (Dick’s prompt): “Here’s what my customer journey is. Help me map… AI solutions against each of those.” Pick the stage with the biggest constraint, not the easiest.
  2. Run Roetzer’s 90-minute session on that stage: everyone leaves with one or two use cases worth at least 3 hours a week, and at least one built before the session ends.
  3. Name a digital employee for marketing, Peapack-style. Start with a weekly site audit (dead links, stale content, AEO/SEO suggestions) that posts findings where the team already works.
  4. Measure the downstream metric. Track tickets, hours or conversions for the team affected, not usage of the tool.
  5. Ask IT for one horizontal approval, e.g. read-only CRM to your assistant, and a fast lane for low-risk marketing tools.

Open Questions

  • Learning-lab cadence: Kaput’s intro to Ep 242 calls Baptist’s learning labs “weekly”; the guest says “bi-weekly AI learning labs.” The guest’s account is more likely correct.
  • Garbled names. Captions mangle several: Peapack’s CTO (surname), Baptist’s CMCO (surname), HubSpot’s growth leader (“Flanigan”, most likely Kieran Flanagan) and Peapack’s AI partner (“Verapath”). Verify before quoting.
  • HubSpot support metrics are unclear. The support-side statement (“industry standard resolution rates… 60% CSAT is high”) is garbled; no clean number was recorded.
  • Sponsored, self-reported numbers. None of the figures are audited. The series is presented by Google Cloud, and HubSpot’s CCO is describing HubSpot’s own product.
  • Baseline for AEO growth. HubSpot’s “almost 2,000%” AEO conversion growth has no starting figure, so the absolute impact is unknown.

Graph View

Backlinks