In this Mucker Growth Series session, we invited friends from one of our portfolio companies Haley Chute, CPO & CMO, and Harpreet Marwaha, COO & CFO of Octagos Health, to share how they've embedded AI across their entire organization — from a top-down mandate to a Claude-powered platform connecting every department. A recap of what they covered, as well as the full webinar recording, is below.
Making AI Non-Optional
Octagos is in the business of helping cardiologists remotely monitor patients with surgically implanted devices — pacemakers, loop recorders, defibrillators. They're processing close to a million patients across 300+ clinics, using AI to triage and contextualize the arrhythmia data those devices transmit around the clock. While they built an AI-based product, for this session we wanted to learn how they use AI internally across their various business operations functions, like marketing, sales and product design.
The drive to use AI started with their CEO, Dr. Shanti Bansal — a cardiologist with a computer science degree who's also a self-described tinkerer. Late last year, Claude adoption started accelerating at the executive level. Leadership was texting each other: "Hey, look what I built with this." Vibe-coded dashboards were being shared.
That led to an all-hands where the mandate became explicit: AI use is not optional.
Getting there wasn't instant. Most people don't want to change how they work, and even when they say they're going to try AI, what often happens is they use it for basic chat prompts and call it a day. Octagos took a more deliberate approach. They required everyone to complete an AI training course and share their certificate in the company-wide Teams channel. Then they assigned internal champions — engineers and leaders who did screen shares with colleagues, showed them their first win, and helped them build confidence with the tools.
One Platform, Connected to Everything
Once the mandate was in place, the bigger challenge was enablement. Each team had to figure out what to offload, what to connect, and how to make it work in a regulated industry.
Octagos landed on a single governed platform: Claude, with different license tiers based on data access needs. Teams that touch protected health information (PHI) run on Claude Enterprise, which is HIPAA compliant. Teams that don't — say, someone iterating on a marketing email — are on the Claude Teams plan, which costs less and is more than sufficient for the work.
Under that sits an MCP layer that connects Claude to their actual systems. The easy connections were the ones with official MCPs already built: HubSpot (CRM), QuickBooks (billing), Power BI (dashboards), Microsoft 365, Figma (design), Google Drive, Brex (expenses). The hard ones required building in-house — their own database connector (read-only, secured), a custom connector for their clinical platform, and others.
For a healthtech company, this matters more than in most industries. Patient data carries strict access controls that don't disappear just because you're running AI on top of your systems. Octagos uses dummy data (AI-generated) in sandboxes, maintains clear read-only boundaries for non-clinical staff, and treats PHI access as a governance question first.
Why Claude specifically? Two reasons: HIPAA compliance made it the obvious choice for a regulated industry, and the breadth of existing MCP integrations meant they could move fast without building everything from scratch.
Claude for Marketing: Deck Design, Pipeline Analysis and Competitive Intelligence
Before Claude, the marketing team owned all branded output. If sales needed a pitch deck, they'd submit a brief. Marketing would build it in PowerPoint, send it back, get comments, revise, repeat. The deck would ship days later.
Now, sales builds it themselves.
Haley's team fed Claude Design their entire design system — the actual logo, brand fonts, color palette, component library — and packaged it as a skill that anyone can access. A sales rep types a prompt ("build me a pitch deck for this customer based on these talking points"), and gets back a fully on-brand deck, not a generic AI-generated approximation. The first version took some effort to build, but anyone with an existing brand template can replicate this as a starting point.
The result: what used to take days of back-and-forth now ships in minutes. Marketing didn't lose control of the brand — they just moved from being the hands to being the architects of the system.
For pipeline analysis, Octagos connected Claude to their CRM, digital advertising platforms, emails, and calendar. They now get automatically generated updates pushed to Teams when new leads come in, and their Monday readouts are pre-built PowerPoint slides they can copy and paste directly. No one hand-builds these anymore.
The competitive briefing piece goes even further. Every morning, Haley gets a brief that scans what her competitors posted on their website, what appeared on their social media, what came up in a sales email or a meeting that was flagged. It also updates their competitive intelligence database automatically. The information isn't just surfaced — it's maintained.
And then there's the morning briefing, which Haley recommends as the first thing any startup should build. The setup is straightforward: tell Claude which systems to scan (email, calendar, Teams messages, competitive signals), set priorities ("anything from Preet, always surface it; anything related to contracts, always surface it"), and schedule it to run before you start your day. The result is alert triage: you start the day knowing what matters, instead of spending 30+ minutes sifting. Haley doesn't read her email until the end of the day unless it's been flagged. The whole setup takes about 30 minutes to build.
Claude for Product Development: Design, Mockups, Engineering & Product Analytics
On the product side, Haley's most important point was this: the AI-enabled state they're in now didn't happen overnight. They tracked their evolution by quarter, starting from early Q4 2025.
Early on, ideation was manual — mockups drawn on paper or built in PowerPoint. They pulled in ChatGPT for ticket cleanup. Design was outsourced. Reporting lived in spreadsheets.
By late Q4, they were using Google AI Studio for simple mockups and had brought Figma in-house. Through Q1 2026, they moved to Claude Desktop and eventually Claude Code. Prototypes became HTML-based instead of static images. Tickets started getting auto-generated first drafts. Figma got connected via MCP. Dashboards got vibe-coded.
Now, the product team runs an end-to-end workflow entirely through AI:
- Product manager types a prompt ("generate a new user modal")
- Claude Design produces a mockup using the design system
- Back-and-forth refinement, then push to Figma via MCP
- Figma design gets converted to frontend code via Claude Code
- Code and visuals get packaged into an Azure DevOps ticket via CLI
- Engineering takes the ticket — which already contains the code, visuals, and context — and ships it
The product team — people who were never in a terminal window doing dev work — is now writing in plain language through Claude Code. UI/UX review doesn't disappear from this process, but it's downstream, not upstream. The designer isn't blocking speed; they're reviewing output.
One often-overlooked benefit: non-engineers can now access the codebase as a source of truth. A marketer at Octagos has read-only access to their product code via Claude, which means she can ask questions about how the software works without needing a SQL background or an engineering ticket. That kind of self-sufficiency across the org — non-technical people getting technical answers without slowing down engineers — compounds over time.
Lessons Learned
Lesson 1: Connect AI to everything, or the value is limited.
The "siloed version" of AI — where each person uses ChatGPT for their own tasks, disconnected from the company's data — isn't that different from what was available before. The inflection point happens when all teams access the same systems through the same platform. Disparate data becomes connected insight. Non-technical people answer questions that previously required an engineering sprint. Account managers can get a full picture of what's happening with a customer — JIRA tickets, phone call logs, clinical notes — in one place, without logging into four different tools.
Lesson 2: Keep humans in the loop.
AI slop is a real problem, and it's insidious because it looks competent. The failure mode is: you put a prompt in, think you know what it says, and forward it to someone. Then you're in a meeting with a slide you haven't actually read. The discipline Haley's team built is simple: read your AI outputs before you send them. Working with AI is like working with an intern you've never met — you don't yet know their blind spots, so you have to check everything. The fix isn't to stop using AI; it's to build validation into your workflow. Starting a task in a fresh prompt (instead of continuing a context-heavy conversation) often produces cleaner output. Having one Claude window review another catches errors you'd otherwise miss.
Lesson 3: Make it scalable with governance.
Early adoption feels like creativity. Scaling it requires protocols. Who has access to what systems? Who owns the accuracy of a shared artifact? How do you ensure agents are operating within your company's values, especially in a regulated industry? At Octagos, they built internal champions not just to onboard people but to maintain standards — someone whose job includes greasing the wheels, answering questions, and catching problems before they propagate. That role doesn't have to be a full-time hire; it can be a significant portion of an existing person's responsibilities.
Lesson 4: Watch the economics.
This one surprised them. Increased token usage resulted in going from a $10K monthly bill to now close to a $40K month bill. The increase didn't come from misuse; it came from success. Successful adoption and overspend look identical until you measure by team and workload.
They implemented three things to help optimize usage and manage cost: they moved non-PHI users from Claude Enterprise to the Claude Teams plan, rebuilt the heaviest dashboards as standard engineered assets in Power BI (pulling from Claude on a schedule rather than continuously), and moved recurring dashboard refreshes out of Claude entirely.
From Preet's role as the CFO, his view on justifying that: "On a purely numbers basis, while the cost is going up, I also look at all the other areas where it's offset — delayed hires, accelerated roadmap, lower churn." The math isn't just cost; it's cost against time-to-revenue, increases in productivity and speed, and hiring decisions you didn't make.
Their long-run playbook for AI economics:
- Prove it on the frontier — the fastest path to knowing whether it works for your use case
- Optimize what repeats — once a workflow is proven, move it to the cheapest model or tool that handles it reliably
- Keep your exits open — open-source and on-prem options matter as volume grows and data sensitivity increases
The summary they closed with: Make it a mandate. Pick one platform and go deep. Connect real systems. Ship skills, not prompts. Keep humans in the loop. Watch the economics.
