Most content about AI in companies focuses on engineering. This makes sense: coding is where AI tools are most mature. But if you only use AI in engineering, you're leaving massive leverage on the table.
At Horizon, every function has its own AI stack. Not because we're AI enthusiasts (though we are), but because each function has repetitive, high-volume work that AI handles better than humans. This post walks through exactly what we use, where, and why.
Sales
Day AI is our CRM. It's 100% agentic and AI-first. It listens to calls, automatically moves deals through the pipeline based on meeting content, assigns owners, suggests next steps, and lets us query our sales data in natural language. The key value: the pipeline reflects reality because it updates from actual conversations, not from reps remembering to update fields.
Howdy handles meeting coordination. It's an AI assistant that schedules, reschedules, and manages calendar logistics. This sounds trivial until you realize how many hours per week a sales team spends on scheduling. For us, it's fully automated.
Superhuman for email. The AI features let us draft responses faster and maintain consistent tone across the team. The speed matters in sales where response time correlates with close rate.
Design and Customer Success
Devin, Cursor in Slack, and Claude in Slack handle quick wins. When a CS team member spots a small UI issue or a customer requests a minor change, they can tag an AI agent directly in Slack. The agent proposes a fix, a Product Engineer reviews it, and it ships.
This is the workflow that surprised us most. Non-engineers creating PRs through natural language in Slack. The documentation layer (AGENTS.md, agent_docs) makes this possible because the AI has enough context to write correct code without an engineer specifying every detail.
Gamma.app for design presentations. Every presentation we create (including the one this article series is based on) is built with Gamma. Fast iteration, consistent design, no time spent on slide formatting.
Product Management
This is where AI has the highest density of use cases for us.
Notion AI for meeting preparation, note extraction, and post-meeting action items. Before every client meeting, Notion AI generates a briefing from previous notes. After the meeting, it extracts action items and distributes them. This replaces the manual process of reading through past conversations to prepare.
PostHog AI for product analytics. We ask questions in natural language ("How many insights did admin users generate last week?") and get charts, dashboards, and data breakdowns without writing SQL or configuring analytics tools manually. When we need a new metric tracked, we describe it and PostHog AI creates the insight and adds it to the relevant dashboard.
Day AI (again, but for product) shares meeting information from customer calls. When our sales team talks to prospects, the product team gets structured summaries of what features were requested, what pain points were mentioned, and what competitors came up. This flows automatically from call recordings to product context.
ChatGPT for summarizing long-form content. Videos, documents, articles, books. When someone on the team finds a relevant resource, they run it through ChatGPT to produce a summary that gets shared in Slack. This turns a 45-minute video into a 2-minute read.
Slack AI for summarizing conversations. When discussions happen in Slack channels over multiple days, Slack AI generates summaries that catch up anyone who missed the thread. This is particularly valuable for our distributed team across time zones.
Retool with AI-assisted queries for understanding database content and building internal monitoring dashboards.
Marketing
ChatGPT for structuring and aligning content. Newsletter drafts, blog post outlines, and social media content all start with AI-generated first drafts that get refined by humans. The AI handles structure and consistency. Humans handle voice and strategy.
We have a prompt template for our product newsletter that includes our audience definitions, tone guidelines, and content objectives. The AI produces drafts that are 80% ready, and a human brings them to 100%.
The Pattern Across All Functions
If you look across these use cases, a pattern emerges. We're not using AI for "thinking" in any function. We're using it for three categories of work:
Data synthesis. Taking large amounts of information (call recordings, Slack threads, analytics data, documentation) and producing compressed, actionable summaries. Humans can't read 102 meeting recordings from last week. AI can.
Template execution. Applying known patterns to new inputs. Writing a newsletter follows a template. Generating a meeting prep follows a template. Fixing a bug in a well-documented codebase follows a template. The AI handles the execution, humans defined (and refine) the templates.
Coordination logistics. Scheduling meetings, distributing action items, routing information between teams. These are high-volume, low-judgment tasks that AI handles reliably.
We don't use AI for strategy, judgment calls, or creative direction. Those require human context and values that models don't have. The line is clear: AI handles volume and speed, humans handle judgment and direction.
What This Costs
We're a startup of about 15 people. Our total AI tooling spend across all functions runs between $3K and $5K per month (all tools combined across all teams). The ROI calculation is straightforward: compare the hours saved against the subscription costs. A single CS team member resolving bugs via Slack AI agents saves 10+ engineering hours per week. That alone covers the entire tooling budget.
The harder cost is adoption. Each tool requires someone to set it up, define the workflows, and train the team. We invested significantly in this during the first few months. Now it's self-sustaining because each team member has internalized which AI tool to use for which task.
Starting Point for Other Teams
If you're an engineering leader reading this and thinking "we should do this beyond engineering too," here's the sequence we'd recommend:
First, identify the highest-volume repetitive tasks in each function. Not the strategic work. The work that feels like it should be automated.
Second, find the AI tool that fits that specific workflow. Don't try to use one tool for everything. ChatGPT is great for summarization but bad for CRM automation. Day AI is great for sales but you wouldn't use it for product analytics.
Third, build the workflow and templates. The tool is just infrastructure. The real value is in the defined process that the tool executes.
Fourth, measure. Hours saved per week per person. Error rate before and after. Speed of task completion. Real numbers, not impressions.
Every function compounding small AI efficiencies adds up to a fundamentally different operating speed.