YC Root Access

Parahelp: The End-to-End AI Support Agent

Parahelp is building an end-to-end AI support agent. One that doesn’t just reply, but actually resolves tickets by taking real actions inside a company’s tools. The team recently raised an $18M Series A to scale this vision as their agent now closes thousands of tickets every day. In this conversation with YC's Jared Friedman, founders Anker Ryhl and Mads Liechti share how they went from cold-emailing 700 prospects during YC to powering support at companies like Perplexity, ElevenLabs, and Replit. They talk about the mid-batch pivot that changed everything and how their new agentic architecture lets them onboard customers in hours instead of weeks. Learn more about Parahelp at https://parahelp.com. 00:00 – Intro: What Parahelp Is Building 00:46 – The First Version (and Why It Didn’t Work) 03:12 – YC, Cold Emailing 700 Prospects, and Finding the Problem 06:05 – The Mid-Batch Pivot That Changed Everything 08:40 – Landing Perplexity as Customer #2 11:28 – What “End-to-End Resolution” Actually Means 14:20 – How the Agent Takes Real Actions (Refunds, Fixes, Workflows) 18:05 – Building the Agentic Architecture 22:18 – Scaling to Thousands of Tickets Closed Every Day 29:04 – Onboarding Customers in Hours, Not Weeks 34:40 – Raising the $18M Series A 40:52 – Advice for Founders

ARAnker RyhlguestMLMads LiechtiguestJFJared Friedmanhost

Dec 2, 2025·47m·Watch on YouTube ↗

📖 CHAPTERS

1. 1

   Parahelp in one sentence: an AI support agent that resolves tickets end-to-end

   The host sets the stage: Parahelp is a YC-born agentic AI company pushing beyond “drafting replies” into actually closing customer support tickets. The episode will cover the founders’ origin story, the tech behind their agentic system, and their recent Series A.

2. 2

   From small-town Denmark to building together: the founders’ early arc

   Anker and Mads describe growing up near Copenhagen, meeting through a youth entrepreneurship community, and teaming up as builder + designer. They built multiple consumer apps during high school, forming a product-obsessed, rapid-building mindset.

3. 3

   The first real startup: an NFT investing app—and why it failed

   Their first venture was an “easy NFT investing” app with Apple Pay, even partnering early with Stripe in Europe. By the time they launched, the NFT market had collapsed and remaining users didn’t value product quality—only price movement—forcing a pivot.

4. 4

   Pivot discipline: 3-week sprints, postmortems, and discovering Silicon Valley

   After the NFT product, they adopted a structured pivot process: three-week build-and-sell sprints followed by written postmortems. They visited San Francisco, explored hacker houses, and became convinced to pursue a bigger B2B opportunity.

5. 5

   YC entry and the co-pilot version that didn’t create urgency

   In Berkeley, they met founders who pushed them to apply to Y Combinator. Their initial YC-era product was a co-pilot for support, but customers weren’t “disappointed without it,” and sales cycles were long—signaling the pain wasn’t acute enough.

6. 6

   The mid-batch pivot: going all-in on end-to-end support resolution

   A model capability jump (Sonnet 3.5) convinced them agentic tool use could be reliable with enough sampling and judging. They pivoted mid-batch to full end-to-end resolution, then used aggressive outbound to quickly validate demand.

7. 7

   Landing Perplexity as customer #2 and validating the wedge

   Perplexity became Parahelp’s second customer during the YC batch—an unusually fast proof point. Working with a hyper-growth AI-native company validated that solving complex, high-stakes tickets end-to-end was a compelling differentiator.

8. 8

   What “end-to-end resolution” means: tools, policies, and human approvals

   The founders define end-to-end as understanding company-specific policies and taking real actions (e.g., refunds) via integrations like Stripe. Sensitive actions can require Slack approvals, while the agent continues communicating like a human rep during pending decisions.

9. 9

   Reliability strategy: sampling, judging, tool validation, and continuous evals

   They explain why “tool support” is not trivial: you need guardrails, deterministic validation, and evaluation systems to avoid regressions—especially with sensitive tools like billing. Early on they manually monitored everything; at scale they rely on automated evals and monitoring.

10. 10

    Scaling in production: thousands of tickets closed per day and enterprise trust

    Parahelp now participates in thousands of daily ticket resolutions, with some customers trusting it even for large enterprise accounts. They describe a progression similar to training human agents—starting small and graduating to high-value conversations, including Slack-based support.

11. 11

    Knowledge and memory: the Parahelp Assistant, memory files, and org-wide adoption

    Beyond ticket handling, Parahelp becomes an internal knowledge hub. The Assistant proactively requests missing info, maintains “memory files,” prevents duplication, and enables teams across the company (support, sales, engineering, leadership) to query customer intelligence.

12. 12

    Next-gen agentic architecture: agents that generate evals and optimize policies

    They split the system into two major components: the Parahelp Agent (in the ticketing system) and the Parahelp Assistant (configuration/testing/research). With the arrival of a stronger model (Opus 4), the Assistant can generate evaluation sets on the fly—removing the forward-deployed engineering bottleneck.

13. 13

    Research flywheel and mass actions: improving resolution rate and proactive follow-ups

    A research agent analyzes large ticket histories (searching tens of thousands, deeply reading subsets) to produce reports and suggested policy/tool changes. This enables a continuous loop of discovering patterns, updating policies, testing, and publishing—plus bulk follow-ups to clear backlogs or reach users who requested features.

14. 14

    Demo: modes, diff-based policy edits, sandbox testing, and publish workflow

    They walk through the Parahelp Assistant UI: ask → configure → test → deep research modes, each with distinct prompts/tools. The Assistant edits policy “files” via diffs, generates realistic test contexts (including outage simulations), runs test agents, and only publishes when scenarios pass—turning support automation into a controlled release process.

15. 15

    Series A: why they chose Jack Altman / Old Capital and board-level partnership

    They describe raising an $18M Series A led by Jack Altman (Old Capital), with additional participation from new and existing investors. The deciding factor was alignment and high-trust collaboration—especially important because it’s their first board addition.

16. 16

    Founder advice: learn SV history, embrace action, and treat failure as fuel

    They close with advice to aspiring founders—especially those outside major startup hubs. They emphasize reading Silicon Valley startup history, building confidence through doing, and reframing failure as low-risk experimentation (especially in countries with strong social safety nets).