Why we built custom AI hires for our own therapy platform before selling them

Founders rarely set out to build internal tools with the explicit intention of commercializing them. Usually, you are just trying to stop the bleeding. In early 2024, our team was running Pasul, an online therapy platform based in Romania. We had a growing roster of licensed therapists and a rapidly increasing number of patients trying to navigate scheduling, billing, and preliminary intake questions. Building a two-sided marketplace is notoriously difficult, but managing the daily logistics of healthcare is entirely unforgiving. The administrative overhead was crushing us. We did not want to hire an army of human administrators because the margins in mental health care are tight, and we needed to allocate our revenue toward better practitioner payouts. That immense operational pressure forced us to start building operational AI employees from scratch. We had to automate our own business just to survive the growth phase.

operational bottlenecks at our romanian mental health platform, pasul

Running a mental health platform introduces a unique set of operational constraints that standard software cannot easily solve. At Pasul, our core product was the secure, reliable connection between a licensed therapist and a patient who needed immediate support. But around that core product sat a massive perimeter of logistical noise. Patients would email us at 11pm on a Sunday asking if a specific therapist had experience with postpartum anxiety or cognitive behavioral therapy. Practitioners would send us urgent Slack messages at 7am on Tuesday because a calendar sync issue had double-booked their morning slots.

Every single manual intervention required a human team member to open a database, look up a patient record, verify payment information, and write a careful, compliant response. When you are processing hundreds of these micro-tasks every week, the cognitive load on the founding team becomes unbearable. We found ourselves acting as highly paid, highly stressed dispatchers. We spent our days routing messages, issuing refunds for missed sessions, and putting out scheduling fires instead of improving the actual clinical experience or expanding our provider network.

The breaking point came in late spring when we realized our customer support response time had slipped to over twelve hours. In mental health services, making a vulnerable patient wait twelve hours for a basic administrative answer is unacceptable. It damages trust before the therapy even begins. We needed a system that could handle the intake and triage process instantly, consistently, and securely. Hiring more people would only add management overhead and training delays. We needed a programmatic solution.

the failure of out-of-the-box software wrappers in handling patient privacy

Our first instinct was to buy a solution rather than build one. The market was flooded with conversational AI tools and customer service widgets promising to automate support in minutes. We eagerly signed up for several of these out-of-the-box software wrappers, hoping for a quick fix. The results were disastrous.

The primary issue was patient privacy and data routing. When a user asks a sensitive question about their mental health history or medication side effects, you cannot pipe that query through a generic third-party API wrapper that logs conversation data for future model training. We needed absolute control over the data pipeline to maintain compliance. Second, the commercial tools were incredibly rigid. They operated like standard live chat widgets with a thin layer of language generation slapped on top. If a patient asked a complex question that required checking a therapist's specific credential database and cross-referencing it with real-time calendar availability, the off-the-shelf bots would hallucinate or default to a generic fallback message asking the user to email support.

We quickly realized that a generalized tool built to answer shipping inquiries for e-commerce shoe brands could not handle the nuance of healthcare administration. We needed agents that understood our specific database schema, respected strict data compliance rules, and knew exactly when to escalate a conversation to a human operator. The generic tools lacked the deep integrations required to actually do the work. They were toys, and we needed infrastructure.

prototyping systems for automated analytics and visibility updates

Since the market could not provide what we needed, we opened our code editors and started writing custom scripts. We began with isolated, single-purpose automations. Our very first prototype was not even a customer-facing bot. It was an internal analytics aggregator. Every morning, we needed to know how many new patients had registered, which therapists had open capacity for the week, and where our marketing spend was actually converting into booked sessions. Pulling this data manually from Stripe, our custom backend, and our marketing platforms took a founder two hours every single day.

We built a basic system that queried our databases overnight, synthesized the raw data, and generated an automated morning brief delivered to our team Slack channel by 8am. Getting those two hours back every morning felt like a revelation.

Once we saw how reliably a custom script could handle internal reporting, we moved to visibility and patient acquisition. We needed to ensure that when someone searched for specific therapy modalities in our region, Pasul showed up accurately in AI search engines and traditional results alike. We built a prototype to manage our structured data and monitor our digital footprint. This early system constantly scanned how our practitioner profiles were being indexed and flagged any discrepancies in our local search presence. It was crude, but it worked. We stopped flying blind. We had a reliable pulse on our operations and our market presence without spending hours compiling spreadsheets or checking search rankings manually.

what we discovered when real business operations ran on custom AI agents

The success of our internal reporting and visibility prototypes gave us the confidence to tackle the real bottleneck, which was direct patient communication. We built a custom intake agent designed specifically for the Pasul platform. We connected it directly to our secure PostgreSQL database, gave it strict boundaries on what it could and could not say regarding medical advice, and deployed it to handle initial patient inquiries.

The impact was immediate and profound. The agent could instantly answer questions about session pricing, therapist specializations, and availability. It could guide a new user through the preliminary intake form, collect their preferences, and slot them into a practitioner's calendar without any human intervention. We watched our average administrative response time drop from twelve hours to three seconds.

But we also discovered something unexpected about how users interact with highly competent, custom-built systems. Because the agent was specific, accurate, and deeply integrated into our operations, patients trusted it to handle their logistics. They did not treat it like a frustrating automated phone tree. They provided clear information, booked their sessions, and moved on with their day. On the backend, our therapists loved the system because it eliminated the administrative back-and-forth they previously had to manage between sessions. The founding team finally had the breathing room to focus on growth, clinical quality, and platform stability. We had successfully replaced an entire department of hypothetical administrative hires with a few pieces of deeply integrated, purpose-built software.

transitioning from individual custom code to a repeatable studio system

Other founders in our network started noticing how efficiently Pasul was running. They would ask us how we were managing such a high volume of patients and practitioners with such a lean core team. When we opened our laptops and showed them the custom agents we had built, the reaction was always the same. They wanted the exact same system for their own businesses.

We realized that the operational pain we experienced at Pasul was not unique to mental health platforms. Every founder-operated business hits a ceiling where administrative overhead chokes growth. Whether you are running a regional logistics company, managing a portfolio of real estate properties, or operating a multi-location retail brand, you eventually spend more time managing the system than building the business.

This realization prompted a massive pivot. We took the isolated scripts we built for Pasul and began architecting them into a unified, repeatable framework. We stripped out the healthcare-specific logic and built a modular system where we could inject any company's specific knowledge base, operational rules, and database integrations. We moved from building one-off internal tools to running a dedicated software studio.

We packaged the core functions we had perfected into distinct roles. Our internal reporting script evolved into our chief-of-staff AI. Our patient intake prototype became our customer service AI. Our digital footprint monitor matured