Andrew Babkin, Azati. Building production software & AI for regulated industries for 24 years.
Have you ever made a mistake that cost you hundreds of thousands of dollars and three years of your team’s work and, still, if asked whether you would do it again, you’d say “yes”? That’s the leader’s spirit.
We build enterprise AI for a living. Three years ago, I turned my company into an AI lab; before selling AI transformation to clients, we would go through one ourselves. We became part of the AI failure statistics, recovered and ended up with something better than we had planned.
Thirty Months, 24 In The Wrong Direction
The plan was to put an AI layer over billing, staffing, pipeline and cash flow. For roughly 24 of the 30 months, it took to get somewhere useful, we were actually building yet another CRM. We weren’t alone: In an April 2026 Gartner, Inc. survey of 782 infrastructure and operations leaders, only 28% of AI use cases fully met ROI expectations.
The clearest symptom was my favorite feature, CV-Tinder, a module that ranks available engineers for a role on a 100-point scale. Almost nobody used it. A 30-second message to a tech lead felt easier. I had pushed it out partly to show motion.
The fix wasn’t a redesign. We moved the AI into the work chats where people already spend their day, and that decision reshaped the whole system.
An Operating Layer, Not A Chatbot
Today, the system sits above everything we run: billing, CRM, the task tracker, Google Workspace, GitLab and the staffing tool. Each user gets a personal AI agent plugged into that layer. When a salesperson receives documents for legal review, their agent notifies a legal chat, opens a task with the files attached, assigns our lawyer and sends the email.
The part I’m most curious about went live in October 2026. As people work, their agents turn repeated tasks into reusable skills. A supervising agent reviews them and promotes the valuable ones to a company catalog, so a new hire’s agent can use on day one what a colleague’s agent learned over months. In the first days, one engineer’s agent alone contributed 29 skills.
To be clear, we haven’t finished rolling it out even to ourselves. Meanwhile, we grew almost threefold over those three years, and many of our engineers learned AI on a live system rather than on a client’s project.
What Technology Leaders Can Take From This
1. Adding A Layer Instead Of Replacing Systems
Connecting AI to the platforms you already run meets far less resistance than ripping them out. Modern systems with APIs connect easily. For older ones, an agent can learn by watching how people click through them. It’s slower, but it works.
2. Putting AI Where People Already Work
Your real rival isn’t a legacy platform or a skeptical stakeholder. It’s habit. If people have to open a new tool to get an answer, most won’t. Map where your teams ask questions today—whether in chats, by email or on quick calls—and put the AI there.
3. Letting Autonomy Grow Gradually
Start with the agent working alongside a person and then doing the task itself and escalating whenever it’s unsure. Each employee decides when to hand over more. Decisions that carry legal or financial responsibility stay with a human.
4. Making Adoption Your First Metric
A system that’s right 95% of the time and that nobody opens is a failure. Before launch, agree on who should use the tool, how often and for which task. Review usage weekly, and when it drops, find out why within days. It’s usually a workflow problem, not a model problem.
5. Not Releasing To Show Momentum
This is the lesson I struggle with most. A release made for optics costs you twice: once to build it, and again to win back trust after nobody uses it. Before every launch, ask: Who will use this next Monday, and what will they use it in place of?
6. Experimenting Widely
Testing on yourself isn’t only for smaller players. According to a Wall Street Journal report, Johnson & Johnson let employees experiment broadly with GenAI, an approach that CIO Jim Swanson described as letting a thousand flowers bloom, and ended up with nearly 900 use cases. A review showed that 10% to 15% of them drove about 80% of the value, and J&J narrowed its focus to those. Experiments are cheap only if you’re disciplined about shutting them down.
Testing Everything On Yourself First
I sell AI transformation, so I run it on my own company first. That’s how a provider earns the right to advise. I know how to save money, mitigate the risk and avoid the bruises.
How have your own in-house AI efforts gone? What surprised you most?
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?


