I Built a Personal AI Agent — Here's Where It Broke
Most people talking about AI agents have never actually shipped one to a team. I did — and the moment I tried, everything I thought I knew about building AI agents fell apart.
This is not a tutorial. It's an honest account of building a personal AI agent that worked brilliantly in isolation, hit a wall the moment real people touched it, and what that taught me about what enterprise AI deployment actually requires.
If you're a developer, technical founder, or engineering lead experimenting with agentic AI — this is the story you don't usually read.
The core difference: personal AI agents optimise for a single user; enterprise AI agents require auditability, RBAC, cost transparency, and modular skills — and these are architectural differences, not scale differences.
Table of Contents
- Why I Built an AI Agent in the First Place
- What the Agent Actually Did
- How I Built It: LLM as the Reasoning Core
- Where It Worked — and Why That Made Me Overconfident
- The Moment It Broke: Trying to Roll It Out to My Team
- What Happens When You Try Open-Source Alternatives
- Personal AI Agent vs Enterprise AI Agent: The Real Differences
- What Enterprise AI Agent Deployment Actually Requires
- FAQ
- Key Takeaways
Why Did I Build an AI Agent in the First Place?
My sales team was spending most of their working hours doing things that weren't selling.Before every outreach call, someone had to research the prospect — company background, recent news, leadership changes, funding rounds, likely pain points. Then update the CRM. Then track what was sent, when, and what was replied to. There were tools for most of these tasks — Apollo for prospecting, various scrapers for company research, news aggregators, LinkedIn. But each tool was completely disjointed. A rep would open five tabs, copy-paste between them, and spend 45 minutes on research before a 20-minute call.
That felt like exactly the kind of problem an AI agent should solve.
So I built one.
What Did the Agent Actually Do?
The agent took a single input — a company name or website URL — and did the following autonomously:- Searched the web for recent news and company context
- Called the Apollo API to pull contact and firmographic data
- Scraped the target company's website for positioning, products, and leadership
- Pulled industry research and relevant trends
- Synthesised everything into a structured research brief — use case fit, talking points, likely objections
That felt like magic. And for a while, it was.
How Did I Build It? LLM as the Reasoning Core
The architecture was simpler than you'd expect. At the centre was an agentic harness — a lightweight orchestration layer that could take a problem statement and string together the right tools in the right order based on interim results.Claude was the reasoning engine. Given a problem, it would decide which tool to call next, interpret what came back, and determine whether the answer was complete or whether another tool call was needed. Web search, API calls, website parsing — Claude coordinated all of it.
What is an agentic harness?
An agentic harness is an orchestration layer that connects an LLM to external tools and APIs. Rather than following a fixed script, the LLM reasons about which tool to call next based on what it already knows. This is what separates a true AI agent from a simple chatbot or automation.
For a deeper look at agent architecture patterns, see How to Design AI Agents: A Practical Architecture Guide.
The harness was built in Python. LLM handled the reasoning. The tools handled the execution. It took about two weeks to get to a state where the output was genuinely useful.
Where It Worked — and Why That Made Me Overconfident
For my own use, the agent was transformative.I could research 10 companies in the time it previously took to research one. The output quality was consistently better than manual research because the agent never got tired, never skipped a step, and always cross-referenced multiple sources.
The result: Research time per prospect dropped from ~45 minutes to under 2 minutes. That's not an improvement — that's a category change in how outreach works.
I was convinced. I wanted to roll this out to my team immediately.
That's when I hit the first wall.
The Moment It Broke: Trying to Roll It Out to My Team
When I tried to roll the AI agent out to my sales team, the problems surfaced almost immediately — and they had nothing to do with the quality of the AI output.Problem 1 — No traceability. When it was just me, I knew what the agent had done because I'd watched it run. The moment a team member used it, I had no idea what happened. Which tools were called? What data was pulled? What did the agent decide and why? There was no audit trail. In a sales context — where data accuracy and compliance matter — that was a dealbreaker.
Problem 2 — No cost visibility. Every API call costs money. LLM calls cost money. Web search calls cost money. When I was the only user, I could eyeball the bill. The moment multiple people were using it at unpredictable frequencies, I had no way to see costs per user, per task, or per team. I couldn't optimise what I couldn't measure.
Problem 3 — Skills kept breaking. Each new use case revealed a missing capability. A new industry needed a different research source. A different company type needed a different scraping approach. Every expansion of the agent's scope required code changes. There was no modular way to add new skills without touching the core harness.
Problem 4 — No access control. Every team member had the same access to everything. There was no way to say "this person can run prospect research but cannot access CRM write operations." For a solo tool, that's fine. For a team tool handling real customer data, it's a security gap.
What Happens When You Try Open-Source Alternatives
At this point I looked at open-source agentic frameworks. I tried one of the more popular options — I'll call it OpenClaw for brevity.The capability was genuinely impressive. But the security posture was not.
Running it felt like handing a junior employee a master key to every system in the company and saying "figure it out." There were no guardrails on what the agent could access, modify, or delete. No audit logging. No way to constrain scope. For internal experimentation, acceptable. For anything touching real customer data or production systems — deeply unsafe.
I was not going to open that to my team.
Personal AI Agent vs Enterprise AI Agent: The Real Differences
This experience made the distinction concrete for me. It's not a matter of scale — it's a matter of fundamentally different requirements.| Dimension | Personal AI Agent | Enterprise AI Agent |
|---|---|---|
| Users | 1 | Many, with different roles |
| Auditability | Optional | Non-negotiable |
| Cost tracking | Eyeballed | Per-user, per-task |
| Access control | None needed | RBAC required |
| Skill modularity | Hardcoded | Pluggable |
| Security | Trusted environment | Zero-trust required |
| Compliance | N/A | Regulatory requirement |
This is the gap most developers don't see until they try to cross it. Orchestrik was built specifically to close this gap for enterprise teams.
For a deeper look at how front-runner companies are thinking about this transition, see What Front-Runner Companies Are Doing to Scale Agentic AI Safely.
What Does Enterprise AI Agent Deployment Actually Require?
Based on what I learned building and breaking my personal agent, here is what I now consider non-negotiable for enterprise AI agent deployment:1. A defined use case with clear boundaries Don't try to automate everything. Pick one high-value process, define what the agent can and cannot touch, and start there. Scope creep in agentic systems is dangerous.
2. Auditability at every step Every agent action must be logged — what tool was called, what data was accessed, what decision was made, and by which agent on whose behalf. This is not optional in any regulated environment.
3. Role-based access control (RBAC) Different team members need different permissions. An agent acting on behalf of a sales rep should not have the same access as one acting on behalf of a system administrator. Access must be scoped to role, not just to user.
4. Cost transparency Enterprise deployments need to know what every agent run costs — per user, per task, per team. Without this, AI costs become invisible until they're a problem.
5. Modular skills New capabilities should be addable without touching core infrastructure. A pluggable skill architecture means the agent grows with the business instead of requiring a rebuild every time requirements change.
6. Security by design Data residency, encryption, and access controls need to be designed in — not bolted on. Especially critical for Indian enterprises dealing with DPDP compliance and clients who require on-premise deployment.
This is the architecture I eventually built toward with Orchestrik — an enterprise AI agent platform designed specifically around these requirements, for teams that need to deploy agents safely across real business operations.
For a governance framework you can take to your leadership, see Enterprise AI Agents: A Safe Governance Playbook for CXOs.
FAQ
What is an AI agent?
An AI agent is a system where a large language model (LLM) autonomously decides which actions to take — calling tools, accessing APIs, browsing the web — to complete a goal, rather than just answering a question. Unlike a chatbot, an agent can take multi-step actions without a human directing each step.How is an AI agent different from a chatbot?
A chatbot responds to inputs. An AI agent acts on them. An agent can call external tools, chain multiple steps together, make decisions based on interim results, and complete tasks end-to-end without human intervention at each step.Can I build an AI agent?
Yes. LLM APIs support tool use and multi-step reasoning, making it a capable reasoning core for an agentic harness. For personal or prototype use, it's relatively straightforward to build. Enterprise deployment requires additional infrastructure around auditability, access control, and cost management.What is the difference between a personal AI agent and an enterprise AI agent?
A personal AI agent is optimised for a single user's needs — speed and output quality. An enterprise AI agent must additionally handle multiple users with different roles, full audit trails, role-based access control, cost tracking per user, and security compliance. These are architectural differences, not just scale differences.What is RBAC in the context of AI agents?
RBAC stands for Role-Based Access Control. In an enterprise AI agent context, it means different users or roles have different permissions for what the agent can do on their behalf. A sales rep's agent might be able to read CRM data but not write to it; a manager's agent might have broader access.What is Orchestrik?
Orchestrik is an enterprise AI agent platform built for Indian businesses that need to deploy agents securely, with full auditability, RBAC, on-premise options, and modular skill architecture. It handles the infrastructure layer that most developers hit when trying to scale a personal agent to a team.Key Takeaways
- Building a personal AI agent is genuinely achievable and can deliver dramatic productivity gains for individual use
- The moment you try to roll an AI agent out to a team, four problems surface immediately: no traceability, no cost visibility, no access control, and brittle skill architecture
- Open-source alternatives offer capability but rarely offer the security posture enterprise environments require
- Personal AI agents and enterprise AI agents are not the same product at different scales — they require fundamentally different architecture
- Enterprise AI agent deployment requires: defined scope, full auditability, RBAC, cost transparency, modular skills, and security by design
- The gap between "it works for me" and "it works safely for my team" is where most agentic AI projects stall — and where purpose-built enterprise platforms like Orchestrik start
Social hook: Research that took my sales team 45 minutes now takes 90 seconds. But the moment I tried to roll it out to the team, everything broke — and none of it was the AI's fault.
References
- Anthropic Claude API Documentation — tool use and agentic patterns: https://docs.anthropic.com
- Apollo.io API Documentation: https://apolloio.github.io/apollo-api-docs
- DPDP Act 2023 — India's Digital Personal Data Protection Act: https://www.meity.gov.in
