AVA is AI workforce for small businesses.
Problem
SMBs and solo operators often struggle to consistently find the right customers for their products.
A huge amount of time goes into prospecting, researching companies, qualifying leads, writing outreach and following up. Worse, much of that effort can be spent on people who were never a good fit in the first place.
I've experienced this firsthand.
I've sent hundreds of cold emails and DMs with little to show for them. I've spent money on Instagram, Facebook and Google ads that produced a few clicks, an occasional sale, and plenty of uncertainty about what was actually working.
The problem isn't simply getting more leads.
It's finding the right people and reaching them without spending an enormous amount of human time doing it.
The idea
AVA is an agentic workforce orchestration platform that lets SMB owners create AI agents that understand their business and perform specific operations on their behalf.
Instead of treating AI as a chatbot, I'm exploring a model where agents can own meaningful pieces of operational work.
For the first version, I'm focusing on marketing — specifically lead qualification and outbound email.
Why I'm building it
AVA came from a problem I've repeatedly experienced while building my own businesses and products.
Getting something built is one problem. Getting it in front of the right people is another. That led to a simple question:
What parts of this work could an intelligent system actually own?
Not write an email for me. Actually do the work. That question became AVA.
What I've built so far
AVA is still a work in progress, but the underlying agent system is already taking shape.
- Agent orchestration and subagent execution
- Tool calling
- Human-in-the-loop approval flows
- Pause and resume for long-running agent runs
- Observability
- Database persistence
- Integration tests around core execution workflows
Core workflow
An agent can execute a multi-step workflow, call tools and reach completion.
Approval-gated workflow
When an agent attempts to use a tool that requires approval, it generates an approval request and pauses the run. The user reviews the request and either approves or rejects the tool call. Once that happens, the run resumes, the tool call is resolved (executes on approval, ignored on rejection), the tool result is fed back into the LLM, and the run proceeds as usual.
Architecture
AVA uses an orchestrator + subagents architecture. The orchestrator owns the broader business context and acts as the interaction plane between the user and the agent workforce. Subagents own specialized tasks and remain unaware of unrelated capabilities.
For example, an outreach agent knows how to perform outreach-related work. It doesn't need to know that an accounting agent exists or what it does. This gives each agent a narrower responsibility while keeping business-level coordination centralized.
Human in the loop
Human approval is part of the architecture rather than a UI feature bolted onto the system. Some agent actions have consequences that justify review. For example, a business owner may want to verify that an email accurately represents their business before it is sent to a prospect. The system therefore allows an agent to reason about an action and prepare it without necessarily granting the agent unrestricted authority to execute it.
Product decisions
There is a huge amount that could be built around AI for small businesses. I deliberately chose not to start by building an AI-native operating system that handles every aspect of the business.
Start with marketing
I chose marketing because it creates visibility. A customer cannot buy from a business they don't know exists. And once customers come in, the rest of the business naturally follows. That makes customer acquisition one of the highest-leverage problems an SMB can solve.
Start with lead qualification
I chose lead qualification because it helps businesses capture more value from attention they are already receiving. It also creates information about the business's ideal customer profile. The better the business understands which prospects are worth pursuing, the better it can eventually target outbound efforts.
The progression I'm exploring is:
Capture attention → qualify demand → learn the ICP → target similar prospects → improve conversion.
Deliberately leave things out
For the first marketing slice, I chose not to integrate social platforms and let agents automatically publish content or respond to DMs. Those capabilities may eventually make sense. But they increase both product complexity and the number of ways an agent can behave incorrectly. I'd rather make one workflow genuinely useful than build a broad system that does everything poorly.
Early validation
AVA is still early, but it has already moved beyond a purely technical experiment. Three companies have expressed interest in beta testing AVA with their marketing workflows. The next question is no longer whether I can build the system. It's whether I can make it useful enough that another business wants to rely on it.
What's next
The immediate goal is to make lead qualification and outbound genuinely useful and reliable before expanding into additional workflows. That means improving:
- lead discovery
- qualification
- outbound
- approval flows
- observability
- evaluation
The larger experiment is figuring out what happens when a small business can delegate meaningful operational work to agents.