AI workflows for research, analysis and executive briefs.
KLJ Ventures is my AI systems practice, and I designed, built and operate the system of 120+ specialized AI agents that does its research, analysis and writing. Its clearest work product is a client deliverable, a strategic growth blueprint of roughly 60 pages with an accompanying executive briefing, and I approve every output before it leaves.
- Role
- Founder. I designed the architecture, built the system, and operate it daily.
- Dates
- 2026 to present
- Scope
- 120+ specialized AI agents across 14 functional domains, in daily use within my practice. A one-person firm.
- Outcomes
- A client deliverable: a roughly 60-page growth blueprint and an executive briefing. 2 analytical judgment calls from that work, described below.
- Evidence
- Adoption and business results belong to the client.
What the system produced for a client
The blueprint and briefing came from one client engagement. 2 calls from that work are described here without the client’s name or figures, because the reasoning is the part that transfers.
An acquisition cost with an undefined denominator
In client analysis I held an acquisition-cost figure as a channel cost until its denominator was defined. A cost per acquired customer means little until it is clear who counts as an acquired customer, and labeling it a channel cost kept the figure in the analysis without letting it stand for more than it showed. In the same section, I drew months with no reported data as a dashed line, so the chart did not suggest a trend for months nobody had measured.
3 ways to reach one revenue target
In a client analysis I laid out 3 illustrative revenue compositions that would each reach the same target. Each implied a different company, hiring plan and financing story, and they were labeled as illustrations, not forecasts or recommendations. The shortest path to the target and the stronger financing profile appeared to belong to different compositions. I think naming that tension was the most useful thing the section did.
The blueprint also labels claims by how well they are supported, so a reader can tell a verified figure from a company-reported one or an open question.
Work I would put my name on
KLJ Ventures is a one-person firm, so the volume of research and writing has to come from a system. Getting an AI model to produce a draft is easy. Getting it to produce work I would sign, repeatedly, is harder. An assistant that is right 90% of the time and gives no signal about which 10% is wrong still costs 100% of the review.
Using AI to write more first drafts is where I started, and it only produced faster first drafts. The problem worth solving was making review reliable enough that my own attention could go to judgment.
How work moves through the system
An agent here is a role defined in its own file, with a name, instructions, permitted tools and a model tier, that Claude Code runs when work is routed to it. I designed and operate 120+ of them across 14 functional domains. Work moves like this:
- A request is classified by type and routed to the domain that owns it, where the research starts.
- Specialized agents draft the analysis or the document.
- The agent that drafts is not the one that approves. Separate reviewers check the work, and on anything consequential a reviewer is instructed to look for reasons to reject it.
- In my most heavily used pipeline, 150+ automated checks are defined for written deliverables, and each deliverable runs the checks that apply to it before it reaches my review.
- I approve every output before it leaves the system.
- When something fails in real use, I document it, work out which class of error it belongs to, and encode the fix as a check or a written rule, so the same class of error is caught if it recurs.
- Separation
- The agent that drafts is not the one that approves.
- Traceability
- A claim that does not trace to the canonical source is rejected.
- Authority
- A human approval gate everything passes through before it leaves.
A request arrives and is classified by type, urgency and which domain owns it.
Nothing starts until the system knows what kind of work this is.
The check count covers automated, deterministic checks only. Reviews by separate review agents and the written rules the agents follow are not counted in it. Most of the checks record, in their own documentation, the failure that produced them. Checks and reviewers reduce risk, which is why the last step is mine.
The full map of the 14 functional domains, and the agents each one runs, is on the homepage.
Treat every fabrication as a bug
The common framing is that language models hallucinate, that this is inherent, and that the person reading the output should check it carefully. That framing locates the problem in the technology. I located it in the system built around the technology, because that is the part I can change.
I decided to treat every fabrication as a bug: a specific defect with a cause, a reproduction case and a fix that could be encoded, so the same class of error is caught if it recurs.
The tradeoff is cost. Every check adds time to each run and something to maintain, so the set of checks mostly grows, and checks that stop earning their cost get retired.
What the evidence shows
- A client deliverable: a strategic growth blueprint of roughly 60 pages and an accompanying executive briefing designed to be read in one sitting.
- 120+ specialized AI agents across 14 functional domains, in daily use within my practice.
- 150+ automated checks defined for written deliverables in my most heavily used pipeline.
These figures describe how the system is built and what it produced. The evidence is the client deliverable and the counts above. Adoption and business results belong to the client. The system is operated by one person, for one practice.
Design the checks in from the start
The first version treated review as a light final step. The rebuild put review and automated checks at the center, and I would design them in from the first version. It is the same conclusion I reached about the intake system and meeting cadence on the AMC Networks account: put the structure in place before the problems make the case for it.
Hiring for a Chief of Staff or Strategy & Operations leader?
Let’s discuss the role.