The AI-Augmented Solo Operator Model: Six Months In
In June I published a log of my first 90 days running PMC solo on AI. Six months in, the core rule still holds. What changed is what I stopped trying to be, what I built instead, and the check every claim now passes before it goes out.
On this page
In June I published a log of my first 90 days running Pfender Marketing Co. as one person with AI doing the volume work. It was written close to the work, and it reads like it. Six months in, the core of it holds up. Some of it I would say differently now, and a few things I have stopped doing entirely.
This is the updated version. The original numbers are at the end, for the record.
The rule that held
The rule has not changed: AI handles volume work and research synthesis. I handle positioning, client-facing judgment, and anything where a wrong call compounds over time.
It started with an inventory. In my first two weeks I listed every task I was doing by hand on a recurring basis, 63 of them, and sorted each one into one of three buckets.
First drafts, data pulls, competitive scans.
AI first. I edit rather than originate.Strategic recommendations, client positioning calls, product decisions.
Me, in full.Reformatting deliverables, writing versions of the same email, maintaining systems that should run without me.
AI first. This is where the time came back.What changed is how literally I apply it. At 90 days the rule was a principle I checked against when something felt off. Now it decides the workflow before the work starts. If a piece of work ends in a recommendation that affects a client's budget or positioning, it goes through me in full, not as an edit on a draft. I learned that in month three, when a recommendation that should have been specific to one client went out generic, and the client noticed.
What I stopped trying to be
For the first few months I was building my own internal tooling alongside the client work: a CRM and pipeline system shaped around how I work. With AI coding tools it came together fast, and it ran.
What I had not appreciated was everything that comes after something runs. Maintenance, fixes, the next feature, the edge cases nobody planned for. I still use what I built for some things. But eventually I admitted what I had been avoiding: I do not want to be a software developer. I want to be the person who fixes a business's marketing.
So the pipeline, forms and booking for PMC now lean on an off-the-shelf platform, and the hours I was spending maintaining my own system go to marketing, for clients and for PMC itself.
The lesson was not that building it was a mistake. AI made building cheap. It did not make owning cheap. Every hour spent keeping my own infrastructure running is an hour not spent on the problem a client is paying me to solve.
What I built instead
The building did not stop. It moved to things a client or prospect actually touches.
The website. For as long as I have been doing this, I built sites the way most marketers do, on a platform: WordPress, Wix, Squarespace, and most recently Webflow. This month I moved pfendermarketing.com off Webflow onto a site that lives in a code repository and deploys on Vercel. Every page and post is a file I own, and the repository is connected to the AI tools I work with, so I can change anything on the site directly instead of working around what a platform allows. I directed the build; AI did most of the typing. How my whole toolset changed in six months is a longer story, and it has its own piece.
The site scanner. The free scan on the site grades a page in five areas: found on Google, found in AI answers, easy to call or book, fast on a phone, and knows what is working. It is the operating rule applied to a tool, and it took me until this month to apply it all the way.
Fixed rules already calculated the grades. But every finding a visitor saw, the two or three things worth fixing first, was written by a language model. On September 2, two scans of my own site came back with their grades and no findings at all, only a line saying there were not enough to feature. The model had failed, and the report had nothing to fall back on.
- Grades
- Fixed rules
- Findings
- A language model
- Model fails
- No findings shown
- Grades
- Fixed rules
- Findings
- Fixed rules, quoting the measured number
- Model fails
- Nothing lost
So the findings moved to fixed rules too. Each one appears only when the scan has the evidence for it, and it quotes the number it measured. The numbers should never depend on a model's judgment. Since September 28 the wording does not either: AI now only reads a screenshot of the first screen on a phone, and every word of the report comes from fixed rules. If you want to see a real result, I ran it on this site for the article on why websites do not get leads.
The problem written standards did not solve
In month three I fixed voice drift by writing the standard down: a brand voice document, now a linter, attached to every content workflow. That worked. The writing stopped drifting toward a generic register.
What a voice standard does not catch is a sentence that sounds right and is not true. This month I had the whole site and blog reviewed by an outside editorial pass. It found platform facts that had changed since publication, a statistic quoted without the limits of the study behind it, and claims about my own work that did not match my records. I corrected the facts, took one post down, and every revised post now shows the date it was updated.
Even that did not catch everything. A day later I had every number in one post checked against its original source, and one sentence said the opposite of what the study found.
The rule that came out of it is simple.
- A fact about a platform or a studyNeeds the primary source: the vendor's own page, or the study itself. Not someone's summary of it.
- A claim about my own workNeeds a record I can point to.
- NeitherThe sentence goes.
What the model still does not do well
The same three things as at 90 days, and I have not seen that change.
It does not handle a client situation I have not seen before. The output is a competent summary of adjacent situations, useful as input and not sufficient as a recommendation. The judgment about which adjacent situation is truly analogous has to come from me.
It does not repair a relationship. When an engagement is going wrong partly because expectations were never set, the model can help draft the message. It cannot tell me what the relationship needs.
It does not produce differentiated positioning. Ask it for positioning options and you get competent, interchangeable ones. The position worth building comes from what is specific to that business, and finding that is the work.
What the first 90 days said
At 90 days I put three numbers on the change. All three compared real post-AI output against a remembered baseline, not a logged one, and I said at the time which ones I would defend.
What was countedDeliverables shipped per week, each reviewed before it went to a client.
The baselineMy recollection of the same measure before AI. Not logged.
What was countedTime to complete recurring tasks.
The baselineRemembered. Part of the gain is work that disappeared rather than got faster.
What was countedProspects surfaced by the outreach system.
The baselineRemembered, and what that coverage is worth was not known.
I said then that I would start logging throughput properly. Six months in, I do not have a figure I would stand behind in public, so there is no new number here. When there is one, it will come with the baseline and the method next to it, which is the standard I would hold anyone else's number to.
If you are running solo
The advice from 90 days stands. Look for the volume work you have filed as judgment work because it is attached to something you care about. That is where the first leverage is.
Two things I would add at six months. Write down what "right" looks like before you let a model produce more of it, both for voice and for facts. And be careful what you build for yourself. AI will build it for you quickly. It will not maintain it for you.