How it works
We talk. I build. You approve.
The whole engagement runs on one weekly rhythm. Your side is about 2 hours a week. Mine is the rest.
The weekly rhythm
One cycle, repeated. Each round the system knows more.
1
We talk
One call a week. We pull what you know out of your head: procedures, product rules, where you research, how you decide. I record nothing you do not approve; everything lands as written notes you can read.
2
I build
Between calls I find and test the tools, construct the agentic knowledge base, and wire the agentic workflows to your real accounts: email, files, store, inventory. The system works from real orders, not pasted screenshots.
3
You approve
You review what the system drafted, in minutes, not hours. Anything wrong gets corrected in the knowledge base, so the same mistake cannot happen twice. Everything syncs to your computer and stays yours.
The five stages
Behind the rhythm, a repeatable playbook.
- DiscoverIntake, access, and a map of how you actually work today: every workflow, every tool, every place the business depends on your memory.You get: the map of how your business runs.
- Audit and prioritizeThe 3-5 highest-return opportunities, ranked by impact against effort. You see the roadmap before anything gets built.You get: the ranked roadmap.
- BuildThe agentic knowledge base plus the agentic workflows, one at a time, each documented as an SOP with a human checkpoint.You get: the working build, documented.
- Train and hand offRecorded walkthroughs and written SOPs. The test of done is that you can run it without me.You get: the walkthroughs and SOPs, yours to keep.
- OptimizeIf we keep going: monitoring, tuning, and new workflows as the tools evolve, scoped one phase at a time.You get: each phase scoped and priced in writing before it starts.
Two ways to work with us
Learn it, or have it built. Usually both.
Month one is the same engagement either way; we set the mix on the first call and adjust as we go.
Coaching-forward
You drive, I navigate.
The weekly calls become working sessions: you build with my hands off the keyboard, so the skill stays with you. Best if AI fluency is itself the goal.
Build-forward
I build, you approve.
I do the construction between calls; your job is the extraction interviews and the reviews. Best if the hours back are the goal and the tooling is not.
Either way the price is the same, everything is documented, and everything is yours. Not sure which you are? see what changes in 30 seconds, then bring the question to the call.
What gets written down first
What a knowledge base actually is.
A written brain for your business. Product rules and compatibility. Compliance requirements the system always reads before it replies. Lead times, policies, dealer terms, pricing logic.
every one of them reads from
Your voice, trained from your real emails. All of it stored as plain files you can open, read, and edit.
This is what stops the AI from ever guessing. When you ask "does this fit?", the answer comes from your fitment file, with a citation, or the system says it does not know. There is no third option.
- Plain files, not a platform. No proprietary database, no export fees, no lock-in. If we stop working together, you keep everything and it keeps working.
- It compounds. Every correction you make during review becomes a permanent rule. Month three is dramatically better than week one, automatically.
- It is the foundation for everything after. Ads, content, and agents all read the same brain. Build it once, use it everywhere.
- Add agents, and it becomes an agentic knowledge base. The written brain plus agents that read it and run agentic workflows, multi-step jobs drafted from your files for your approval. The files keep the agents honest; the agents turn the files into work done.
A real file, open it
It is a file you own. Here is one.
One page from a working knowledge base, in three kinds of business. Plain files your AI actually reads, and so can you.
Pick a business
Ask it something a customer would
This is one page from a business's files. It opens like any document, and the AI reads it before it answers.
# Fitment: MX-2 handlebar kit
## Fits
- Sur-Ron Light Bee X (2021-2026): direct bolt-on
- Talaria Sting R (2023-2026): requires riser adapter RA-11
- E-Ride Pro SS: direct bolt-on, confirmed by teardown photo (files/teardown-eride.jpg)
## Does NOT fit
- Talaria Sting L1e (road model): stem diameter differs. Do not suggest workarounds.
- Rawrr Mantis: untested. If asked: say untested, offer to check with supplier (2-3 days).
## Reply rules
- Always name the exact model and year back to the customer before confirming fit.
- If the bike is not listed here, the answer is "not written down yet," never a guess.
- Lead time if out of stock: 12-18 days (supplier-lead-times.md).
Composites of real client files, anonymized. This is the whole trick: plain files, readable by you, binding on the AI. When a customer asks "Does this fit my 2023 Talaria Sting?", the draft reply is built from this file and cites it, or says the answer is not written down yet. There is no third option.
What protects you
A system, not a favor.
Grounded, never invented. Every answer traces to a source in your files, and no claim ships that cannot be backed up.
Verified before it counts. Nothing is done because it looks right. It is tested working, on your real data, before I report it.
Documented and yours. The system is written down as we build it, so it is never trapped in my head, or yours.
Questions
The honest answers.
How much of my time does it take?
About 2 hrs/wk: one call plus quick reviews. If you cannot give it that, the system still gets built, just slower, because the extraction calls are where your knowledge comes out.
What stops the AI from making things up?
Structure, not hope. Every draft is grounded in your files, cites its source, and waits for your approval. When the knowledge base has no answer, the system says so instead of improvising. Corrections you make become permanent rules.
What do I need to provide?
A Claude plan, from $17/mo billed annually ($20 monthly), moving up only when you are using it fully. Claude is the AI assistant made by Anthropic; if you have used ChatGPT, it is the same kind of tool.
Plus access to the tools you already run, and about 2 hrs/wk. Most owners find their total software spend drops, because the system replaces $300-400/mo point tools.
Do I own what gets built?
Yes, completely. Plain files on your computer. Every workflow documented. If we stop after month one, you keep a working foundation and the knowledge of how to run it. Every vendor wants to own your business's brain; this is built so you can leave, which is exactly why you can trust staying.
What happens when it gets something wrong?
You catch it at the approval step, because nothing sends without you, and the fix becomes a permanent rule in the knowledge base. A correction made once holds forever, so the same mistake does not come back next week. Most tools need telling twice; this one is the first that does not.
What goes wrong with projects like this?
Two things, and naming them is fair. First, verification debt: if you approve drafts without reading them, errors compound. The review step is small but real.
Second, most AI projects stall because nobody owns the system after setup. That is why the handoff stage exists and why everything is documented. IDC research finds roughly 88% of AI pilots never reach production, and the cause is governance gaps, not the AI.