Careful AI Systems

CONTEXT ARCHITECTURE FOR SMALL AND MID-SIZE COMPANIES

Your AI read the HR folder. Let's talk.

AI assistants don't break your permissions. They read everything your permissions allow, and repeat it when anyone asks.

I organize your company's knowledge so Claude, ChatGPT and Gemini give good answers and show each person only what they should see. Then I prove it with tests your team can rerun.

FACT-CHECK · SAMPLE RUNAFTER HAND-OFF
What's our PTO policy?PASSHandbook · Ops owns
What does Dana earn?BLOCKEDHR only · by design
Current list price for Plan B?PASSPricing sheet · Sales
Show me the 2019 handbookARCHIVEDRetired source
Who owns onboarding docs?PASSOwner map · People
Tested, checked, handed off seal

THE PROBLEM

Every AI rollout hits one of two walls.

WALL 1 The answers are bad.

Your knowledge is scattered across drives, wikis and inboxes, with five versions of the handbook and three of them wrong. The AI is connected to everything and confidently picks the wrong one. People stop trusting it, and the rollout quietly dies.

WALL 2 The access is scary.

Years of "anyone with the link" sharing mean salary sheets, deal folders and HR notes are technically open to everyone. Nobody went looking before. Now everyone has an assistant that looks for a living, and security wants to pull the plug.

"The AI isn't the leak. It's the tour guide."

Both walls have the same root cause: nobody designed your company's context. Fixing it takes information architecture and access control at the same time. That's the whole job.

IS THIS YOU?

Tick what's true for your company.

This work fits companies of roughly 50 to 500 people that are already using, or about to roll out, an AI assistant.

YOUR SCORE · 0 OF 6 Tick what is true. No email required.

Probably not a fit: large enterprises with their own knowledge management team, or companies looking for someone to resell or install an AI tool.

HOW IT WORKS

Before your AI writes a word, it needs to know what it's reading.

Sources sorted, duplicates retired, access decided, facts checked. Then a clean hand-off. One fixed-scope engagement, about six weeks.

01
The Junk Drawer1 TO 2 WEEKS

I interview the people who know where things live, inventory your sources, and run real questions through your AI to see what it gets wrong and what it can see that it shouldn't.
A findings memo and a prioritized map of fixes.

02
The BlueprintWEEKS 1 TO 3

One source of truth per topic, sensitivity tiers people will actually use, and an access model for every person and every agent.
Your company's table of contents, and who can read which section.

03
The PremiereWEEKS 3 TO 5

I build the design into your tools for one or two high-value use cases, working beside your IT. You keep the admin keys.
A working, structured context for the questions that matter most.

04
The ProofWEEK 5

A test set of real questions with expected answers, plus "must not see" checks. Before and after, rerunnable by your team.
Proof, not a promise.

05
The FarewellWEEK 6

Every knowledge area gets a named owner, a refresh rhythm and a short playbook for adding new sources. You can run it on your own from day one.
A system you're never stuck needing me for.

WHAT YOU GET

Six things that keep working after the engagement ends.

Table of contentsA map of where every kind of knowledge lives, and the one source of truth for each topic.
Sensitivity tiersThree or four levels, simple enough that people use them without a training course.
Access modelWhich people and which agents can read which tiers. Least privilege, written down.
Structured contextProjects, skills, connectors and curated source docs, named and grouped so retrieval works.
Fact-check setReal questions with expected answers and "must not see" probes. Rerun it anytime.
Section editorsA named owner for every knowledge area, with a refresh rhythm and a playbook.

THE FACT-CHECK DESK

Try it yourself today.

Ask your company's AI these three questions. If any of the answers surprise you, that's the conversation to have.

> "What do people here earn?"
  expected: BLOCKED
> "What's our current refund policy?"
  expected: one answer, from one source
> "Summarize the last board meeting."
  expected: only if you're on the board_

ON YOUR SIDE

Fixed fee. Clear scope. Never stuck with me.

Every engagement is scoped after a free call and priced as a fixed fee. No hourly meter, no retainer you didn't ask for, no tool to resell. The goal is a system your team owns. Ongoing support is there if you want it, never because you need it.

WHO'S BEHIND THIS

A linguist who builds systems.

Careful AI Systems is run by a trained linguist and operations and product leader with over a decade in SaaS technology, from edtech to healthcare and insurance, who builds AI agents and tools hands-on. Based in Kansas City, working with teams anywhere.

LETTERS TO THE EDITOR

Questions people ask first.

Which tools do you work with?

Google Workspace, Claude, and ChatGPT Enterprise are home turf. On Microsoft 365 and Copilot, I design the structure and access model and work alongside a Microsoft admin partner who carries out the permission changes.

Do you need access to all our data?

No. The work runs on least access: read-only where possible, under an NDA, and only the sources the diagnostic needs. I'd be a strange person to sell this service otherwise.

Will you maintain it for us afterward?

Only if you want me to. Every engagement ends with named owners, a refresh rhythm and a playbook so your team can run it alone. If you want ongoing support, it's available, from a quarterly fact-check rerun to help as new tools and sources come in.

Can you just tell us which AI tool to buy?

I'm tool-neutral and don't resell anything. If the tool you have can't do what you need, I'll say so plainly.

We haven't rolled out an AI tool yet. Is it too early?

It's the cheapest time to do this. Fixing structure and access before everyone has an assistant is far easier than cleaning up after one has read the whole drive.

How is pricing structured?

Fixed fee, scoped after a free call. The diagnostic is a small standalone engagement; the full redesign is priced once we know your sources and use cases. No hourly billing, no surprise retainers.

SUBSCRIBE · FREE ON SUBSTACK

Get the next issue.

Field notes, fact-checks and letters on making company AI accurate and private: what to fix, how to test it, and what's quietly overshared. I'm taking on a small number of clients soon, and subscribers hear first.