Michelle
Jan 24, 2026 - Aug 18, 2026
9
Decisions
0
Active Todos
3
Patterns
Categories
Decisions (9)
Extend the existing Claude PII policy unchanged to the new AI systems - and settle that HubSpot data is not restricted customer data
Brady Dibble raised a data-access question about additional AI systems on 2026-08-15, treating the data he had lost access to as uniformly restricted. Peter refused to answer it per-system. He ruled that (1) the things you lost are not all the same - some are restricted and some are not, (2) Hubspot data is not customer data, and a list of customers is just fine while PII is not, (3) as far as I am concerned the same rules can apply to the two systems you just listed, and (4) there should be no barrier to moving approved data to them, but they should not open the door to more data than we currently allow to Claude. He pointed Brady at the existing policy, noting Michelle has it.
Institute the inspectable-function rule for AI data access as company-wide hygiene - and refuse to make it either a shared library or retroactive
At the Friday AI Committee, Peter re-delivered the rule he had set once before and watched not land. He used Mini-Me as the worked example: AI helped create the capability, but AI is not part of the runtime function of accessing the repositories. Claude cannot get to Slack except through that function. It does not have the keys. He asked for exactly one thing - if you are pulling data in or out of a CIQ system, do it through a function that you write and could inspect - and explicitly declined two obvious extensions. He would not make it a shared implementation, because there is no value in building a library of these things and because forcing an MCP server would block people from experimenting with technology being stood up today. He would not apply it retroactively, because a year from now we are going to have 100 times as many tools running around the company, so he is way more worried about the giant tsunami of stuff that is coming. He named three reasons rather than one: inspectability, 100 percent repeatability, and teaching the company the right pattern, since Claude is going to make everybody at the company a developer and that is incredibly terrifying if we do not teach people to be better developers. He extended the bar to local models when asked, and licensed Michelle to publish the message with his name attached.
Block the AI-cost workstream on complete per-individual measurement before any optimization - so each persons correct plan can be determined
In the 7/23 Steve Wallace 1:1 Steve reported several threads coming out of Mondays AI meeting: efficiency coaching for heavy users, model-switching guidance, and usage reporting to department heads. Peter cut across all of it and set a sequencing requirement - first get a dashboard that shows every individual in the company, including himself. His evidence was his own absence from it: he shows as zero on every cost dashboard despite Claude running 8 to 10 hours a day, with CIQ apparently paying around 200 dollars a month for him. He rejected the manager-level rollup as a starting point and rejected pushing justification down to department heads. All of the costs, not just some of them.
Require all agent/LLM access to systems of record to go through inspectable static code, never direct LLM access
Peter opened the 7/24 Jira Automation call with Karl Lowenbjer by setting a condition before taking Karls agenda. The rule: no LLM gets direct access to a system of record. In every instance the LLM writes code that accesses the system, and that code is inspectable by a human or by another LLM. He stated he wants this true for everything Karls team builds, and that as long as it is true he is comfortable. He explicitly accepted that the filters themselves may be wrong - we might make a mistake in terms of what we pull in and out and that is fine, I do not mind mistakes - but the mechanism is non-negotiable. Karl confirmed Atlas already works this way. Peter then closed the topic without further discussion.
AI/Data Security Audit Commitment to Greg
When Greg raised concerns about CIQ leaking data through AI agents/bots/services, Peter committed to getting Michelle's oversight team to do an assessment/audit of what's running and with what access.
Gate Fuzzball SaaS GTM on Working Prototype — Assign Michelle
Agreed to move forward on Fuzzball SaaS GTM but gated it on having a working prototype first. Suggested Michelle as the resource to stand up a prototype on top of existing Fuzzball, accepting she's not the ideal person but can make progress. Rejected doing it 'out of band' — keep in normal product flow.
Shared Mini-Me Source Code to Internal Org
Shared Mini-Me source code by creating repo at ctrliq/min-me in CIQ GitHub org, after Ryan, Nathan, and Michelle independently asked for it. Proactively noted never having seen the code, setting quality expectations. Requested internal-only repo visibility.
AI Governance Single-Track Pivot for ISO 42001
Pivoted AI governance from dual-track (internal vs products) to single rigorous model because CIQ products (RLCAI, Fuzzball, Werewolf) now directly integrate AI, changing the liability profile.
Require mandatory tagging of all fully AI-generated content
AI Committee established policy that all fully AI-generated content must be tagged to manage user expectations. Applies only to fully AI-generated content, not human-reviewed or AI-assisted work. Format and placement of tags is flexible.
Related Patterns (3)
Lead by Example with New Tools
When championing new tools or processes, personally use them and share results rather than just advocating. Learning by doing and demonstrating value through example is more effective than mandates.
Protect Engineering Focus Through Process
When faced with requests that would disrupt engineering focus (from sales, governance, product, or other stakeholders), establish processes that protect engineering ability to innovate while still satisfying legitimate concerns. Prefer systematic solutions over ad-hoc responses.
Constrain the Mechanism, Not the Access
When AI or automation raises a governance concern, Peter does not restrict who or what may reach the data. He constrains HOW the reach happens, so the pathway becomes auditable and deterministic. The failure mode he designs against is non-determinism, not leakage - a wrong filter in inspectable static code is a bug you fix once, while a model that sometimes pulls the wrong field is an unbounded, unauditable liability. He will therefore accept imperfect filtering and eventual leakage as the price of adoption speed, but will not accept a non-inspectable pathway.