How we judge the pilot: how often customer success managers act on the recommendations, and whether flagged accounts recover more often than similar accounts without one. Targets are set before launch.
A shared platform that lets each department build AI assistants that draft, research and recommend, with access, privacy, quality and cost controlled in one place rather than separately in every project.
Why: act on churn risk and growth signals earlier, give staff back the time they spend gathering context, and stop each team buying and securing its own AI tools.
Six layers. Requests from Claude and from the assistants we build pass through the checkpoint in layer 2.
Slack, Claude, Microsoft 365 Copilot and reporting. No new app to learn.
Checks who is asking and what they may see, removes personal data, blocks attempts to trick the AI, checks answers against their sources, and records each step and its cost.
Built by departments from approved templates. Each has a named owner and an off switch.
Playbooks, product knowledge, our internal policy knowledge base, and account history, limited to what the person asking may see.
Claude, running in the company's own AWS account. Simple tasks use cheaper models.
CRM, call summaries, billing and product usage, defined once so “ARR” means the same everywhere.
Limits: in Claude chat the checkpoint can block a message but cannot remove personal data from it. Copilot stays under Microsoft's own controls. No-code assistants are limited to internal data until they are routed through the checkpoint.
The pilot, step by step.
How we judge the pilot: how often customer success managers act on the recommendations, and whether flagged accounts recover more often than similar accounts without one. Targets are set before launch.
Four scheduled reviews of the platform's own work. Each looks back further than the last. None of them changes anything on its own.
Reviews yesterday's results, rejected suggestions and user edits. Proposes specific fixes.
Spots repeated problems across teams, such as the same edit made again and again or rising cost.
Compares quality and cost against real outcomes, such as accounts that recovered.
Decides which assistants to expand or retire, and checks whether past fixes actually helped.
Claude Enterprise, ThoughtSpot's AI, Tray's built-in AI and parts of Databricks process data outside those regions. Customer product data is excluded by default and enters only with an approved project, then stays in-region. Our own CRM records, including EU and Australian contacts, are processed in the US under our existing transfer terms.
A web page or email can hide instructions aimed at the AI. Everything it reads is screened before use.
Each vendor's terms on keeping and training on our data are checked before confidential data reaches it. For AWS, a written opt-out from using our content to improve its services is a gate in the first phase.
During the pilot, spending is limited to Customer Success and every assistant run has a usage cap. From the rollout, each department has a budget with an alert at 80% and a stop at 100%. Only the accountable executive can raise a budget.
One deletion service removes a record from every copy the platform holds within 30 days, with proof.
First measurable results by the end of month 8.
Claude Enterprise, sign-in and access groups, safety checks, activity records.
The checkpoint, company data and definitions, the policy knowledge base connected, in-region processing ready for customer product data if a project is approved.
Recommendations for customer success managers in Slack, nightly and weekly reviews, go or no-go decision.
Marketing, Finance and Support build their own, with department budgets and the monthly review. First quarterly review in month 12.