.png?width=172&height=34&name=JSC%20(1200%20x%20240%20px).png){kind=small}
On Episode 2 of the Funnel Vision podcast, I interviewed Jordan Blackwell, co-founder of Philly-based AI startup Siah Labs. A few days later I caught his Philly Tech Week presentation with co-founder and wife Deanna, where they outlined a process to help owners avoid the fate of 92% of businesses — closure — and instead reach independence, succession, or sale.
The opportunity is massive. But as I left the Fitler Club, one thing was nagging at me. To make a business legible to AI, you have to feed it the operational details that power the insights — pricing logic, client lists, methodology. There are real protective paths: enterprise contracts with zero-retention APIs, dedicated AI security platforms, or consultancies like Siah Labs that set standards for what to expose and keep humans in the loop on sensitive material. But most founders won't pursue them, and will instead use consumer chatbots and feed them whatever is necessary to get the job done. Could a competitor get their exact data? No. Will they be able to approximate trade secrets more easily using AI? Undoubtedly.
There's a name for this tension — strategic illegibility — and an architectural answer: a digital twin. Founders at Twitter, Zapier, and HubSpot have all championed the pattern. It's a structured replica of your business, useful enough for analysis and automation, stripped of the identifying details that make the underlying data sensitive.
The opportunity is massive. But as I left the Fitler Club, one thing was nagging at me. To make a business legible to AI, you have to feed it the operational details that power the insights — pricing logic, client lists, methodology. There are real protective paths: enterprise contracts with zero-retention APIs, dedicated AI security platforms, or consultancies like Siah Labs that set standards for what to expose and keep humans in the loop on sensitive material. But most founders won't pursue them, and will instead use consumer chatbots and feed them whatever is necessary to get the job done. Could a competitor get their exact data? No. Will they be able to approximate trade secrets more easily using AI? Undoubtedly.
There's a name for this tension — strategic illegibility — and an architectural answer: a digital twin. Founders at Twitter, Zapier, and HubSpot have all championed the pattern. It's a structured replica of your business, useful enough for analysis and automation, stripped of the identifying details that make the underlying data sensitive.
Every implementation I found assumed enterprise budgets. Palantir. Salesforce Data Cloud. Snowflake. Nothing for the agency on HubSpot or the SMB founder who's never written SQL. So I built one.
 |
This is the "Digital Twin" for SMBs. Airtable, Make.com, Claude via MCP. Forty dollars a month. An afternoon to set up the basics, a weekend to expand it.
|
 |
Here's the walkthrough. |
The friction to plug AI into your customer data has dropped to nearly zero. At the same time, both major consumer AI tools defaulted to training on user content this year. ChatGPT Free, Plus, and Pro all default to data sharing on. Anthropic flipped its consumer Claude policy requiring users to opt out or have their conversations used for training, with retention extending from 30 days to five years. You could just toggle this setting off and wait for the next time they try to end-run their own rules. Or you could create a replica they can't touch. Build it once. Use it forever.
› Run pipeline analysis on hundreds of deals without sending client names to Claude. Ask "which deals are at risk" and get patterns back with tokens — you map tokens to real records yourself.
› Surface ICP signals from closed-won customers without exposing who they are. The AI sees firmographic patterns without real names and emails.
› Make win-loss analysis, content performance correlation, and segmentation routine instead of risky.
› Set a team norm where "use AI on customer data" stops being a judgment call. The boundary lives in the system, not in everyone's head.
› Set a team norm where "use AI on customer data" stops being a judgment call. The boundary lives in the system, not in everyone's head.
› Set up the Airtable base. Four tables — Contacts, Companies, Deals, Tokens. Each entity table has fields mirroring HubSpot (lifecycle stage, deal stage, industry, dates, amounts) plus a Token field. The Tokens table holds the mapping between real values and the anonymized identifiers Claude will see. (Free Airtable tier handles the basic version).
› Build the contacts sync in Make. Webhook trigger → HubSpot Search CRM Objects → loop through results → check Tokens table for existing match by HubSpot ID → if none, create a new token → replace text fields with contact_<fieldname>_<token>, pass dropdowns and dates through unchanged → write to Airtable. The hardest part is the loop pattern if you haven't built one — budget extra time the first time. Repeat the pattern for Companies and Deals.
› To avoid the server cost and time of a constant sync, build an on-demand trigger to update your digital twin when you need it . A Slack slash command (/sync-ai) hitting the Make webhook works well — type the command, fresh data lands in Airtable in 30-60 seconds.
› When you are done, connect Claude to Airtable via MCP. Then test. Run manual analysis on real data and AI analysis on twin data to see it in action.
Vendor policies change — Anthropic flipped its consumer training defaults last September, retention rules shift, tools ship with permissions you didn't ask for — but when AI only ever sees anonymized data, those changes stop being your problem. The pipeline analysis you didn't run, the win-loss patterns you didn't surface, the segmentation you weren't sure was safe — all of it becomes routine because the safety question has already been answered by the architecture.
At last week's Knowledge 2026, CEO Bill McDermott opened with a real incident: an AI agent at a client gained elevated permissions and deleted a production database in 9 seconds. ServiceNow's answer is AI Control Tower — auto-discovery of every AI asset across the enterprise, plus a one-button kill switch. Learn More...
Cohesity announced an agentic rollback product in March, joining Rubrik (August 2025) and Cisco's native rollback capabilities. The tools integrate with ServiceNow and Datadog to monitor for anomalies, then trigger API-driven restorations of agents, agent memory, vector databases, and enterprise data. Gartner predicts 40% of enterprise apps will include agents in 2026. Learrn More...
A pending California bill would require employers to give workers 90 days' written notice before deploying any AI tool, and publish an inventory of every AI tool in use. Penalties run up to $500 per employee per violation. Learn More...