Data as a Policy Input
Updated
IntermediateWhat you'll learn
~15 min- Build a label taxonomy small enough to use and wire auto-labeling to it
- Connect data sensitivity to access decisions - the pillar's actual Zero Trust move
- Sequence DLP from audit to block without triggering the business revolt
The pillar that got a forcing function
Data has always been the least mature Zero Trust pillar — everyone agrees it matters, nobody’s labels survive contact with real users. Then AI assistants arrived and did something accidental and clarifying: they made oversharing instantly visible. The copilot that cheerfully summarizes a salary file for anyone who asks didn’t create the permission problem — that file was always readable by half the org — it surfaced it, at chat speed. Every 2026 data-pillar program is riding that forcing function, and yours should too: “make the AI rollout safe” gets budget that “fix file permissions” never did.
Labels: the taxonomy that survives is small
The data pillar’s unit of policy is the sensitivity label — stamped on the document, traveling with it, machine-readable by every downstream control. The design rule learned by every failed rollout: four or five labels, not fifteen. A taxonomy people can hold in their head gets applied; one that needs a decision tree gets ignored, and unlabeled data is policy-invisible data.
Public - the website, the outage mapInternal - default for org content; the do-no-harm tierConfidential - customer PII, financials, most of the crown jewelsRestricted - regulator case files, credentials, safety-critical OT docsHumans mislabel and forget, so auto-labeling carries the volume: pattern-driven service-side rules stamp the obvious classes (account numbers, SSNs, credential material) and default-label inheritance covers repositories at rest — pre-AI hygiene for the mountains of existing content the assistants are about to index. Humans adjust edge cases; machines do the bulk. Meridian’s lake makes it easier than most estates: the Cloud Modernization data-contract discipline means datasets already have owners and dictionaries — the label is one more column in a governed catalog, not a greenfield crusade.
The Zero Trust move: labels drive decisions
A label that only colors a footer is decoration. The pillar earns its place in this training when sensitivity becomes a signal into the PDP, closing the loop with everything you’ve built. The label-to-signal decision structure looks like:
Label × Identity/device signal => DecisionConfidential × unmanaged device (CA08) => browser-only, no downloadRestricted × any session => phishing-resistant step-up (2.2)Restricted-app × session age => re-evaluation ceiling: hourly sign-in frequency (4.3 uses this)Restricted × elevated insider-risk level => block external share + alertany × AI assistant context => assistant honors label ACLs; provenance-untrusted content (4.1) never joins the promptThat third row is adaptive protection — the newest pattern in the set: an insider-risk engine scores behavior, and the score swaps the user between DLP/access policy tiers dynamically. A user exhibiting exfil-shaped behavior gets the strict policy set automatically; everyone else keeps frictionless defaults. That’s tenet 4 (dynamic policy) reaching the data pillar — the same closed-loop shape as Lesson 3.1, with risk flowing from behavior instead of endpoint telemetry.
DLP without the revolt
Data-loss-prevention controls enforce the labels at the exits (mail, upload, USB, print). The rollout discipline is the module’s recurring one, because DLP block-mode on day one is the classic business revolt, legitimate workflows break loudly, executives escalate, and the program retreats to audit mode permanently, which is worse than sequencing honestly. The four-step rollout structure looks like:
1. AUDIT - watch flows per label/exit pair; find the legitimate-but- noncompliant workflows (there are always some)2. FIX PATHS - give those workflows sanctioned routes BEFORE blocking the unsanctioned ones (secure share links replace the vendor email attachments; Lesson 4.1's washroom handles inbound)3. NOTIFY - warn-and-continue with a justification prompt; the data tells you what block-mode will cost4. BLOCK - per label-tier, crown jewels first, with the exception register (owner + expiry) for the stragglersSwitch your workspace agent into plan mode on your data-contract dictionary and CA baseline, then start with:
Plan our four-label data sensitivity rollout wired to our data-contractdictionary, CA baseline policies, staged DLP exit channels, and AI assistantindexing rules. Ask clarifying questions about our repositories and exitchannels first.Check that the plan or draft covers:
- Four-label taxonomy (
Public,Internal,Confidential,Restricted) with one-line definitions, three concrete examples each, and service-side auto-labeling rules plus repository default inheritance - A
label x signaldecision matrix wired directly to our CA baseline policy names (CA08browser-only on unmanaged devices, phishing-resistant step-up, hourly session frequency, and adaptive insider-risk blocks) - The
audit -> fix paths -> notify -> blockDLP sequence per exit channel (mail, upload, USB, print) with the promotion metric gating each step - AI assistant indexing rules specifying which labels assistants may index and how provenance-untrusted content from Lesson 4.1 is excluded from context windows
If the plan jumps straight to DLP blocking or leaves AI indexing unscoped, follow up with:
Add the sanctioned-path replacement and promotion metric before each DLPblock stage, and exclude Restricted and provenance-untrusted files from AI context.Meridian's copilot pilot is a hit — until someone asks it to 'summarize everything about substation access codes' and it obliges, quoting a Restricted operations doc stored in a broadly-shared team site. The knee-jerk proposals: ban the copilot, or delete the document. What does the data pillar actually prescribe?
Key takeaway
Four labels people can remember, auto-labeling carrying the volume, and the labels wired into the PDP so sensitivity changes decisions — download rules, step-ups, adaptive tiers, and what the AI assistants may index. Sequence DLP audit→fix→notify→block, and thank the copilot for the budget: it turned invisible oversharing into a demo. One question left in this module: what keeps a session honest after the decision.