Here is a scene every AI team knows by heart. The modelling sprint starts on Monday, the ticket for table access goes in the same morning, and the approval lands sometime next sprint — usually after somebody senior gets cc’d and annoyed. Meanwhile the one person who did get access last quarter still has it, doing nothing with it. That gap between how fast you build and how slowly permissions move is exactly what DataSecOps was invented to close, so let us diagnose why the lag happens and what a sensible fix looks like.

The Real Problem Is Not the Model, It Is the Waiting Room

Teams assume their access process is slow because security people are cautious. Usually the opposite is true: the process is slow because it is manual, undocumented and personal. Approvals live in inboxes and hallway conversations. Nobody can say which policy justified a grant, because there was no policy — there was a favour. And when an analyst changes teams, the old grants follow them around like an old phone number nobody bothered to delete. The cost compounds quietly: sensitive columns sit untagged, so every new notebook that touches them inherits an invisible liability, and the audit question “who approved this?” gets answered with a shrug. The same stall shows up in the wider platform story: as our guide to the xOps family explains, every “ops” label is really an agreement about who does what automatically. If access decisions stay outside that agreement, the whole pipeline idles at one specific checkpoint, and no amount of GPU spend fixes it.

A Vendor Word With a Surprisingly Precise Meaning

Credit where due: the term did not drift in from a conference banner. Ben Herzberg, chief scientist at Satori Cyber, defined it publicly on 4 August 2021 as an agile method of keeping security embedded while data and its users keep changing. Two honesty notes belong next to any use of the word. It is a coinage from a vendor, not from a standards body — the long-established name for the same practice is data security governance. And the vendor itself was absorbed: Commvault bought Satori Cyber on 28 August 2025, with the preliminary figure put at about USD 28.3 million. Gartner never adopted the label either; its 2025 market guide files the tooling under data security platforms. None of this makes the practice less real. It just means you should describe it precisely rather than brandish the buzzword, and a good plain reference to keep on hand is this atlas page defining DataSecOps.

Drawing the Line Around DataSecOps

The scope is narrower than the name suggests, and that is a virtue. Inside the fence: finding data wherever it hides, tagging it by sensitivity, deciding policy per row, column or query at the moment of access, then applying masking, tokenisation, format-preserving encryption or pseudonymisation, managing the keys, watching what leaves through exports and copies, and auditing who read what until someone formally reviews it. The unit of control is refreshingly small — a row, a column, a single query. Outside the fence sit neighbouring disciplines people often blur into it: DataOps owns pipeline flow, PrivacyOps owns lawful basis and data-subject requests, IAMOps owns identities, and the telemetry side of security analysis goes by SecDataOps. That particular split is editorial rather than official, but it keeps meetings short.

The Fix: Treat Access Like a Code Review

Once the scope is clear, the mechanics borrow the best habit software teams already have. An access request arrives as a structured thing — dataset, purpose, duration — instead of a Slack plea. The data owner reviews it the way a maintainer reviews a pull request, and approval triggers the pipeline: tests run, the grant applies itself, and an expiry date rides along with it. Routine requests clear the same day, which quietly removes the incentive to route around the process. Anything unusual needs a ticket, a recorded risk sign-off and a hard deadline. Then a quarterly review walks the whole grant list and withdraws whatever nobody has used or justified. Waiting room problem solved, without hiring an army of approvers.

The full cycle, in the order teams usually install it:

1. Sweep every warehouse, lake and shadow spreadsheet so nothing sensitive stays invisible.
2. Tag what the sweep found, with people confirming the machine’s guesses.
3. Write the rules as code — per row, per column, per query — and version them.
4. Enforce at the moment of access, not in a quarterly audit trail.
5. Watch the exits: exports, copies and shares are where data actually leaves.
6. Read the access history, then sit down quarterly and cancel what has gone quiet.

Classifiers Need a Human Second Opinion

One warning before anyone promises full automation. Discovery scanners, including Google’s and Presidio’s, are good and useful and never completely accurate — they will miss odd formats and invent labels for innocent ones. Mature setups therefore keep people in the loop to confirm the IDENTIFIER and SENSITIVE tags before policies key off them. That is not a failure of the tooling; it is the design. A mislabelled column either hides data your analysts need or exposes data the regulator cares about, and both mistakes cost more than a five-minute human check. Teams that already read how AI sharpens business decisions will recognise the pattern: model proposes, human disposes.

What Breaks When Nobody Closes the Door

Still tempted to defer all of this? Look at 2024. Attackers tracked as UNC5537 went after customers of a major cloud warehouse using stolen credentials, and the post-mortem arithmetic was ugly. Close to four in five of the hit accounts — 79.7% — showed earlier infostealer contact on their logins, multi-factor authentication was absent, and some passwords had sat unchanged for up to four years. Notification letters eventually went out from roughly 165 organisations. The warehouse platform itself was never breached — the data left through ordinary COPY INTO commands run with legitimate logins. In other words, the front door worked exactly as designed; the keys were the problem. A request workflow with expiring grants and a quarterly revoke pass shrinks that attack surface to a fraction, and pairing it with the automation choices covered in our comparison of automated and manual data pipelines keeps the safeguards from becoming the new bottleneck.

DataSecOps in One Paragraph

Diagnose honestly: your delay is a manual, favours-based approval habit, not cautious engineers. Then install the loop — find the data, tag it, encode policy per row and column, enforce at query time, watch the exits, read the access history, review and revoke on schedule — and let humans confirm what scanners merely guess. The label is young, the vendor behind it now belongs to Commvault, and the standards bodies call the practice something else. The relief your AI team feels next Monday, when access arrives the same day it was requested, will not care what the thing is called.