The governance gap is a data problem

Executives believe they have this under control. Their staff disagree.

Ask a room of executives what worries them about AI agents, and the answers are governance and oversight, then compliance. Ask their employees what they actually did last week, and you get a different picture entirely.

Both sets of numbers below are Okta's own, from two different studies. They are tagged separately throughout, because they are not interchangeable and one number in particular appears in both meaning different things.

The two 58% figures are not the same number. On the left, 58% is the share of leaders naming governance a top concern (Okta, AI at Work 2025). On the right, 58% is the share of executives reporting an AI security incident or close call in the last 12 months (Okta, AI Agents at Work 2026). Two studies, two questions, one coincidence.

Read together, they describe an organization whose leadership is confident about a state of affairs that does not exist.

The gap has a specific shape: confidential documents, pasted into unapproved tools

This is not abstract governance anxiety. Okta asked what employees actually put into AI tools.

Across knowledge workers who already use AI(Okta, AI Agents at Work 2026)

• 36% shared HR-related information
• 29% shared confidential company documents
• 16% shared login credentials and passwords

Among those using unapproved tools, the numbers climb(Okta, AI Agents at Work 2026)

• 54% shared internal messages and emails
• 45% shared HR-related information
• 39% shared confidential company documents, including financials and contracts
• 28% shared banking and payment information
• over 20% shared login credentials and passwords

For a consultancy, "confidential company documents, including financials and contracts" is not a category of risk. It is the deliverable. It is the thing under NDA.

Note who Okta surveyed: knowledge workers across more than a dozen industries, in data-oriented roles. This is not a healthcare finding or a finance finding. Wherever the work product is confidential — client engagements, regulated submissions, sealed commercial terms — the same behaviour is already happening.

Why it happens — and why stricter policy won't stop it

Okta asked employees why they bypass sanctioned tools. The answers are not defiance:(Okta, AI Agents at Work 2026)

• 80% — easier to use their own accounts
• 78% — their team already uses it; it's considered normal
• 57% — the approval process is too slow or difficult
• 49% — the approved tools do not meet their needs
• 6% — didn't know approval was required

Only 6% is an awareness problem. The rest is a product problem.

Okta's own recommendation is unambiguous on this point. Among the four steps they set out for closing the gap, the second is to make the secure path the easiest path: they argue that tightening policy will not fix shadow AI and will simply be disregarded wherever it slows people down, and that the fix is removing friction from the approved route until it is also the quickest one. Their reasoning is that employees follow the sanctioned path when it is also the efficient path.

That is the entire premise Untraceable was built on, arrived at independently by an identity vendor with nothing to sell you here.

The uncomfortable corollary for regulated work: a sanctioned AI tool that strips out the confidential content is, by construction, one of the tools that "does not meet their needs." Redaction retains roughly 37% of baseline AI writing quality; crypto-tokenization roughly 32% (derived from Tremblay & Harricharan, 2026 — see note below). You cannot close a shadow-AI gap with a tool people have a rational reason to abandon — what redaction costs you, measured method by method.

Where we fit — stated precisely

We are not an identity platform. We do not issue credentials, manage non-human identities, scope agent permissions, or discover shadow agents. Okta and its peers do that, and you should use them.

We are the data layer underneath. Identity controls govern who may ask, and what they may reach. Untraceable governs what actually leaves the building when they do. The confidential value is replaced with a meaning-preserving cloak before transmission — so even a fully authorized, correctly logged, entirely sanctioned AI interaction carries no confidential data inside it.

The two layers are complements, not substitutes. Okta answers "where are my agents, what can they connect to, what can they do." Untraceable answers "and what did they receive."

01 — Data loss

There is nothing to lose, because nothing was sent.

Asked about AI generally rather than about agents, Okta's leaders ranked data privacy first, at 68%, and security risks second, at 60% — the two highest AI-related concerns in that survey(Okta, AI at Work 2025). That is a separate question from the agent-concerns ranking above — no three-year horizon, and about AI as a whole — so the two are not additive and should not be read as one list.

Every other control here is probabilistic. DLP rules can be worded around. Enterprise agreements are contractual promises about retention and training — enforceable after disclosure, not preventive. Redaction protects the data by destroying the context the model needed.

Cloaking is architectural.

• Documents open locally. Cloaking runs locally. Only cloaked text is transmitted.
• Real values are restored on the operator's screen after the response returns.
• We never see the confidential value or the key.
• Intercept the traffic and you get informative stand-ins — never a name, a figure, an identifier.

And it doesn't cost you the tool: Advanced Cloaking preserves roughly 94–98% of baseline AI writing and reasoning quality, against ~37% for censoring-style redaction and ~32% for crypto-tokenization.

Most data-loss controls reduce the probability of exposure. Cloaking removes the thing that could be exposed.

02 — Compliance and regulatory requirements (50% of leaders(Okta, AI at Work 2025))

Use frontier AI without asking your regulator to look away.

For consultants and regulated experts, "compliant AI use" is not one obligation with one owner:

The first two rows apply to every consultant alive. The rest stack on top depending on the field.

An enterprise AI licence covers the vendor's obligations to you. It does not cover your obligations to your client. Your Copilot agreement is not a party to your NDA — why an enterprise licence answers the wrong contract.

Cloaking sidesteps the negotiation entirely. You are not arguing that a vendor's contractual assurance is strong enough to satisfy a sponsor, an ethics board or a privacy commissioner. You are demonstrating that the confidential value was never disclosed.

03 — AI governance and oversight (58% of leaders(Okta, AI at Work 2025))

Governance you can hand to a compliance officer.

Governance fails for a mundane reason: nobody can see what happened. Hence 90% executive confidence sitting on top of 52% unsanctioned use(Okta, AI Agents at Work 2026).

Untraceable makes the data question answerable.

• Audit trail on every account. Every value, every cloak, every AI submission — recorded. Cloak labels and run identifiers only, never raw values. Retained in your own SharePoint.
• Protocol Certification. Breach testing that certifies the cloak held, producing a permanent report you can send to a client or hand to a compliance officer. The Untraceable Protocol
• VaultAudit (Enterprise). A live dashboard across every project and team member: cloak-coverage metrics, operator accountability, organization-wide oversight.
• Sentinel auto-detection. Flags confidential values before submission rather than after disclosure.

And the second-order effect, which is the one Okta's data actually argues for: a sanctioned tool good enough that people prefer it is itself a governance control. 49% of shadow users said approved tools don't meet their needs(Okta, AI Agents at Work 2026). Fix that, and the visibility problem shrinks on its own.

If you are writing the policy that sits on top of this, the AI use policy template is a one-page starting point built around the question that actually matters: which data can go into which tool.

Scope note: this is governance of what data reaches AI. Not agent identity governance, permission scoping, or agent discovery. Those remain your identity platform's job, and we are designed to sit beside them.

What this looks like in practice

Two examples, two confidentiality regimes, one engine. The method is not field-specific: the same cloaking runs over a consulting engagement and a clinical study report alike.

Same engine, same guarantee, two different confidentiality regimes. One document. Three problems answered at once: nothing to lose, nothing disclosed, everything logged.

Untraceable is built for consultants and regulated experts generally. HEOR, payers and RWE, pharma, medical writing and biostatistics are the fields where the field-specific layer — Hybrid RAG and the Grounding Codex — is live today, with ESG and management consulting in progress. The cloaking itself is not field-bound.

A note on sources. This page cites published research by Okta, Inc. as independent third-party evidence, with attribution. Okta is not a partner, sponsor, endorser or affiliate of Untraceable, and has no commercial relationship with us. Their research is used illustratively because an identity vendor reaching this conclusion independently is more persuasive than our making the same claim unaided. Okta has not reviewed, approved or supported this page, and any interpretation of their data is ours alone. Okta and its research titles are trademarks of Okta, Inc., used here for identification only. Figures are reproduced from Okta's published summaries; see the sources above for the exact study behind each one. Each figure was re-verified against the primary source on 1 September 2026.

Ready to close the gap? Request a demo · Security architecture · The Untraceable Protocol