AI Governance · Data Protection · Enterprise AI
Your employees are not the threat. The unguarded gap between their keyboard and a public AI model is. Here is why anonymization has moved from a nice-to-have to a requisite, and how we built Navitra OS to close that gap.
| By Pratik Batra, Founder & CEO | November 2026 · 8 min read |
Picture an ordinary Tuesday. A finance analyst pastes last quarter's vendor payment ledger into a chatbot to "clean up the formatting." An HR partner drops three candidate CVs into an assistant to draft a comparison. A developer shares a stack trace that happens to include a database connection string. A bid manager uploads a draft tender response, client name and pricing intact, to tighten the executive summary.
None of these people meant any harm. Each of them got their job done faster. And in each case, corporate data left the building without anyone in IT, security or compliance knowing what went, where it went, or whether it can ever be pulled back.
That is the defining data-protection problem of the AI era. It is not a breach in the traditional sense. There is no attacker, no malware and no alert. It is a quiet, continuous, well-intentioned leak, and most organisations cannot see it.
The scale of the problem, in numbers
The research published through 2026 points in one direction. AI adoption by employees is running far ahead of the controls designed to govern it.
| 66% of office professionals have used AI at work while believing company policy did not allow it. PagerDuty, 2026 | 67% of AI users reach AI through personal, unsanctioned accounts on corporate devices, measured from network traffic. Verizon DBIR, 2026 | 33% of surveyed employees admitted entering sensitive information into unauthorised AI tools. Gartner, Feb 2026 |
Netskope's 2026 research puts the average organisation at roughly 223 generative-AI data policy violations every month. And it is not only free consumer tools. Gartner's 2026 labour market survey found that the large majority of employees who have an employer-provided AI tool are also using a personal one for work. Giving people an approved assistant does not, on its own, stop data flowing elsewhere.
Why "just block it" no longer works
The instinctive response is to block AI domains at the firewall. It rarely holds. Employees who have experienced a genuine productivity gain will find a way back, through a phone, a home laptop, a browser extension or a personal account. Blocking does not remove the behaviour. It removes your visibility of it.
Traditional data loss prevention struggles too. DLP was built for files and email attachments, not for a paragraph of free text typed into a chat window, a screenshot of a dashboard, or a code snippet with a secret buried on line 47. And policy documents, however well written, depend on every employee making the right judgement on every prompt, every day.
The realistic goal is not to stop people using AI. It is to make sure that what reaches the model is safe to send, regardless of which model it is or how the employee got there. That is precisely what an anonymization engine does.
| You cannot un-send a prompt. Once data reaches a model you do not control, the only protection that counts is the one applied before it left. |
What an anonymization engine actually does
An anonymization engine sits at the boundary between your people and the AI models they use. Every prompt, file and attachment passes through it before it is sent. Done properly, it works in five stages:
| 01 | Detect Identify sensitive entities in real time: names, emails, phone numbers, national IDs, account and card numbers, customer records, salaries, credentials, API keys, internal project codenames and client identifiers. |
| 02 | Mask or tokenize Replace each sensitive value with a consistent placeholder, so "Acme Holdings owes AED 2.4M" becomes "[CLIENT_A] owes [AMOUNT_1]". The model still understands the structure of the request, but never sees the real data. |
| 03 | Apply policy Decide, by role, department and data class, whether to mask, warn, block, or route the request to a private model instead of a public one. |
| 04 | Re-identify on return When the answer comes back, restore the original values for the authorised user only. The employee gets a fully useful response; the model provider never held the real data. |
| 05 | Log everything Record what was detected, what was masked and which policy fired, so security and compliance teams can finally answer the question they currently cannot: what data has gone to AI, and where? |
The critical design principle is that the employee experience barely changes. If anonymization makes AI slow, clumsy or less useful, people will route around it, and you are back to shadow AI. The engine has to protect the data without dulling the tool.
Anonymization is necessary, but it is not the whole answer
A standalone redaction tool solves one problem and creates another: yet another point product for an already stretched security team. In practice, the organisations asking "what data are we leaking to AI?" are asking several other questions at the same time:
| → Which AI tools and models are our people actually using, sanctioned or not? |
| → Who is uploading documents, code and data, and should they be? |
| → How much are we spending on tokens across Copilot, Claude, GPT and the rest, and is any of it wasted? |
| → Can we prove to an auditor or regulator what our AI usage looks like? |
These are not separate problems. They are four views of the same flow of data between people and models. Which is why we believe anonymization belongs inside a single governance layer, not bolted on beside it.
How Navitra OS reflects what we have learned
We did not design Navitra OS in a vacuum. It came out of a real enterprise mandate: a large organisation running Microsoft Copilot Studio, Claude and GPT models side by side, needing one window to monitor usage, control budgets, stop unauthorised uploads of data, code and documents, and do all of it without taking away the AI capabilities their people value.
The result is what we call the operating system for enterprise AI governance. Our tagline says it plainly: Every AI. Governed. Anonymization is built into that control plane as a first-class capability, alongside everything else a CIO and CISO need.
| Protect Anonymization at the prompt boundary Sensitive entities are detected and masked before any prompt or file leaves your perimeter, with policy-driven block, warn or redact actions on uploads of data, code and documents. | Govern One control plane, every model A single window across multiple LLMs and providers, with role-based access control and the ability to block unauthorised AI usage, rather than a separate console per vendor. |
| Monitor Visibility you can audit Usage, tracking and security monitoring by user, team and model, so you can answer the regulator's question with evidence instead of assumptions. | Optimise Budgets and token economics Budgets by department, prompt optimisation, context caching and smarter model selection, so governance pays for itself instead of becoming another cost line. |
For organisations where no data may leave the premises at all, such as government entities, regulated financial institutions and critical infrastructure operators, Navitra OS also serves as the governance, monitoring and budgeting layer of our private, on-premises enterprise AI offering. In that model, anonymization becomes a second line of defence: data stays in-house, and even inside the perimeter, each team sees only what its role permits.
That is the understanding Navitra OS is built on. AI usage is not going to slow down, and it should not. The organisations that win will be the ones that let their people use AI freely because they have the controls to make it safe.
| Five questions for your next leadership meeting 1. Can we list every AI tool our employees used last month, including personal accounts? 2. If a client asked whether their data had ever been sent to a public model, could we answer with evidence? 3. Is sensitive data masked automatically, or are we relying on each employee's judgement? 4. Do we know what we spend on AI per department, per model, per month? 5. Are our AI controls making people safer, or just pushing them towards shadow AI? |
If more than one of those answers is "not really," you are in the same position as most enterprises today. The good news is that it is a solvable problem, and it does not require choosing between productivity and protection.
| See what your data is saying to AI Explore Navitra OS, or talk to our team about an AI governance and anonymization assessment for your organisation. |
Related reading: more insights from the Profecia Links blog · our enterprise technology services
Top comments (0)