Is Customer Data Safe with AI Tools? A Guide for Project Management

Why Data Safety Is the Hidden Barrier to AI Adoption in Project Management

If your project management team is excited about AI but hesitating to actually turn it on, you're not being paranoid — you're being rational. The dilemma is real: AI promises measurable gains, with PMI research showing AI-enhanced organizations reporting 61% of projects delivered on time versus 47% without. Yet adoption remains stubbornly low.

The numbers tell the story. Only 22% of project managers say AI tools are deployed and in use, while 39% have plans to deploy — and survey data shows 26% cite ethical, privacy and security concerns as a barrier to adoption. The same source names the fear explicitly: cybersecurity concerns and "the risks of putting your data into a model that you don't completely own," with the added observation that addressing these concerns adequately "seems very expensive at the moment."

That fear is legitimate, and it's compounded by a regulatory gap. The shift toward AI tools is happening so quickly that regulation is lagging behind — a particular problem in regulated industries like healthcare and financial services, where customer data carries legal obligations under frameworks like GDPR and CCPA.

What makes this barrier "hidden" is that it rarely shows up in vendor demos. It surfaces in the fine print. Consider what leading platforms actually disclose:

• ClickUp states its AI is not trained on Workspace data and holds zero data retention agreements with OpenAI, Anthropic and Google — but still flags AI Agents as a genuine leak risk, noting users might see information in agent responses they could not otherwise access.
• OpenAI, Anthropic, Google and Microsoft all state business data is not used for training by default on paid business plans — but protections "depend on the agreement you sign and the control set you enable."
• Organizations deploying Microsoft Copilot are investing in data classification, access governance and tenant cleanup before rollout, because embedded AI will surface whatever users already have access to.

In other words, the deciding factor is not the model itself but how AI connects to your internal data and how easily that access can be governed. Contract detail matters more than marketing language — and "by default" is not the same as "guaranteed."

The risks are growing, not shrinking. Gartner expects more than 40% of agentic AI projects to be canceled before the end of 2027, partly due to inadequate risk controls. And shadow AI — employees using consumer-grade tools on personal accounts before enterprise agreements exist — creates data exposure no license can retroactively fix.

The practical takeaway: treat data safety as a gating requirement, not an afterthought. Verify no-training and no-retention terms in writing, clean up permissions before rollout, and maintain human oversight of AI outputs. The same scrutiny applies to the agent skills and connectors you attach to your AI stack — which is why SkillzBench is building free security scanning and trust scores for agent skills, so teams can see which components are safe before they connect anything to their project data. Data safety is not paranoia; it's the entry ticket to AI's productivity gains.

What Vendors Actually Promise — and Where the Fine Print Bites

The good news is that major AI vendors have quietly strengthened their contractual privacy posture. The bad news is that the strongest protections live in signed agreements, not marketing pages — and one of the biggest leak risks comes from the AI agents everyone is rushing to deploy.

Start with the positives. ClickUp states its AI is not trained on Workspace data and holds zero data retention agreements with OpenAI, Anthropic and Google, meaning prompts and outputs aren't stored by the underlying model providers. Similarly, enterprise plan comparisons confirm that OpenAI, Anthropic, Google and Microsoft all commit to not training models on business data by default on paid business tiers.

Those defaults sound reassuring, but they come with caveats that matter in practice. As the same analysis puts it, enterprise privacy protections "depend on the agreement you sign and the control set you enable" — for regulated industries, contract detail matters more than marketing language. In other words, the sales page is not the contract.

Then there are the geographic and access-control wrinkles:

• ClickUp notes data is processed in the US even for EU-hosted customers, which may complicate compliance for European teams.
• Embedded AI like Microsoft Copilot "will surface whatever users already have access to," which is why organizations invest in data classification and tenant cleanup before rollout.
• ClickUp's permission-aware AI only protects data if your underlying permissions are actually correct.

The sharpest warning concerns AI agents. ClickUp itself flags that a user with access to an item but not an agent's full configured knowledge "might see information in the response that they could not access otherwise" — and connecting a personal app to an agent exposes that app's data to anyone who can interact with it. That's a genuine leak vector, not a hypothetical one. It helps explain why Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027, partly due to inadequate risk controls.

The takeaway for project managers: treat vendor promises as a starting point, then verify. Get no-training and no-retention terms into the signed agreement, confirm where data is processed, and pilot agents with restricted knowledge sources before granting them broad access. If you're evaluating agent skills and connectors alongside your PM platform, the free directory and MCP tool being built at SkillzBench is designed to help you see exactly which skills an agent runs — and which ones carry risk.

Governance Beats the Tool: The Deciding Factor in Keeping Data Safe

Here's the uncomfortable truth: the AI tool you choose matters far less than the governance you build around it. STACK Cybersecurity puts it bluntly — the deciding factor is not the model itself but how AI connects to internal data and how easily that access can be governed.

That's because embedded AI doesn't create most data problems; it reveals them. Organizations deploying Microsoft Copilot are investing in data classification, access governance, and tenant cleanup before rollout, precisely because Copilot will surface whatever users already have access to. If your permissions are messy, AI makes the mess visible — to everyone.

This is why governance must precede tool selection. PMI's experts recommend a dedicated governance team — data scientists, compliance officers, and legal experts — supported by a specific control set:

• Role-based access control (RBAC) so people see only what their role requires
• Multi-factor authentication (MFA) across every system AI touches
• Encryption of data at rest and in transit
• Retention and deletion policies aligned to GDPR and CCPA — holding data too long creates liability
• Continuous compliance monitoring, because even an algorithm can inadvertently expose unauthorized data if not properly monitored

PMI's framing is stark: get governance right and you unlock AI's potential while staying secure; get it wrong and you risk regulatory fines, operational failures, and breaches. The stakes are real — Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027, partly due to inadequate risk controls.

Agents deserve special caution. ClickUp warns that a user with access to an item but not all of an agent's configured knowledge "might see information in the response that they could not access otherwise" — and connecting a personal app to an agent exposes that app's data to anyone who can interact with the agent. Pilot agents with restricted knowledge sources, and audit who can talk to them.

Then there's shadow AI: employees using consumer-grade tools on personal accounts before any enterprise agreement exists. That creates data exposure no enterprise license can retroactively fix — which is why clear AI policy needs to land before, not after, rollout.

The same governance discipline applies to the agent skills and connectors your teams attach to AI tools. SkillzBench is building a free directory that verifies, security-scans, and benchmarks agent skills so you can see exactly what each one does before it ever touches your data. Governance first, tools second — and verify everything in between.

Your Implementation Playbook: Five Steps to Safe AI in Project Management

Getting AI into your project management workflow safely isn't a single decision — it's a sequence. Do the steps out of order, and you inherit risks that are expensive to unwind later.

Step one: get the data-handling terms in writing. OpenAI, Anthropic, Google and Microsoft all state that business data isn't used for model training by default on paid plans, but as cybersecurity researchers put it, contract detail matters more than marketing language. ClickUp, for example, documents zero data retention agreements with its LLM partners — but that's a vendor's own FAQ, not an independent audit. Put no-training and no-retention commitments in the signed agreement, not just the sales page.

Step two: clean up data classification and access permissions before rollout. This is the step most teams skip, and it's where embedded AI bites. As organizations deploying Microsoft Copilot have learned, these tools will surface whatever users already have access to — so if your permissions are loose, AI makes the problem visible at scale. That's why many organizations now invest in data classification, access governance and tenant cleanup before enabling AI features, and some disable bundled tools like Gemini at the admin level until data policies are defined.

Step three: treat AI agents as a higher-risk category. ClickUp explicitly warns that a user interacting with an agent might see information in the response they could not otherwise access — and that connecting personal app knowledge to an agent exposes that app's data to anyone who can interact with it. Pilot agents with restricted knowledge sources, review who can talk to them, and remember that Gartner expects more than 40% of agentic AI projects to be canceled by end of 2027, partly due to inadequate risk controls. If you're evaluating agent skills for your stack, SkillzBench is building exactly this kind of scrutiny — security scanning, version pinning and sandbox benchmarking for skills before they ever touch your data.

Step four: stand up governance with a named owner. PMI recommends a dedicated governance team — data scientists, compliance officers, legal experts — backed by RBAC, MFA, encryption, audit logs, retention policies aligned to GDPR and CCPA, and continuous compliance monitoring. Their warning is blunt: without oversight, even the best governance frameworks are just theoretical exercises.

Step five: run the data-readiness pre-flight check. Before tool selection, verify five signals: a single source of project truth, completion data captured within 48 hours, twelve months of plan-versus-actual history, a named owner for AI output, and a pre-agreed kill criterion. As the guidance goes, data cleanup precedes tool selection, always — and it matters, since 38% of failed AI use cases trace back to poor data quality or limited data availability.

One more thing to budget for: people. 68% of project managers say they haven't received enough AI training, so pair every rollout step with enablement. Get the sequence right, and customer data can stay safe while the AI does real work.

Going Further Safely: Vetting the Skills and Agents You Connect

Your PM tool's AI might be locked down, but the moment you bolt on third-party skills and agents, you've opened a new door to customer data. Published research found that roughly 26% of agent skills carry at least one vulnerability — a number that should give any project manager pause before connecting external capabilities to systems holding client information.

The risk is well documented. ClickUp explicitly warns that agent responses can expose information users could not otherwise access, and that connecting app knowledge to an agent exposes that data to anyone who can interact with it (according to the vendor's own security FAQ). Gartner, meanwhile, expects more than 40% of agentic AI projects to be canceled by the end of 2027, partly due to inadequate risk controls.

The lesson is simple: skills and agents deserve the same scrutiny you applied when vetting your PM vendor. That means checking what data each skill can reach, how it's maintained, and whether anyone has actually validated it works as claimed.

When evaluating skills before they touch customer data, look for:

• Security scanning — automated vulnerability checks on every skill, not just a README and good intentions
• Version pinning — each skill version tied to a specific commit SHA, so you know exactly what you're running and can detect upstream changes
• Independent benchmarking — skills tested in a sandbox with the same model, with and without the skill, so you see real performance lift rather than marketing claims
• Trust signals — a composite score built from security, quality, license and maintenance data, kept separate from performance metrics so ratings can't mask risks

This is exactly the gap SkillzBench is being built to close. Its free, independent directory will continuously crawl and dedupe skills from GitHub, then security-scan, license-check and benchmark every one in a sandbox — publishing the measured Lift alongside a 0–100 trust score. A free MCP connector will let you attach vetted skills to any agent, with per-agent loadouts so each agent only sees the tools you explicitly choose. And because upstream skills change, SkillzBench will track those changes, re-scan and alert you when something shifts.

Everything — the directory, the MCP connector, benchmarks and alerts — will be free, with advertising on the blog keeping scores and listings untouched by commercial influence. As cybersecurity guidance puts it, the deciding factor is not the model itself but how easily its access can be governed. Treat connected skills the same way, and your customer data stays yours. SkillzBench is pre-launch — join the waitlist at skillzbench.com to get early access when the directory goes live.