Is It Safe to Use Company Data with AI? A Practical Security Guide
Before you worry about which AI tool to trust, you need to know which data you're even allowed to put in front of it - here's the classification system and the five contract clauses that actually matter.
You're about to paste a client contract into ChatGPT to summarize it, and for half a second you hesitate. That hesitation is correct - but the question most people ask next is the wrong one. They ask 'is this AI tool safe,' when the question that actually matters is 'what category of data is this, and am I allowed to put it anywhere outside my own systems at all.'
Is it safe to use company data with AI? The honest answer is: it depends entirely on what the data is, not which tool you picked. Teams that get burned almost never get burned by a sophisticated breach. They get burned because nobody drew a line between 'this is fine to paste anywhere' and 'this should never leave our building.'
Classify first, panic about tools second
Before you evaluate a single AI vendor, sort your data into three tiers. This takes an afternoon with your team, not a consultant engagement.
- Public - marketing copy, published blog drafts, job postings, anything already visible on your website. No restriction needed.
- Internal - meeting notes, internal SOPs, draft financials, non-identifiable business data. Fine for most reputable paid AI tools, not for free consumer tiers.
- Sensitive - customer PII, health records, payment data, anything under NDA, anything regulated (PDPA, GDPR, HIPAA-equivalent). This needs a specific, verified data processing agreement before it touches any third-party AI - or it shouldn't touch one at all.
If your team can't tell you which tier a piece of data falls into in under ten seconds, the tool you're using doesn't matter - you've already lost the plot.
At itsherbs, the TCM clinic chain this agency spun out of, we hold records for 30,000+ patients across 10 locations - health data, which sits at the most sensitive end of any classification system in Malaysia. Patient notes never go into a consumer AI tool, full stop, regardless of how convenient it would be. That line was drawn before we ever evaluated a single vendor's security page.
What to actually read in a vendor's terms of service
Marketing pages love the word 'secure.' It tells you almost nothing. When a vendor's sales team says their platform is safe for business use, make them point you to the actual clause for each of these five items:
- Training use - does your input get used to train their models by default, or only with explicit opt-in?
- Retention period - how long is your data stored after the session ends, and is that configurable?
- Data residency - which country are servers physically located in, and does that matter for your compliance obligations?
- Subprocessor list - which third-party infrastructure (cloud hosting, analytics, logging tools) touches your data downstream?
- Deletion rights - can you request full deletion on demand, and how long does the vendor commit to completing it?
If a vendor can't answer all five in writing within a day, that's your answer about whether they take enterprise data seriously, regardless of what their homepage claims.
Free tools vs. paid tiers - what you're actually buying
Here's the part most teams get wrong: they assume paying for an AI subscription buys them better output. Often what it actually buys is a different data contract. Free consumer tiers of most major AI tools default to using your conversations to improve their models - that's the trade you're making for a free product, and it's a reasonable one for drafting a birthday card, not for a client proposal.
Business tiers - ChatGPT Team or Enterprise, Claude for Business, and equivalents - generally carry a contractual exclusion from training use, along with SOC 2 attestations and admin-level controls over retention. That contractual exclusion, not some extra layer of magic encryption, is the real product you're buying when you upgrade. If your team is still running sensitive work through free personal accounts, the upgrade pays for itself the first time it prevents one bad outcome.
When on-premise deployment is actually worth the cost
Every few months someone asks if they should just run their own local LLM to avoid the whole question. Usually the answer is no. On-premise or self-hosted deployment only makes financial sense when three conditions hold at the same time: you're processing a genuinely large volume of sensitive data daily, a regulator or client contract explicitly mandates it, and you have the engineering capacity to maintain, patch, and monitor that infrastructure indefinitely - not just set it up once.
Miss any one of those three and you've built an expensive, slower, harder-to-update version of a tool you could have rented safely for a fraction of the cost. We've seen teams spend months and a meaningful chunk of budget standing up local infrastructure for a use case that a properly configured enterprise AI contract would have covered for a monthly fee.
The companies that get this right don't have better lawyers. They just did the unglamorous work first: sort the data, read the five clauses, match the tool to the tier. Everything downstream from that - which vendor, which plan, whether to self-host - is a much smaller decision than it feels like before you've done the sorting.
Frequently asked questions
Is it safe to use company data with AI tools like ChatGPT?
It depends on the tier of data and the plan you're on. Public and internal business data is generally fine on paid business tiers (ChatGPT Team/Enterprise, Claude for Business), since those exclude your input from model training by contract. Sensitive data - customer records, health data, anything under NDA or regulation - should only go through a tool with a verified data processing agreement, or not at all. The tool matters less than whether you classified the data before deciding.
How do I know if an AI vendor is training on my data?
Check the data processing agreement or enterprise terms for an explicit training opt-out clause, don't rely on the marketing page. Free consumer tiers of most major AI platforms train on inputs by default unless you manually opt out in settings. Paid business and enterprise plans typically state in the contract that customer inputs are excluded from training. If a vendor can't point you to the specific clause within a day of asking, treat that as your answer.
Should I pay for an AI enterprise plan just for data privacy?
Yes, if your team regularly handles internal or sensitive business data - the price difference is usually small compared to the risk of a free-tier data leak. The real value of upgrading to a business tier isn't better output quality, it's the contractual exclusion from training use plus admin controls over retention and deletion. If your team is currently pasting client or financial data into personal free accounts, the upgrade cost pays for itself the first time it prevents a bad outcome.
How do I choose a vendor or consultant to help secure company data for AI use?
Look for someone who can show you their own data classification system in practice, not just a compliance checklist they sell. Ask them to walk you through how they'd sort your specific data into tiers, which five clauses they'd check in a vendor's terms, and under what conditions they'd ever recommend on-premise deployment over a paid SaaS tool. A good advisor will tell you when you don't need a complex solution - if they recommend expensive infrastructure for every client regardless of data volume, that's a signal to look elsewhere.