Staff are already using AI tools, whether or not the company has approved them. Verizon's 2025 Data Breach Investigations Report found that 15% of employees routinely used generative AI on corporate devices, and 72% of them signed in with non-corporate email addresses. A short AI usage policy turns that hidden activity into something you can manage. This guide lists the clauses to include and the public frameworks behind them.
Why a small company needs an AI policy now
The risk is not the technology itself but unmanaged use. In Verizon's 2025 data, 15% of employees accessed generative AI systems on corporate devices at least once every 15 days. Of those, 72% used personal email addresses as their account and a further 17% used corporate emails without integrated sign-in. That means company data is going into accounts the company does not control.
Small firms are also behind on governance. The UK Cyber Security Breaches Survey 2025/2026 found that 31% of businesses use, are adopting or are considering AI, but only 24% of those have practices in place to manage AI risks. A two-page policy closes much of that gap.
The frameworks to borrow from
You do not need to invent categories. NIST's AI Risk Management Framework 1.0, released in January 2023, organises the work into four functions: Govern, Map, Measure and Manage. Its Generative AI Profile, NIST AI 600-1, published on 26 July 2024, lists 12 risks that generative AI creates or makes worse, including confabulation, data privacy, information security and intellectual property.
In India, MeitY released the India AI Governance Guidelines on 5 November 2025 under the IndiaAI Mission. They set out seven guiding principles and recommendations across six pillars, and the government said the approach relies on existing laws where possible rather than a new AI law. For a small company, that means the DPDP Act and IT rules already apply to AI use.
Clause 1: approved tools and accounts
List the AI tools staff may use and require company-managed accounts with single sign-on for each one. Ban personal accounts for work content. This one clause addresses the biggest finding in the Verizon data, where most generative AI use ran through personal email identities.
Include a simple request route for new tools, with a named approver and a short checklist: where the vendor stores data, whether inputs are used for training, what retention and deletion options exist, and whether the tool supports your sign-in provider. NIST's value chain risk covers exactly this: third-party models and data that are poorly vetted reduce your accountability.
Clause 2: what data may go into AI tools
Define three tiers. Public information can be used freely. Internal information, such as draft documents and code, can go only into approved company accounts. Restricted information, including customer personal data, health records, payment details, credentials and contract terms, must not be pasted into general-purpose AI tools at all unless a specific, reviewed system is built for it.
This clause carries legal weight in India. The DPDP Rules, notified on 14 November 2025, begin an 18-month phased rollout, and the Act allows penalties of up to ₹250 crore for failing to keep reasonable security safeguards over personal data. NIST AI 600-1 also warns that models can leak or infer sensitive personal data.
Clause 3: human review of AI output
NIST lists confabulation, where a model produces confident but false content, as a core generative AI risk. Your policy should say that a named person is responsible for any AI-assisted output that leaves the company, whether that is a customer email, a proposal, code or a social media post.
Set stricter rules for high-stakes work. Legal, medical, financial and HR content needs review by someone qualified, and figures or citations must be checked against the original source. NIST also flags human-AI configuration risks such as over-reliance and automation bias, so remind staff that review means reading, not skimming.
Clause 4: security for AI features and agents
If you build AI into products or workflows, the policy needs a security section. Prompt injection, listed first in the OWASP Top 10 for LLM Applications 2025, lets crafted input make a system ignore its instructions or leak data. Sensitive information disclosure comes second. NIST AI 600-1 also names prompt injection and data poisoning under information security.
Practical rules: AI agents get the least access they need, never broad admin keys; any action that sends money, deletes data or contacts customers requires human approval; system prompts and connected data are treated as confidential; and AI features are tested for injection on authorised systems before launch.
Clause 5: intellectual property, disclosure and incidents
NIST notes that generative AI may reproduce copyrighted material or trade secrets. Require staff to check that AI-assisted content does not copy third-party work, and to disclose AI assistance to clients where contracts or professional rules require it. Do not paste a client's confidential material into a tool the contract does not allow.
Finally, treat AI misuse as a security incident. If restricted data went into an unapproved tool, staff should report it the same day. For Indian entities, CERT-In's April 2022 directions require cyber security incidents to be reported within six hours, so the internal route must be faster than that.
A one-page policy outline
Keep the policy short enough that people read it, and review it every six months, since tools change quickly. The table below maps each clause to the NIST risk it addresses. XDQ Labs helps small teams draft AI policies and test AI features on systems they are authorised to assess.
| Policy clause | What it says | NIST AI 600-1 risk addressed |
|---|---|---|
| Approved tools | Company accounts with single sign-on only | Value chain and component integration |
| Data tiers | No restricted or personal data in general tools | Data privacy |
| Human review | Named owner checks all external output | Confabulation; human-AI configuration |
| Security | Least privilege, approvals, injection testing | Information security |
| IP and disclosure | Check for copying; disclose where required | Intellectual property |
| Incidents | Same-day internal reporting | Information security; data privacy |
Frequently asked questions
What should a small company AI usage policy include?
At minimum: a list of approved AI tools with company accounts, rules on which data may be entered, a requirement for human review of external output, security rules for AI features and agents, intellectual property and disclosure rules, and a same-day route for reporting misuse.
Can employees use personal AI accounts for work?
A sound policy says no. Verizon's 2025 report found 72% of employees using generative AI on corporate devices signed in with non-corporate emails, which puts company data outside company control. Require company-managed accounts with single sign-on instead.
Does India have a law on AI use by companies?
There is no dedicated AI law. MeitY's India AI Governance Guidelines, released in November 2025, favour using existing laws where possible. That means the DPDP Act, whose penalties reach ₹250 crore for weak security safeguards, and CERT-In rules already apply to AI use.
What are the main risks of generative AI for businesses?
NIST AI 600-1 lists 12 risks. For most small companies the most relevant are confabulation (false but confident output), data privacy, information security issues such as prompt injection, intellectual property, and third-party value chain risks from tools and models you did not vet.
Sources
- 2025 Data Breach Investigations Report: Executive Summary, Verizon, 2025
- NIST AI 600-1: Generative AI Profile (full text), NIST, 2024
- Cyber security breaches survey 2025/2026, GOV.UK (Department for Science, Innovation and Technology), 2026
- AI RMF: Generative Artificial Intelligence Profile (NIST AI 600-1), NIST, 2024
- MeitY unveils India AI Governance Guidelines (press release), Press Information Bureau, Government of India, 2025
- Digital Personal Data Protection Rules, 2025: backgrounder, Press Information Bureau, Government of India, 2025
- What are the OWASP Top 10 risks for LLMs?, Trend Micro, 2025
- The CERT-In Cyber Security Directions: More Questions Than Answers, Cyril Amarchand Mangaldas, 2022
