Many companies begin governing AI after something goes wrong, such as a confidential file pasted into a public chatbot or a customer-facing answer that nobody can explain. By that point, dozens of tools and pilots are usually already in use, and each one needs a review after the fact. Setting a light framework early is faster and cheaper than retrofitting one.
Effective AI governance scales with risk: a low-risk internal drafting tool needs a short checklist, while an AI system that affects customers, money or health needs documented testing, human oversight and a named owner who is accountable for its results.
This article sets out risk tiers, six guardrails and a 90-day rollout plan that a mid-sized company can follow, and shows how they relate to current regulation and standards. It is not legal advice, and it does not cover governance for companies that train their own foundation models.
Key takeaways
- Start with an inventory of AI tools and use cases, since a company cannot govern systems it has not listed.
- Sort every use case into a risk tier and match the controls to the tier, so low-risk tools are not slowed down by high-risk paperwork.
- Give each AI system a named business owner who is accountable for its results.
- Guardrails are technical as well as written, and they include access rules, data filters, logging and human approval steps.
- Under the EU AI Act as amended in 2026, obligations for stand-alone high-risk systems apply from 2 December 2027.
- A working framework fits on a few pages and is reviewed every quarter.
What an AI governance framework covers
A framework answers four practical questions. What may the company use AI for? Who decides? How does the company know a system works? What happens when it fails?
Two terms are often used interchangeably. Guardrails are the controls inside a system, such as input filters, output checks, permission limits and approval steps. Governance is the decision layer above them that sets which guardrails each system needs and who is responsible for keeping them in place.
A framework works best when it sits next to existing security, privacy and quality processes and reuses them. One added step in the existing vendor review or release checklist gets followed more reliably than a separate AI approval process.
Start with risk tiers
Risk tiers keep governance proportionate. Most companies need three.
| Tier | Typical examples | Controls required |
|---|---|---|
| 1. Low | Drafting internal emails, summarizing your own notes, code suggestions reviewed by developers | Approved tool list, data handling rules, basic staff training |
| 2. Medium | Internal knowledge assistant, document extraction with human review, drafting sales emails | Tier 1 controls plus a named owner, a test set, access control, logging and a quarterly review |
| 3. High | AI that informs decisions on credit, hiring, insurance claims or medical triage, and customer-facing agents that take actions | Tier 2 controls plus a documented risk assessment, human approval of decisions, accuracy and bias testing, an incident plan and legal review |
Some tier 3 uses carry legal obligations of their own. The EU AI Act classifies AI used in recruitment and in creditworthiness assessment, among other areas, as high-risk, with requirements for risk management, data quality, documentation and human oversight.
Planning to roll out AI across several teams? We help set up the technical guardrails, from access control and logging to evaluation and human review, as part of the build.
Explore AI servicesSix guardrails to set before you scale
Each guardrail below has an owner and a simple signal that shows whether it is working.
1. An AI inventory
List every AI tool, feature and pilot with its owner, purpose, data, vendor and risk tier. Include AI features inside software the company already pays for, since these are the easiest to miss.
Owner: IT or the CTO's office.
Signal it works: a new AI tool cannot be purchased or connected without being added to the inventory.
2. An acceptable use policy
Keep it to one or two pages that staff will actually read. The policy names the approved tools, lists the types of data that must never be entered into them and explains when a person must check AI output before it is used.
Owner: legal and HR together.
Signal it works: staff can answer the question "may I paste this into the assistant?" without asking anyone.
3. Data rules enforced in systems
Decide which classes of data may reach which tools. Confidential client data might be allowed in an enterprise assistant covered by a no-training agreement and blocked everywhere else. Enforce these rules through access controls and data loss prevention, so they do not depend on people remembering them.
Owner: security and data protection.
Signal it works: blocked data types are stopped automatically, and each attempt is logged.
4. Testing before launch
Every tier 2 and tier 3 system needs a test set of real cases with expected results, plus tests for known weaknesses. The OWASP Top 10 for LLM applications is a practical checklist for security testing and covers risks such as prompt injection, sensitive information disclosure and excessive agency.
Owner: the delivery team, with sign-off from the business owner.
Signal it works: no release goes out without a recorded test result.
5. Human oversight matched to the tier
Decide where a person must approve each output, where a person reviews a sample and where the system runs on its own. Approval steps work best when the reviewer sees the source material and the reason behind the output. A bare approve button invites rubber-stamping.
Owner: the business owner of the system.
Signal it works: approval and override rates are tracked and discussed at each review.
6. Monitoring and incident response
Log inputs, outputs, model versions and costs. Set alerts for drops in quality, unusual usage and cost spikes, and agree in advance who has the authority to switch a system off.
Owner: engineering, with the business owner kept informed.
Signal it works: the team has run at least one incident drill and knows how long it takes to disable a system.
Who owns what
Governance fails most often when ownership is vague. A small structure is enough for most mid-sized companies.
- An AI steering group of five or six people meets monthly at first. It usually includes a business leader, the CTO or head of IT, security, legal or privacy and one person from a team that uses AI heavily. It approves tier 3 use cases and keeps the policy current.
- A business owner for each system is accountable for its results, its budget and the decision to keep it running.
- The delivery team builds, tests and monitors the system and reports issues to the owner.
- Every employee follows the acceptable use policy and reports problems through a known channel.
The steering group should not approve every prompt or small tool. Tier 1 decisions belong to the teams themselves, within the policy.
What regulation and standards expect
Three references cover most of what B2B companies in the US and Europe are asked about.
The EU AI Act
The AI Act entered into force in August 2024. Bans on prohibited practices have applied since February 2025, and obligations for providers of general-purpose AI models since August 2025. The Digital Omnibus amendments adopted in 2026 moved obligations for stand-alone high-risk systems to 2 December 2027 and for high-risk AI built into regulated products to 2 August 2028, as Orrick's summary of the final text explains. Transparency obligations under Article 50 apply from 2 August 2026.
The NIST AI Risk Management Framework
The NIST AI RMF is a voluntary US framework organized around four functions: Govern, Map, Measure and Manage. Many companies use it as the structure for their internal policy, because it is free, widely recognized and independent of any vendor.
ISO/IEC 42001
ISO/IEC 42001 is a certifiable management system standard for AI. It is useful when enterprise clients or regulators ask for formal assurance, and it fits well alongside an existing ISO/IEC 27001 security program.
A 90-day rollout plan
Days 1–30: see what is already in use
- Build the AI inventory and assign a provisional tier to each entry.
- Publish an interim acceptable use policy and an approved tool list.
- Form the steering group and agree how often it meets.
Days 31–60: set the rules
- Confirm tiers for all use cases and name a business owner for each tier 2 and tier 3 system.
- Enforce data rules in systems through access controls and data loss prevention.
- Start test sets for every live or planned tier 2 and tier 3 system.
Days 61–90: make it operational
- Add testing and monitoring to every live tier 2 and tier 3 system.
- Run one incident drill and the first quarterly review.
- Train staff on the policy, using real examples from your own inventory.
Governance work fits naturally into a wider adoption plan. Our guide on how to implement AI in your business places these steps alongside use case selection and rollout.
What guardrails look like in delivered systems
Guardrails are easier to understand in real projects. Three examples from .wrk's work show different controls for different risks.
- Data residency. A regional telecom operator with more than 12 million subscribers had to keep subscriber voice data inside its own perimeter. The on-premise voice AI platform runs in the operator's data centre, with monitoring, failover and a staged rollout across two pilot sites.
- Human review before publication. For a cybersecurity media company, AI drafts the daily newsletter, and an editorial review step stays in place before anything is sent to subscribers.
- Restricted outputs. In a CMS-based SEO tool, the AI was limited to an approved list of URLs and tested against previously failing cases, which stopped it from suggesting links to pages that do not exist.
Frequently asked questions
What is an AI governance framework?
An AI governance framework is the set of policies, roles and technical controls that decides which AI uses a company allows, who approves them and how their risks are monitored. It usually includes risk tiers, an acceptable use policy, data rules, testing requirements, human oversight and incident response.
How do you implement AI governance in a company?
Start with an inventory of AI tools and use cases, sort them into risk tiers and name an owner for each important system. Then publish an acceptable use policy, enforce data rules in your systems, and add testing and monitoring to medium and high-risk systems. A mid-sized company can put the basics in place within 90 days.
What are AI guardrails?
AI guardrails are the controls built into an AI system that limit what it can receive, produce and do. Examples include input filters, output checks, permission limits, approval steps and logging.
Does a small company need AI governance?
Yes, in a lighter form. A small company needs an approved tool list, clear rules on what data may be entered into AI tools and a named person who decides on new uses. More formal controls become necessary when AI affects customers or decisions about people.
Which frameworks can we base AI governance on?
The NIST AI Risk Management Framework is a common voluntary starting point, and ISO/IEC 42001 is used when formal certification is needed. Companies that operate in the EU also need to map their use cases against the EU AI Act.
To sum up
An AI governance framework works when it is proportionate and owned. List your AI uses, tier them by risk, set six guardrails and review them every quarter, so the company can scale AI without slowing down low-risk work.
.wrk builds AI systems for US and European companies with the technical guardrails included, from access control and logging to evaluation and human review.
Scaling AI beyond a few pilots? Tell us what is already in use, and we will help you build the guardrails into the systems themselves.
Explore AI services