Building AI Guardrails Without a Rulebook: What IT Leaders Are Doing
- Carrie Pallardy
- 915 views
No rulebook exists for AI guardrails, so IT leaders are building their own around governance, model choice, data control and human oversight.

There's no settled playbook for AI guardrails yet, and that's the uncomfortable part. Boards are asking about them more often as AI agents start doing work beyond what they were originally set up to do. Yet the CIO who sits down to draft them has very little to copy. The frameworks that do exist give a starting point, not a finished set of rules for a specific company.
Starting from frameworks that stay general
Most IT leaders begin with established references. The AI Risk Management Framework from the National Institute of Standards and Technology and ISO 42001 both give teams a structure to work from. The trouble is that they're written to apply across many uses, so they don't say much about any single application.
Ashley Casovan, managing director of the AI Governance Center at the International Association of Privacy Professionals, sees the pattern in the questions teams bring to these frameworks. Organisations ask whether they have a governance board and whether someone is clearly responsible or accountable. Those are sensible questions, but they don't tell you what a claims-processing bot or a sales assistant actually needs. "How you actually navigate to what types of standards or guardrails you need at what time for your particular context or use of AI is becoming increasingly difficult," she said.
Her read is that the noise in the market makes this harder. Vendors, consultants and analysts all have views, and enterprises chasing AI at scale often end up with a pile of advice that doesn't connect to their own risks. Mitigating both enterprise risk and human risk is still a work in progress for most organisations.
Who sits on the governance committee
Here's the thing: the guardrails that matter depend on the company's own context, which is why the committee itself becomes the first real control. Casovan's advice on membership is specific. "You need people with a policy background, with a legal background, with an ethics background, with a technical background that understands how those systems work," she said.
That mix matters because the committee has to answer hard questions. How should safe and responsible use be defined for each business case? How do you test outputs for accuracy and bias? Which existing internal policies already cover AI, and where do new ones need writing? A committee made only of technologists will tend to miss the policy questions. One made only of lawyers will struggle to test whether a model actually behaves as promised.
Guardrails built into the product and the testing
Paula Goldman, chief ethical and humane use officer at Salesforce, works on the trustworthiness question from the vendor side and in conversations with other businesses. She describes guardrails in three layers. Some are policies about how customers may use the product. Some are built directly into the product. Some come from testing, and that testing includes deliberately trying to break things before a flaw shows up in production.
That last point is worth dwelling on. Red-team style testing isn't new in security, but applying the same mindset to model behaviour is still uneven across enterprises. If your organisation is buying AI features rather than building them, you can still ask your vendors how they test, what failure cases they look for and what they do when they find one.
Choosing models without overbuying
The range of AI models available makes selection harder than it should be. The priciest options aren't always the right fit, and many teams pick a model based on a demo rather than on their own workload.
Goldman says Salesforce uses a benchmark aimed at CRM use cases to compare models. The idea is to skip generic tests. "Rather than testing models generically in whatever the consumer context might be, we're testing them for business goals and business use cases," she said.
Prashant Hinge, chief information and transformation officer at insurer MSIG USA, keeps the decision simpler. "It is very clear you don't need a frontier model to solve all your problems in any given workflow," he said. He suggests choosing models that deliver business value with the right security, the right provenance and a sensible cost.
Planning for the model that fails
Picking a model for today's needs is one problem. Planning for the day that model disappears or changes is another. AI vendor lock-in is an enterprise risk that calls for its own guardrails, whether the trigger is a vendor's financial collapse, a sale of the company or a product change nobody saw coming.
Hinge's answer is flexibility. "Always plan for swapping your models and your agents overnight," he said. In practice, that means keeping prompts, tools and integrations portable enough that you can move workloads without rebuilding them. It's easy to say and expensive to retrofit, so it's worth raising early.
Data control and vendor scrutiny
Data privacy comes up constantly in Goldman's conversations with leaders. Their core demand is simple. "They want to know that they are in control of their data, that their data is their data, and that they don't have these third parties training on their data," she said.
Hinge argues that a contract alone doesn't settle the question. He recommends asking vendors for concrete examples of how their policies are applied, then digging into third- and fourth-party relationships to see who else touches the data. Reading the terms carefully matters, but so does testing whether the practice matches the paperwork.
There's also the question of how much data to share at all. Even with sound privacy and retention terms, Hinge doesn't think handing a large frontier model every piece of enterprise data makes sense. "You're just teaching the frontier model your organizational ways of working," he said. "You want to own that layer. You want to own the brain of your organization." That view points toward smaller language models for certain tasks, rather than sending everything to externally run frontier systems.
Where the human sits as agents get more autonomous
A human in the loop is still the guardrail most enterprises cite when asked about AI outcomes. The hard part is working out where that person belongs, and how their role shifts as agents act with less supervision.
Goldman is candid about the tension. "The whole point of AI agents is that you don't have a human looking at every single data point," she said. "But you definitely need a record of every action and data point. You need anomaly detection." In other words, you give up line-by-line review in exchange for audit trails and alerts that flag when a person needs to step in.
Hinge is more concerned about people trusting model output too readily. If staff aren't questioning what the system tells them, he wants a technical layer of protection behind the scenes. That's a sensible design principle: assume people will accept AI output by default, and build checks that don't depend on them remembering to push back.
Accountability is still unsettled
Enterprises want to show their AI is safe, fair and accurate, and that takes accountability behind every guardrail. The problem is that accountability hasn't been standardised. Casovan points out that outside a company, there's little that would hold a specific individual personally answerable for AI outcomes. For now, that question falls to internal policy, or the lack of it, and to whatever the market does with a product. "If there's a lack of consumer uptake in a product, then the head of the AI product development or the CEO would ultimately be held accountable, but that would end up becoming a board decision," she said.
Hinge sees the same gap from inside a company. "Personal accountability matters because now you can make bad decisions faster and at scale," he said. As AI governance lags behind adoption, risk grows, and so does the pressure to name someone responsible when something goes wrong. Goldman's view is that one leader shouldn't carry all of it. Enterprises need broader stakeholder involvement, and IT leaders are still very much involved in the guardrails built into the general-purpose tools their businesses use.
Domain leaders are also being asked to set guardrails for their own areas, which puts oversight closer to the work. Casovan argues they still need executive leadership behind them. "That is not something that you're just going to let engineering teams manage and take care of. They shouldn't have that burden," she said.
What to do this quarter
If you're starting from scratch, my read is that the order matters. Stand up a committee with policy, legal, ethics and technical members. Pick two or three live use cases and write guardrails for those, rather than a universal policy that nobody can apply. Test models against your own workloads. Get vendor data terms in writing and check how they're applied in practice. Build logging and anomaly detection before you give agents more reach. And decide now who is accountable when a model makes a bad call, because waiting for an incident to answer that question is the expensive way to learn it.