· How to run AI with confidence,
· How to keep litigation exposure in check,
· How to hold spend to a plan, and
· How to know what is actually running in the environment.
Leaders are expected to answer them.
Four insights stand out from these conversations, and the common theme is having visibility in one place. Because you can’t manage what you can’t see. Here is what we’re seeing heading into year-end:
1. AI erased the line between data and privacy.
For years, the organization chart kept data and privacy apart. In general, IT owned the data. Legal and compliance owned privacy. The nature of AI requires both to be part of the same conversation, to stay compliant and to protect the organization from litigation.
A single AI misstep can produce two separate claims. The first begins as a data problem. A model that drifts, or one trained on flawed data, produces skewed decisions. Skewed decisions become bias and discrimination claims, and civil rights law prohibits discrimination whether or not AI made the call.
The second begins as a consent problem. Data collected for one purpose becomes a privacy violation once a model uses it for another. The consent litigation reshaping web tracking, including claims under the California Invasion of Privacy Act (CIPA), state wiretap laws, and pixel class actions, can also reach AI tools, depending on how they collect or transmit communications.
The stakes are already visible in legal departments. In FTI Consulting’s 2024 General Counsel Report, 80% of legal department leaders rated data privacy one of their top five risks, well ahead of ESG. The 2026 Carlton Fields Class Action Survey found that the share of general counsel naming data privacy and cybersecurity as key drivers of future class actions doubled in a year, to 25%.
The organizations handling this well have given data and privacy owners a shared table. In our April 2025 CIOReview feature, we recommended a cross-functional AI governance board with legal, compliance, cybersecurity, HR, and executive leadership. Including department heads on that board is essential, in whatever form fits your organization. Governance depends on understanding how people actually use AI tools, and IT cannot carry that responsibility in a vacuum. It takes collaboration across the business.
2. Governance at this scale takes a platform.
In 2025 we recommended a centralized AI inventory and compliance scanning tools, and that remains the right foundation. Most organizations begin with manual discovery: surveying teams, reviewing SaaS spend and single sign-on logs, and recording what turns up. From there, they turn to the endpoint solutions capturing AI data, which may now include the majority of your cyber, SaaS, and network vendors. Centralizing what those tools capture is essential, and keeping that picture current across hundreds of tools, models, and agents calls for a platform.
Although an enterprise may try to standardize on one AI provider, in reality a typical environment includes models from OpenAI, Anthropic, Microsoft, and Google, home-grown agents, AI features inside licensed SaaS, and tools nobody registered. Each vendor’s dashboard covers its own slice, which leaves the enterprise the task of trying to understand the full picture.
Vendor oversight needs the same central view. Organizations have long told vendors they cannot use company data beyond the contracted service. AI governance requires those terms to be specific: whether a vendor may use your data to train or improve its models, how it handles prompts and outputs, and whose AI touches your information downstream. A governance platform should audit every vendor’s AI policies against those terms and alert you when a vendor’s practices put your data at risk of unauthorized use. The vendor’s contract terms set the rules for how their AI may use your data at the moment of signing, but constant monitoring is needed to track legal changes and the vendor’s use of AI.
Recent agentic AI incidents show why timing matters. OpenAI and Anthropic each disclosed models acting outside their intended sandbox during evaluations believed to be simulated, reaching real infrastructure and real organizations. Both came to light in after-the-fact review. Live monitoring of agents gives leaders the earliest signal, and scheduled, risk-based reviews build on that foundation.
A mature governance platform brings inventory, vendor oversight, token spend monitoring and restriction, and privacy into one place. It provides:
• Continuous discovery of models, agents, and AI-enabled applications
• Auditing of vendor AI policies and data-use terms
• Monitoring from the organization’s own vantage point
• Enforcement of internal policies on data access and autonomous actions
• Detection of new risk as permissions and integrations change
• Real-time token spend monitoring, with the ability to throttle or stop runaway usage
• An audit trail the organization controls, ready for regulators, auditors, insurers, and boards
• The ability to alert, restrict, suspend, or escalate when an agent crosses a line
With one inventory, one policy set, and one audit trail across the AI estate, leaders gain the confidence to approve new AI initiatives, because they can see and steer what is already running.
3. AI spend can move like a bullet train.
AI governance also has a cost side, and it is catching well-run companies off guard. OpenAI reports that average reasoning-token consumption per organization grew about 320x in twelve months as AI spread across products and workflows.
Tokens behave differently from software licensing. Pilot costs are predictable. Production adds longer context, retrieval, retries, tool calls, and agent loops, and each adds tokens. At enterprise scale, small inefficiencies become material. As one lesson from our AI GRC workshop puts it, a team can burn a year of tokens in a month when nobody is watching.
The discipline to manage this already exists in the cloud FinOps playbook, and early adopters are proving it works on tokens:
• Caching. Care Access cut data processing cost 86% and processing time 66% with prompt caching on Amazon Bedrock.
• Model tiering. BPC moved its production chatbot to a lightweight model and reduced chatbot costs by about two-thirds. Frontier models handle hard reasoning, mid-tier models handle everyday work, and lightweight models handle high-volume repetition.
• Prompt discipline. AWS documented a production multi-agent system that cut token consumption about 70% and latency about 50% by removing prompt bloat and duplicated instructions.
• Real-time visibility and throttling, so a usage spike is caught before it consumes the budget.
The strongest programs pair cost with value. Chime tied AI call summarization to more than 250,000 hours saved annually, $700K in efficiency gains, and a five-point increase in Net Promoter Score. Measuring spend against outcomes like these gives leadership teams a clear basis for investment decisions.
Cost belongs in the governance conversation because AI is a usage model. Leaders who can see what their AI is doing, whose data it uses, and what it costs to run are positioned to scale it as far as the business can use it.
4. The most expensive AI in your company is the AI you can’t see.
Every insight above rests on visibility, and two blind spots account for most of the surprises.
The first is token spend no one is watching. It tends to surface on an invoice, after a workflow has run at ten times its usual volume. The organizations caught off guard are usually well run. They lacked a live view of who was using what.
The second is shadow AI, meaning any AI operating outside your review. Part of it arrives embedded in software you already license, switched on inside vendor features and running as a black box. Accountability for automated decisions stays with the organization that relies on them. In EEOC v. iTutorGroup, hiring software that screened out older applicants led to a settlement, and in Mobley v. Workday, a court allowed AI screening claims to proceed against the vendor as an agent of the employers using its tools. Liability follows the decision through the whole chain.
The rest arrives through employees, who adopt new tools faster than any approval process can review them. Each unsanctioned tool moves data outside the perimeter and makes decisions outside policy, unseen by security, legal, and finance.
I was recently playing tennis with a friend who works for a Fortune 500 global oil and gas company, and he asked, “Hey Meg, you work in AI governance, right? Is it okay that I’m putting my work information into my personal Claude at home? It’s private, right?” After I cringed, he got an earful from me.
His question was sincere, and he is far from alone. Most employees don’t know where the line is, so governance has to reach how people actually use AI, through clear policy and training, as well as the tools themselves.
What to Do for 2027
Moving forward, continuous discovery and visibility are how leaders will go from reacting to AI to steering it. When you work across the full IT landscape, privacy, AI governance, and cost control can become part of a single strategy.
This level of operation is a big ask. Organizations of any size will need a comprehensive platform because success demands a live, complete inventory of every model, agent, and AI-enabled application – approved or not, first-party or embedded.
There is good news. Among the organizations we’re helping achieve success, many had more of the foundation in place than they realized. Also, we’ve identified a platform we can recommend. And since we first brought this company into these client projects, Gartner has named Truyo a Leader in both the 2026 Magic Quadrant™ for AI Governance Platforms and the quadrant for Privacy UX.
Organizations that build comprehensive visibility now will expand AI on their own terms. And they’ll have the evidence to show regulators, boards, and customers that it is well run.
It’s time to get the most out of your AI investment.
Meg Toups is Founder and CEO of BlueSky IT Partners (www.blueskyitpartners.com). The company launched its AI Governance Accelerator Program in April 2025.
