Three Years of Enterprise AI Adoption: Why AI Security, Governance and Value Realization Have to Rise Together
We have spent three years working with enterprises on AI adoption. The tools changed almost every month. The pattern did not.
Companies adopt AI fast, risk climbs just as fast, and then they clamp down, which protects the business but quietly switches off the value. The organizations getting ahead have stopped choosing between the two. This is what we have learned, and what we would tell any leader starting out today.
Defense and Offense: Two Halves of the Same Job
Defense is protecting the business: AI security, governance and control over what data goes where. Offense is making AI pay off: adoption, productivity and value realization. Most companies treat them as a trade-off and lean hard one way. In our experience, that is where AI programs stall. Defense without offense gets you a locked-down tool nobody uses. Offense without defense gets you fast adoption until the first leak, or the first budget surprise. They have to rise together.
Three Years In, AI Is a Paradox
Three things are true at once.
It is easy to use. The ChatGPT app reached 1 billion monthly active users in May 2026, faster than any other app, according to Sensor Tower data. Nobody had to ask IT first.
It can be a productivity superpower. Morgan Stanley’s AI tool for deciphering old code had saved its developers more than 280,000 hours by June 2025, per the bank’s head of technology strategy (the article is in Japanese, translated from Business Insider).
It carries a lot of risk. In 2023, Samsung engineers pasted sensitive code and a recorded meeting into ChatGPT on at least three occasions, and the company responded by limiting prompt size. More recently, reports said Uber burned through its entire 2026 AI coding tools budget in four months, with its COO questioning the return on that spend. And LayerX found that 77% of enterprise AI users paste data into chatbot queries, with 82% of those pastes coming from unmanaged personal accounts.
Easy, powerful and risky, all at the same time. That is why enterprise AI adoption is so hard to get right.
The Six AI Adoption Challenges We See in Almost Every Enterprise
When we look across our clients, the challenges fall into two groups. Three are about defense: protecting the business. Three are about offense: getting the value.
Defense: Protect the Business
- Visibility and shadow AI. Who is using AI, which tools, for what, and at what risk? Most companies cannot say, because AI arrives through browser tabs, personal accounts and apps they already own. McKinsey found that executives estimate 4% of employees use gen AI for at least 30% of their daily work, while employees report 13%, roughly three times higher. Deloitte Australia learned what that gap can cost: it agreed to repay the final installment of an AU$440,000 government contract after fabricated citations were found in its report, with AI use disclosed only in the corrected version.
- Security and insider risk controls. Once you can see AI, can you control it? Every prompt can carry sensitive data, and agents now act with real access to your systems. IBM’s 2025 research found that organizations with high levels of shadow AI saw an average of $670,000 in higher breach costs. Agents add a different kind of exposure. SaaStr’s founder reported that Replit’s AI agent deleted a production database during a code freeze, then told him a rollback was impossible, which he says turned out to be wrong.
- Compliance. Can you keep the records regulators expect? Investigations and audits need every AI transaction, often for years, and that record holds your most sensitive material: executive prompts, production code, IP. Rules are arriving. Under the EU AI Act, deployers of high-risk AI systems must keep automatically generated logs for at least six months, and providers must keep technical documentation for ten years. Courts can force the issue too. In the New York Times lawsuit, a judge ordered OpenAI to preserve output log data that would otherwise have been deleted. The order was later lifted, but the question stands: do you have your own record?
Offense: Get the Value
- Cost control. Can you stop rogue agents and runaway token spend? AI is billed by usage, so a few heavy users or one agent running nonstop can blow a budget. Spend limits from each vendor stop at that vendor’s edge, and no single cap spans users, agents and vendors. The controls also tend to arrive late. OpenAI did not launch ChatGPT Enterprise until August 2023, by which point it said teams had adopted ChatGPT in over 80% of Fortune 500 companies, measured by accounts registered with corporate email domains.
- Finding use cases worth investing in. Which uses of AI actually justify the investment? Without usage and results data, pilots cannot prove they deserve production. IDC research for Lenovo found that for every 33 AI proofs of concept a company launched, only four graduated to production. Klarna’s story shows the other side of the problem. It said its AI assistant was handling two-thirds of customer service conversations, doing the work of 700 agents. Its CEO later said cost had been too dominant a factor, leading to lower quality, and the company moved to invest in human support again.
- Insight for leaders. Can your leaders make AI decisions on evidence? Security, IT, legal, finance and the business all need a clear view of usage, risk and results, but the data is scattered across tools and teams, so decisions run on opinion. And accountability does not wait for clarity. When Air Canada’s chatbot gave a customer wrong information about bereavement fares, the airline argued it could not be held liable for what its chatbot said. The tribunal disagreed, finding that Air Canada did not take reasonable care to ensure its chatbot was accurate. Who owns what your AI says, and who can see it?
What We Watched Happen: Three Phases of AI Adoption
Across clients, we have seen the same sequence play out.
Phase 1: the shadow surge. Because AI is so easy to adopt, usage races ahead. People get real productivity gains, and risk rises right alongside, because almost all of it is shadow.
Phase 2: controls. Companies respond with defense. They block and lock down. Risk falls, but so does the value, because the same restrictions that cut shadow usage also cut the use cases that were working.
Phase 3: harmonize. Leaders who get here can see what is being used, protect it, and scale what works. Defense and offense rise together.
That third phase is the new paradigm. You do not have to trade control for value. You need a single record that serves both.
One Record, Both Sides: What an AI Adoption Platform Does
This is what we have been building at Portal26. An AI adoption platform captures every prompt, attachment and response, every agentic interaction, the identity of who issued it, and the licensing behind the tool that was used. Every interaction is retained in a NIST FIPS 140-2/3 vault and classified by user, department, app, use case, risk and token.
From that one record come both sides of the job:
- Defense: visibility and shadow AI discovery, security and insider risk controls, and compliance and audit controls.
- Offense: licensing and cost controls, use case analysis, and ROI reporting.
Because both sides draw on the same evidence, they stop competing. The controls that protect the business are the same data that shows where AI is working.
How to Start: Give Your AI a Checkup
The best way to begin is the way you would begin with your own health: get the checkup first, then decide what to do. In practice that is three steps.
- Assess. Turn the lights on. In two weeks, you get one readout covering shadow AI and risk, security and licensing, token usage and efficiency, top users and adoption, and the use cases ripe for agents.
- Control. Play defense. Act on visibility and shadow AI, security and insider risk controls, and compliance.
- Optimize. Play offense. Control cost, invest in the use cases worth scaling, and give leaders the insight to steer, every quarter.
The data only helps if people can use it. That is where Recon by Portal26 comes in. Recon is the natural-language front end to your AI record: ask a question in plain English and the answer is assembled from every AI transaction your organization has generated. There is no dashboard to learn and no analyst to book, and you can re-ask anything for any period. It lets leaders manage intelligence directly, and it feeds the State of AI report, which shows where you stand on both sides: shadow AI and sensitive data in prompts on defense, and adoption, dormant users, token concentration and use cases on offense.
What Changes When Leaders See the Full Picture
When everyone looks at the same record, each function gets what it needs:
- Security sees risk and has the record to investigate.
- IT sees every tool and agent in use.
- Legal and compliance get audit-ready records.
- Finance sees spend by user and agent before the budget is gone.
- Business leaders see where AI is actually working.
- HR sees how employees are using AI, how proficient they are, where adoption is landing and where training is needed
Decisions run on evidence instead of opinion. And since AI moves at machine speed, governance has to as well.
Where Enterprise AI Adoption Goes From Here
Three years in, the lesson is simple. Defense without offense gives you a locked-down organization that cannot show a return. Offense without defense gives you a breach waiting to happen. The organizations that win are the ones that can see everything, protect what matters and scale what works, from one source of truth.
If you want to know where you stand today, start with the checkup.