Most mid-market companies don't need a dedicated AI compliance or security team to put real AI risk governance in place in 2026. What works instead is a named owner already accountable for the decision, running a short, repeatable review cycle on borrowed time from functions that already exist: IT, legal, whoever handles vendor contracts. That's a governance and accountability question, not a security architecture one, and it's a different discipline than designing or hardening the technical controls themselves. The risk doesn't shrink to match the team size. The way you manage it has to.

Adapting the rigor to the team you actually have

Steering committees or governance structures I've been part of ran on a subset of the staff a textbook version of that function would call for. The rigor didn't come from headcount. It came from being deliberate about what got reviewed every cycle versus what got flagged and set aside, and from treating the review itself as a fixed, protected block of time rather than something that happened whenever someone found a spare hour.

That's the same trade a company managing AI risk without a security function is actually making. The choice isn't between a full program and no program. It's between a lightweight cycle that runs consistently and a comprehensive one that exists on paper and rarely happens, because there was never enough dedicated capacity to run it the way it was designed.

Naming an owner isn't the same as managing the risk

Naming an accountable owner is the first decision in setting up AI risk management, not the last one. Once someone owns AI risk, the question that actually determines whether anything gets managed is what that person does with the time they have. A named owner with no repeatable process still ends up reacting to problems one at a time as they surface, which looks a lot like having no owner at all from the outside.

What 2026 data shows about the gap between naming an owner and actually managing the risk

Netrio's 2026 survey of 401 U.S. IT leaders at companies with 200 to 5,000 employees found only 26% describe their AI use as scaled and governed enterprise-wide, even though the same companies are investing heavily, 88% plan to spend at least $100,000 on AI this year.

Just 42% have formal AI policies with controls that are actually enforced, and only 53% report full visibility into which AI tools are in use at all. In the twelve months before the survey, 42% of these companies had a confirmed AI-related security incident or exposure, and another 31% reported a near miss.

Retool's separate 2026 survey of 307 senior technology and security leaders found the accountability gap sitting underneath those numbers: 44% either have no clear default for who's responsible when an AI tool causes a problem, or haven't decided yet. Only 8% describe their organization's AI governance as strong.

What this means in practice: if your organization has named someone responsible for AI risk but hasn't given them a specific, recurring block of time to actually review it, you're in the majority, and that majority is where the 42% who had an incident or near miss came from. Naming the owner turns out to be the easy part. The review cycle is what's usually missing.

Five things that can substitute for a dedicated AI risk team

Does borrowing a few hours from IT count as a real program?

It's more than most companies currently have, which is closer to nothing. It isn't the same as a fully staffed security function, and it won't catch everything a dedicated team would; deep technical audits and round-the-clock monitoring aren't things a few borrowed hours a month can replace. But measured against the alternative most mid-market companies are actually running today, an unreviewed patchwork of tools and an incident plan that exists only in someone's head, a lightweight cycle that runs on schedule closes most of the real gap.

The goal isn't matching an enterprise security budget. It's not being the company with 42% of Netrio's respondents in the "had an incident and didn't have a plan" column.

A few questions that separate a working cycle from a stalled one

If the honest answer to any of these is no, the risk isn't being managed yet, whatever the org chart says about who's supposed to be doing it.

How this connects to the rest of Security & Risk

Security & Risk is one of the five pillars the RAISE OS™ AI Maturity Assessment measures on its own, and a mid-market company can score reasonably here without ever building a dedicated function, as long as the lightweight cycle above is actually running. What sinks the score isn't the absence of a security team. It's the absence of anything repeatable standing in for one.

Find out where your organization actually stands

You don't need a dedicated AI compliance or security team to manage AI risk well in 2026. Here's the lightweight cycle that closes most of the gap instead.

How do we manage AI risk without a dedicated compliance or security team?

An AI-influenced hiring decision has already been challenged in federal court, and the case didn't get dismissed.

In Mobley v. Workday, a judge let age discrimination claims move forward past the company's motion to dismiss. The judge also authorized formal notice to a class covering everyone 40 or older who applied through the platform since September 2020. And the court allowed a theory that would have sounded like a stretch a few years ago: a hiring vendor's AI system can itself be treated as the employer's agent for discrimination purposes.

Court filings in the case reference roughly 1.1 billion job applications processed through the platform during that window. Whatever happens next in that specific case, one thing is already clear: a legal challenge to an AI-influenced hiring decision is no longer a hypothetical question.

Employers already sense this shift. Littler's 2026 Annual Employer Survey, drawing on more than 300 U.S. C-suite executives, in-house counsel, and HR leaders, found 79% expect AI-related litigation is possible in the next 12 months, and discrimination or bias was the single most-cited legal concern behind data privacy.

A quick note before we go further: Tier8 isn't a law firm, and nothing here is legal advice. This is our read on where the law and the litigation actually stand in 2026, and what it means for how you implement AI in hiring.

What actually happens once a challenge is filed

A challenge to an AI-influenced hiring decision rarely stays a dispute about one candidate. Once litigation starts, both the employer and the vendor whose tool touched the decision are typically named, on the theory that a vendor performing a hiring function is acting as the employer's agent, not a neutral outside supplier.

Discovery then reaches past the individual rejection into the tool itself: what data trained or informed it, what criteria it weighted, and how consistently it produced the same pattern across every applicant who looks like the plaintiff, not just the one who complained. That last part usually turns into a statistical check: did the tool reject people in this group more often than it rejected everyone else?

In other words, this means the case stops being about one candidate's rejection and becomes about a pattern across everyone the tool ever screened. None of this depends on whether anyone intended to discriminate. It depends on what the pattern actually shows and what record exists of how the decision got made.

The part that decides how the conversation goes

In the governance and steering committee structures I've been part of, documenting the rationale behind a decision was never optional. Every consequential call needed a written record of what was considered, who signed off, and why the outcome landed where it did. It was a simple decision log but was a record we could go back to. That discipline wasn't built with litigation in mind. It existed so a decision could be explained months later to someone who wasn't in the room when it happened: a new team member, an internal auditor, a leadership review.

HR teams already do a version of this. Interviewers take notes, hiring managers talk through why one candidate beat another, and some version of the rationale usually exists somewhere, even if it's informal and scattered across emails or a shared drive. What changes once an AI tool is making or influencing the screening decision is that this note-taking has to become formal and consistent. The informal version that worked when a person made every call doesn't hold up the same way once a tool is doing part of the screening at scale.

That's the exact muscle a legally challenged hiring decision tests. An organization that can already produce a clear record of what happened and why is having a hard conversation. One that can't is often trying to reconstruct, after the fact and under pressure, a decision nobody wrote down the first time.

What ends up under scrutiny once a challenge is filed

The legal landscape is a patchwork, not a gap

Federal guidance on AI in hiring didn't get stronger in 2026, it got thinner. The EEOC withdrew its formal AI-hiring guidance in early 2025 under the new administration, and no replacement has taken its place. But withdrawn guidance isn't the same as no law. Title VII, the ADA, and the ADEA still apply to a hiring decision whether or not AI was involved in making it, and Mobley is being litigated entirely on that existing statutory ground, not on guidance that no longer exists.

States are filling the visible gap at different speeds. New York City's Local Law 144 already requires an independent bias audit before an automated hiring tool gets used. Illinois' HB 3773 took effect January 1, 2026, extending the state's existing human rights law to AI-assisted employment decisions. Colorado's broader AI Act, by contrast, got stayed by a federal court in April 2026, with a narrower replacement bill not set to take effect until January 1, 2027, if signed into law. Three states, three different postures, in the same calendar year.

Enforcement doesn't close that gap as cleanly as a rulebook suggests, either. A New York State Comptroller's audit of Local Law 144 enforcement found the city agency responsible for checking employer compliance identified just one issue of non-compliance across 32 companies reviewed, while an independent audit of those same companies found 17, roughly six percent of what was actually there. A quiet enforcement record doesn't mean a low-risk practice. It can just as easily mean nobody's looked closely yet.

What this means in practice: there's no single national rulebook to check your process against, and a quiet enforcement record in your state doesn't mean your hiring tool is compliant or safe. The patchwork puts the burden back on each organization to look at its own practices directly, rather than waiting for a regulator to flag the problem first.

What your own file would show if someone asked today

None of that guarantees a challenge is coming. It just determines what the organization would actually have to work with if one did.

Security & Risk is one of the five pillars the RAISE OS™ AI Maturity Assessment measures on its own, separate from Responsibility & Governance, because a company can have reasonably clear ownership of AI decisions and still be carrying legal exposure nobody has actually tested. Ownership answers who's accountable for an AI-assisted decision; what a legal challenge tests is whether that accountability ever produced a record anyone can stand behind.

That's not a pass/fail security check, it's a broader read on how prepared your organization actually is across all five pillars.

An AI-influenced hiring decision has already been challenged in federal court. Here's what a legal challenge actually tests, and where the law stands in 2026.

What happens if an AI-influenced hiring decision gets legally challenged?

A department lead signs up for a new AI tool with a company card because it solves a real problem faster than waiting on IT’s review queue. Multiply that by every team with a deadline, and the answer to whether a company needs an approved AI tools list becomes obvious: yes. What’s less obvious is that the list itself does almost nothing on its own. A document nobody checks before adopting a new tool isn’t a control, it’s a formality. What actually determines whether shadow AI stays a manageable exception or becomes the default way tools enter the company is what happens after someone wants something that isn’t on the list yet.

I haven’t sat on a committee that maintained an approved vendor list, but I’ve watched both ends of this problem play out. At a company with more than 10,000 employees, getting any new system approved meant a genuinely bureaucratic process, real oversight, but slow enough that it visibly discouraged people from trying anything new.

At a company with around 200 employees, the opposite was true: anyone could adopt any tool, at any moment, with nobody checking. That flexibility felt good in the moment, but IT ended up tracking more than 200 different applications, many doing the exact same job at different price points, with no real handle on what the company was actually paying for or who had access to what.

Neither extreme worked. The large company’s process caught real problems but pushed people toward workarounds anyway. The small company had none of that friction, and paid for it in duplicate spend and untracked access instead. What actually works sits in the middle: real oversight, moving fast enough that going around it is never the easier option. That’s the same balance an approved AI tools list has to strike, on a much shorter adoption cycle than the enterprise software those two companies were dealing with.

IBM’s 2026 Cost of a Data Breach Report, based on 602 breached organizations across 17 industries and 16 countries, found that 43% of security incidents in the past year involved shadow AI, tools nobody had approved, more than double the share from the year before. Over two-thirds of the organizations in the study had no governance process in place to limit that kind of unapproved use at all. Unapproved tools aren’t a rare exception anymore. In a large share of companies, they’re already the majority path AI takes into the business, which makes an unenforced list less like a minor gap and more like an open door with a sign next to it nobody reads.

Mid-market companies aren’t dealing with one or two unofficial tools scattered around. Cledara’s 2026 platform data on companies with 30 to 500 employees found 80% already pay for at least one AI tool outright, averaging 4.1 paid AI subscriptions per company, ranging from 2 to 4 tools at the smallest companies up to 8 to 15 or more once a company reaches the 150-to-500-employee range. Every one of those tools got adopted by someone who needed it faster than an approval process could deliver. A slow list doesn’t prevent that. It just guarantees the tool arrives without anyone reviewing it first.

Four things that actually make a tools list stick

The real gap isn’t the list, it’s visibility

Retool’s 2026 State of AI Governance survey of 307 senior technology and security leaders found that only 5% felt very confident they had full visibility into which AI tools were actually running in production at their company, while 43% said they had no such confidence at all. A list can only govern what someone actually knows exists. Enforcement without visibility isn’t enforcement, it’s a policy nobody has the information to check.

Ask three questions. Is there a specific person, not a committee, who owns approving new requests? Is that review faster than the time it takes to sign up for the tool directly, with no approval at all? And has anyone actually gone looking for AI tools already in use that never went through the process? A no to any of these means the list exists on paper, but the enforcement behind it doesn’t.

Where enforcement fits into managing AI risk

An approved AI tools list is a starting point for Security & Risk, one of the five pillars the RAISE OS™ AI Maturity Assessment measures, but it’s the visible layer of a much larger control, not the control itself. Shadow AI is already inside most mid-market companies whether a list exists or not, and it stays that way until someone owns making enforcement faster than the workaround.

An approved AI tools list only works if enforcement is built into it. Here's what actually makes one stick, with 2026 data on why most companies still can't enforce their own AI policy.

Do we need an approved AI tools list, and how do we actually enforce it?

I don't think most shadow AI use is reckless. It's usually what happens when people are given a real problem and no sanctioned way to solve it fast enough. Shadow AI is any AI tool or use inside your company that leadership hasn't approved, reviewed, or even necessarily knows about, from someone pasting client data into a personal ChatGPT account to a browser extension that quietly adds AI features to a tool you already pay for.

I take this risk seriously because I've been on the wrong side of it. A few years ago I used a free AI tool on work material without realizing the data could be used to train the model. Nothing came of it, but it was a real exposure I created without meaning to, and I only caught it after the fact. For a mid-market company, the risk isn't hypothetical: a 2026 workplace survey found 66% of office professionals had used an AI tool despite believing it violated company policy, and 88% had shared work-related information with a public AI tool.

Why this is a mid-market problem, not just an enterprise one, in 2026

Shadow AI spreads fastest where there's no dedicated team watching for it, which describes most mid-market companies more than it describes large enterprises. A big company might have a security team that notices unusual data flows. A mid-market company usually doesn't, so shadow AI use can run for months before anyone in leadership becomes aware it's happening at all.

That gap is exactly what the 2026 PagerDuty Shadow AI survey captured: 81% of employees believed leadership operated under different AI rules than everyone else, and 77% felt company AI restrictions limited their professional growth. Employees aren't hiding AI use out of malice. They're routing around a policy vacuum because the tools are useful and nobody's given them an approved alternative.

Shadow AI is broader than someone secretly using a chatbot. It includes AI features toggled on by default inside software you already license, browser extensions that summarize or rewrite content using an AI backend, a team adopting a free-tier AI tool without procurement ever reviewing it, and contractors or agencies running your company's data through their own AI tools under no policy that applies to them at all.

Any one of these can exist in a well-run company. Several at once, with no one tracking them, is what turns shadow AI from an inconvenience into exposure.

Concrete, not hypothetical. Among employees who shared information with a public AI tool, 43% shared emails or correspondence, 40% shared meeting notes, 34% shared customer data, and 31% shared financial information or confidential documents, according to the same 2026 survey. Separately, IBM's 2025 Cost of a Data Breach report found that shadow AI was a factor in 20% of breaches at organizations studied, adding roughly $670,000 to the average cost of those breaches.

Those numbers describe two different but related exposures: data leaving the company through everyday, well-intentioned AI use, and the added cost when that exposure turns into an actual breach.

Not on its own, and it can make things worse. The same 2026 survey found 75% of employees would consider leaving for an employer that offered better AI tools and training, a number that climbed to 80% at large companies. A ban without an approved alternative tends to push AI use further underground rather than eliminate it, since the underlying reason people reach for these tools, genuine time savings, doesn't go away because the policy changed.

What actually reduces the risk is narrower: a short list of approved tools, a clear line on what data can and can't go into them, and enough visibility that leadership would actually notice if that line got crossed.

This is close to a problem I watched play out under a different label: shadow IT. Long before "shadow AI" was a phrase, teams in digital transformation programs I worked on would quietly build or use digital tools that solved real problems; a content development tool, a project tracker, a file-sharing app, a scheduling tool, without ever routing it through IT or procurement. Nobody was trying to create risk. They just needed something that worked, and the sanctioned option was slower or didn't exist yet.

The fix was never a blanket ban. It was giving teams a fast, reasonable path to get a tool approved, paired with enough visibility that unsanctioned tools didn't just accumulate silently for years. Some of those unofficial tools turned out to be worth adopting company-wide once someone actually looked at them. Others got retired the moment a reviewer noticed what data they touched. Neither outcome was possible while the tools stayed invisible.

Shadow AI is the same dynamic showing up again, with higher stakes attached to the kind of data now involved. The instinct to reach for a faster tool hasn't changed. What's changed is how much a single unreviewed tool can now see.

How to reduce the risk without killing legitimate use

Start with visibility before restriction: an honest inventory of what AI tools people are already using, gathered without punishing anyone for admitting it. From there, a short approved list covering the most common use cases removes most of the reason to go around policy in the first place. Pair that with a plain rule on what data categories can never go into a public AI tool, and a simple, fast way to request a new tool get reviewed, so the approved list keeps pace with what people actually need.

Shadow AI sits squarely inside Security & Risk, one of the five pillars the RAISE OS™ AI Maturity Assessment measures, and it connects directly to who owns AI governance in the first place. A company can't manage a risk it can't see, and visibility, not prohibition, is what actually closes the gap.

Shadow AI isn't a hypothetical risk. Here's what actually counts as shadow AI, how exposed a mid-market company really is in 2026, and what reduces the risk without banning tools people rely on.

Security & Risk

How do we manage AI risk without a dedicated compliance or security team?

What happens if an AI-influenced hiring decision gets legally challenged?

Do we need an approved AI tools list, and how do we actually enforce it?

What is shadow AI and how big a risk is it for a mid-market company?