I hear a version of this question from almost every owner or operations leader I sit down with. They approved the spend, IT provisioned the seats, and six months later they cannot point to a single measurable result. That is not a failure of the tool. It is what happens when a company treats AI adoption as a purchase order instead of a capability-building program.
Direct answer: your team needs structured AI training before another AI license because capability, not access, is what determines whether the tool produces value or risk. Untrained employees either avoid a paid tool or route around it with unapproved ones, and both outcomes cost you more than the license did.
Most leaders assume the sequence does not matter much: buy the platform, roll it out, backfill training later if it seems necessary. Having watched that sequence play out across enough SMBs and public-sector environments, I can say plainly that it is backwards, and not just a minor inefficiency. Handing a powerful, general-purpose AI tool to an untrained workforce is the decision that creates the exposure in the first place.
When people do not know what a tool is good for, they either underuse it or misuse it. Underuse means the license sits mostly idle while employees keep doing things the old way, and finance can never justify the renewal. Misuse means employees quietly paste client contracts, financial figures, or source code into whatever consumer AI tool solves their immediate problem, whether or not it is the one you paid for. Either way, you have converted a technology purchase into dead weight or unmanaged risk, and you paid for the privilege.
This is the pattern researchers at MIT's NANDA initiative documented at scale. Their 2025 study of enterprise generative AI programs found roughly 95% of organizations saw no measurable profit-and-loss impact from their AI pilots, despite tens of billions in collective spending. The bottleneck was not model quality; it was what researchers called a "learning gap," meaning organizations never built the internal capability to apply the tools to real workflows. The 5% that succeeded were not using better AI. They were using AI inside teams that had actually learned how.
You get shadow AI, and the data on how far that has already spread should change how any leader thinks about sequencing. A national survey covered by the Journal of Accountancy found that 59% of employees use AI tools their employer never approved, and among executives and senior managers specifically, that figure jumps to 93%. Three-quarters of the employees using unapproved tools admitted to feeding them potentially sensitive information: employee records, customer data, internal documents. This is not a rogue-employee problem. It is a leadership-visibility problem.
WalkMe's second annual workplace AI survey, published through SAP's newsroom, quantifies the training side of that gap directly. Only 7.5% of employees reported receiving extensive AI training, up just half a percentage point from the year before, while 78% admitted to using AI tools that were never approved by their employer. Daily AI use climbed 16 points year over year. Training barely moved. That divergence is the entire thesis in one data set: usage is accelerating on its own, and organizations are not building the guardrails to match it.
Microsoft's own Work Trend Index found something similarly uncomfortable for leadership teams: only 39% of employees who use AI at work have received any company-provided training on it, and just 25% of companies planned to offer generative AI training at all that year. Three out of four employees are teaching themselves how to use tools that touch client data, financial records, and proprietary business logic, with no framework for what is and is not safe to type into them.
It is not theoretical. In 2023, engineers inside Samsung's semiconductor division pasted confidential source code, internal meeting transcripts, and equipment test sequences into ChatGPT across three separate incidents within a single month, seeking quick fixes to real problems, according to reporting compiled by CIO Dive and documented in the AIAAIC incident repository. None of those employees set out to leak anything; they were solving a task the fastest way available, without any framework for what should never leave the building. Samsung's response was a full ban on generative AI tools on company devices, an emergency 1,024-byte prompt limit, and a disciplinary investigation into its own engineers.
Samsung is instructive precisely because it is one of the most technically sophisticated manufacturers in the world, and the failure still happened at the individual employee level, driven by ordinary task pressure rather than malice. If a company with Samsung's security resources had this happen three times in twenty days, an SMB with no AI acceptable use policy and no training program is not managing that risk. It simply has not discovered it yet.
The financial exposure compounds the operational waste. IBM's cost-of-a-breach research, cited in that same Journal of Accountancy reporting, found that breaches traced to shadow AI cost organizations $670,000 more on average than breaches involving sanctioned, governed AI systems. That gap exists because shadow AI incidents are harder to detect, scope, and explain to a regulator or a cyber insurance underwriter after the fact. You are not choosing between a licensing cost and a training cost. You are choosing between a training cost now and a larger incident cost later, layered on top of the licensing spend you already made.
Meanwhile, adoption is not slowing down to wait for anyone's governance program. The Census Bureau's Business Trends and Outlook Survey shows overall U.S. business AI use running between 17% and 20% as of early 2026, with larger firms adopting fastest and smaller firms roughly a year behind but closing the gap. The tools are already inside the building, sanctioned or not. The only real decision left to a leader is whether employees learn to use them under a framework you built, or under no framework at all.
In practice, the same three gaps show up in almost every SMB that treats AI adoption as a procurement event rather than a readiness program:
Each of those gaps is a training problem, not a tooling problem, and each one is fixable before you spend another dollar on seats.
Sequence training ahead of the next purchase, not behind it. That means giving employees a working framework for what data can and cannot go into an AI tool, role-specific guidance for how their function should and should not use it, and a channel to ask questions before they improvise an answer on their own. This is the same discipline we walk clients through when we help them build an AI acceptable use policy, because a policy without trained people behind it is a document nobody reads until after something goes wrong.
The leadership teams that get this right treat AI adoption the way they would treat any other operational rollout with real risk attached, the way our team lays out in our leadership playbook for AI adoption: define the use case, train the people who will touch it, govern the exceptions, then scale. Skipping straight to "scale" is exactly why the MIT NANDA researchers found so much spend and so little return. It is also why most SMB AI initiatives fail in ways that are fully predictable in advance.
This is also why I keep telling clients that AI adoption is a governance problem before it is ever a technology problem. The model you choose matters far less than whether your people know how to use any model safely, and that capability has to exist before the rollout, not get bolted onto it after the first incident. It is also the exact question we work through with clients in our guide to whether employees can safely use ChatGPT and Copilot at work, and the honest answer is always: only once they have been trained on how.
Concretely, the order matters more than the vendor. Train people on data handling and role-specific use cases first, build the acceptable use policy alongside that training, then provision licenses once your team already knows what they are for. Reversing that order produces the shallow-usage-plus-shadow-AI outcome documented across every survey above.
| Sequencing choice | What typically happens | What it costs you |
|---|---|---|
| Tool first, training later | Shallow sanctioned usage, high shadow AI usage, inconsistent output quality | Wasted license spend plus unmanaged data exposure |
| Training first, tool second | Employees know what to use, when, and what never to input | Slower initial rollout, dramatically lower incident risk and higher realized ROI |
The right-hand column is not a hypothetical. It is the difference between the 5% of organizations MIT found extracting real financial value from AI and the 95% that were not, and the difference tracked directly to whether the organization built capability before it scaled access.
When we bring on a new SMB or regulated client, we do not start by recommending a platform. We start by mapping what AI tools employees are already using, sanctioned or not, and what data has already moved through them, the same discovery process our team describes in our piece on supervising your AI intern. That baseline usually surprises leadership, because shadow AI is, by definition, invisible until someone looks for it.
From there, we build role-based training tied to how each department actually works, paired with an acceptable use policy that reflects your real regulatory obligations, whether that is HIPAA, CMMC, or general data privacy exposure. Only after that foundation exists do we help evaluate which licensed tools are worth paying for and how to configure them so guardrails are built into daily workflow, not enforced after the fact through a policy nobody remembers reading.
If your organization already has AI licenses sitting underused, or if you suspect employees are getting work done through tools nobody approved, the fix is not a better tool. It is giving your people the training and framework they were never given before the first tool arrived.
If your team is moving faster with AI than your guardrails are, start with structured training rather than another tool. Securafy AI University gives your people role-based AI training with security built into the material, not bolted on afterward.
If you would rather talk through your specific environment first, book a strategy call with Securafy and we will walk your current AI usage, exposure, and the fastest path to safe adoption.