AI is everywhere in leadership meetings, and yet it is almost absent from the teams’ real work. The distance between the two is exactly the size of the job ahead, and its cost has already been measured:
95% of generative AI projects in companies produce no measurable return (MIT Project NANDA, August 2025).
On paper, everything is in place. Copilot licenses deployed, budgets approved, AI committees set up, policies published. On the ground, teams rarely open the tool, don’t know what to do with it, or use it on the sly with their personal account. The result: a low ROI and projects that stop after 6 months.
The good news is that the barriers to AI adoption at work can be identified and named, and above all, you can act on them.
Why does AI fail at work? The 4 major barriers
Barrier 1: AI licenses that go unused
Companies invest in Copilot, Gemini or ChatGPT Enterprise. The rollout is massive, often “for everyone”, with no targeting. Use doesn’t follow. Only 35.8% of users with a Microsoft 365 Copilot license actually use it: nearly 64% of seats remain dormant or significantly underused (Lighthouse Global, analysis of Microsoft adoption data, 2026).
At €360 per user per year, the math is quick: for 1,000 licenses, that is €230,000 of budget producing no value. And when renewal time comes, leadership has neither proof of use nor tangible ROI to justify the spend. The tool was there, but nobody ever knew what to do with it in their daily work.
Barrier 2: no concrete business use cases
For people on the operational side, AI often remains a fuzzy concept. A generic tool presented in a training session, tried twice and forgotten. Because nobody made the link between that tool and their own pain points.
Without precise use cases (writing this type of report, preparing this recurring reporting, answering this category of customer emails), AI remains an abstract promise. And in a company, abstract promises are what die fastest. Teams conclude: “we tried it, it’s useless”, and the momentum fades. But behind the “it’s useless”, there is often a fear nobody listened to. Here is how to defuse it: AI literacy at work: how to defuse fear and lift adoption
Barrier 3: shadow AI with no structure
While leadership teams debate the strategy, the teams don’t wait. They already use AI, but informally, with no framework and without sharing.
49% of employees admit to using AI tools their company hasn’t approved, and 58% turn to free versions, which often lack the security, data governance and privacy guarantees that are essential in a professional setting (BlackFog study, 2026).
In practice: ChatGPT in a hidden tab, improvised prompts, personal accounts, customer data copied and pasted into free tools. Two heavy consequences. First, a risk: security, GDPR compliance, intellectual property. Second, a dead loss: good practices stay in each person’s head, none of them are captured, and the company silently pays for know-how that never spreads.
Barrier 4: a lack of in-house skills
AI isn’t a technology you launch at the push of a button. It is a tool you guide with precise instructions: prompts. Teams don’t know how to turn their expertise into instructions an AI can use. The result?
- Low-quality deliverables from the AI (because the prompts are vague)
- Time lost in trial and error → frustration → people give up
- No organized upskilling: everyone learns alone in their corner
The 4 levers to unblock AI adoption
Lever 1: target adoption on specific business uses
Stop rolling out AI licenses “to everyone”. Target 2 or 3 priority jobs, the ones with the most repetitive pain points and the highest volumes, and concentrate licenses, training and support there for 3 months.
The principle: 200 licenses fully used are better than 2,000 dormant ones. ROI is built through depth of use, not breadth of rollout. And this targeting brings an essential side benefit: the first successes become visible internal references that naturally pull the other jobs along.
Lever 2: identify use cases from the front line (bottom-up)
The relevant use cases aren’t decreed by the leadership team. They already exist in the heads of employees who, every day, spot the repetitive tasks, the low-value reports and the customer responses that could be assisted.
The bottom-up method means bringing these pain points up from the teams themselves, qualifying them (impact, frequency, feasibility), then picking 3 to 5 of them to turn into use cases with the right tools, ready to be shared. That is the logic of the AI idea box: the front line names the problem, and AI becomes the concrete solution. You no longer sell the tool, you solve a pain point identified by the person who lives with it.
Lever 3: govern and structure shadow AI rather than banning it
Banning shadow AI is a losing strategy: teams will keep doing it, just more discreetly. The strategy that works is the opposite: bring shadow AI out of the shadows and into a framework.
In practice: give teams an approved AI tool, a simple policy (data allowed, data forbidden, risky uses) and above all a shared space to exchange good prompts. What circulated in secret becomes a collective resource. The company gets value from what it pays for, reduces its risk of leaks, and turns a behavior it was putting up with into a governed practice.
Lever 4: move from theory to action: train your teams on the right prompts and the right uses
AI training can no longer be a one-hour e-learning module watched in a hurry. It has to be short, repeated, led by peers and anchored in real business cases. A salesperson learns on their sales scripts. A financial controller learns on their reports. An HR business partner learns on their job postings.
To go further on this point, we have laid out a complete method in this article: Training your people on AI: the number 1 lever for adoption.
The right metric isn’t the number of training hours delivered. It is the number of employees able, within 30 days for example, to write a prompt that solves a concrete business pain point. As long as that number doesn’t move, the licenses stay dormant.
Conclusion
The barriers to AI adoption at work are neither technical nor budgetary. They are organizational: poorly targeted licenses, fuzzy use cases, shadow AI nobody builds on, skills left to self-teaching. Each one has its lever. Together, they make a simple method: start from the front line, equip concrete cases, train through peers, measure what moves.
AI will never pay off until the teams have it in hand. Start with them.
FAQ: your questions about the barriers to AI adoption at work
Why isn’t AI working at my company?
If AI isn’t working in your company, it is probably neither a tool problem nor a budget problem. It is an organizational problem: licenses rolled out too broadly with no targeting, no precise business use cases, shadow AI nobody builds on, and employees left alone with the tool and no prompting method. The result: 95% of generative AI projects don’t reach a measurable ROI. The solution lies in targeting specific jobs, finding the use cases bottom-up and training through peers.
What are the main obstacles to AI in 2026?
There are four main obstacles to AI adoption at work in 2026: (1) massive underuse of licenses: only 35.8% of Copilot users actually use it; (2) no concrete business use cases, which leaves AI as an abstract promise; (3) shadow AI with no structure: 49% of employees use AI tools that haven’t been approved; (4) a lack of in-house prompting skills, which produces low-quality deliverables and leads people to give up. These obstacles are organizational, not technical.
What are the solutions to AI problems at work?
The solutions rest on four levers: target 2 or 3 priority jobs rather than rolling out to everyone, bring the use cases up from the front line (the logic of an AI idea box), govern shadow AI rather than banning it by offering an approved tool and a space to share prompts, and train teams in a way that is short, repeated and anchored in their real business cases. The right KPI: the number of employees able to write a prompt that solves a concrete pain point.
