THINK BIG
AI Adoption · Stalled Rollouts · Usage Rate
Rolled Out AI and Nobody Uses It? The Problem Usually Isn't the AI. Work Backward From the One Task Employees Actually Do Every Day
When AI rollouts go unused, it's rarely because the model isn't smart enough. More often, nobody decided what it was for, nobody showed employees how to use it, and it never made it into the daily work. The fix: work backward from one task employees actually do every day.
1. What's happening: bought, installed, activated, and then ignored
Plenty of companies have been through this. The boss sees how impressive AI is, buys a tool, sets up accounts, and holds a kickoff session. Two or three weeks later, employees are still doing things the old way, and hardly anyone clicks into the AI.
It isn't just a few companies. Surveys in Taiwan and abroad show the same pattern:
Taiwan: the biggest roadblock is "we don't know what to use it for." As reported by Taiwanese media, the 2025 Survey on AI Use Among Small and Medium Enterprises, commissioned by Taiwan's Small and Medium Enterprise and Startup Administration (Ministry of Economic Affairs) and conducted by the Industrial Technology Research Institute, collected 1,207 valid responses from companies with 199 or fewer employees. The top challenge SMEs cited when considering AI was "no clear use case yet," followed by not understanding AI, high costs, talent shortages, and difficulty measuring results. The same survey found that only about 15% of SMEs have AI talent, and roughly 75% of that talent are entry-level users of off-the-shelf tools. (Source: China Times Want Rich, 2025-12-26)
Globally: most companies haven't seen results yet. An international management consulting firm surveyed 1,000 senior executives in 2024 and found that 74% of companies had yet to show tangible value from AI. Only 26% had built the capabilities to move beyond proofs of concept and generate real value. (Source: BCG press release, 2024-10-24)
People are using AI, just not the company's. A large software company's 2024 survey of 31,000 knowledge workers in 31 markets found that 75% already use generative AI at work, but 78% of those users bring their own AI tools. Only 39% of people who use AI at work have received AI training from their company. (Source: Microsoft Work Trend Index 2024)
Put the three together and the picture is clear: employees aren't refusing AI. The setup their company handed them just doesn't connect to the work they actually do every day.
2. The debate: is the problem the AI, or the people and processes?
Those who say it's about processes and organization have plenty of evidence. In the same report, the consulting firm describes a "10-20-70" principle: leading companies put 10% of their resources into algorithms, 20% into technology and data, and 70% into people and processes. (Source: BCG press release)
A 2025 preliminary report from a research initiative at MIT reached a similar conclusion. As reported by Fortune, it found that most companies' generative AI efforts were falling short, and that the key issue wasn't the quality of the AI models but a "learning gap" between tools and organizations, along with flawed enterprise integration. (Source: Fortune, 2025-08-18)
Skeptics raise two cautions: 1. Read the numbers carefully. The MIT report is preliminary research, and coverage describes its sample and methods inconsistently. It shouldn't be taken as proof that every AI project fails. 2. Tools still matter. If an AI often gets answers wrong, gives no sources, or responds slowly, employees will try it once or twice and stop. No amount of training can save a tool that isn't good enough.
The two views don't conflict: a good-enough tool is the baseline. Once you clear that bar, what decides whether people use it is where it's used, who uses it, and how.
3. What we've learned doing this work
Think BIG deploys AI agents for businesses and keeps reviewing usage logs after handover. Here's what we've learned, with no client details:
1. Unused systems all show similar symptoms. Employees stick with the old way. They ask once, get a wrong answer, and never ask again. The entry point is buried behind several logins. Nobody knows what to ask. Answers come without sources, so people don't dare rely on them. Permissions are misconfigured, so people can't see documents they should. Almost none of this has anything to do with how smart the model is.
2. Work backward from one task, not forward from a pile of features. Before we start, we look for a task that meets three conditions: it happens every day, its rules can be written down, and the cost of a wrong answer is manageable. Think "sales looking up product specs" or "new hires checking internal policies." Get that one task used every day first, then move to the second.
3. Write acceptance questions the way employees actually ask. We prepare a set of acceptance questions with known answers, phrased as closely as possible to how employees really talk, including abbreviations, casual wording, even typos. A system that aces neatly worded test questions but stumbles on real ones is a very common gap.
4. Put the AI where employees already are. Every new website or account adds friction. When the AI lives inside a chat tool or system employees already use daily, usage tends to be much better. One caveat: when you connect an outside chat channel such as LINE, messages through that channel pass through that service, so make that clear before rollout.
5. Say "I don't know" when it doesn't know, and cite sources when it does. Whether employees use it comes down to trust. When an answer shows which document it came from, employees can click through and check. Saying "I couldn't find that" builds more trust than making something up.
6. After handover, review the logs every week. What employees asked, which questions got weak answers, which documents need to be added: it's all in the logs. Name one internal point person to spend a little time each week reviewing logs and adding material, and the system gets more accurate the more it's used. That's also the most practical way to close the "learning gap" mentioned above.
Our biggest takeaway: whether an AI rollout succeeds is often decided before the work begins. Get clear on which one task the AI is for, and most of the later problems disappear.
4. Our recommendations: answer these five questions before you roll out
1. What's the most time-consuming, most annoying task employees do every day? Ask the front line directly, not just managers. 2. Does that task follow rules? Work with documentation and standard answers is the best first job for an AI agent. 3. What happens if the answer is wrong? Start with tasks where mistakes are manageable. For high-stakes work like external quotes or legal judgments, have the AI draft and a person confirm. 4. Where will employees use it? Put it inside tools they already use, and open as few new entry points as possible. 5. Who reviews the logs and adds material? Name one internal point person. They don't need to be technical, but they need to know the process and be able to make decisions.
| Start with features | Work backward from one task | |
|---|---|---|
| Starting point | What the tool can do | What frustrates employees every day |
| First month | Lots of features; nobody knows where to start | One task used by someone every day |
| Acceptance | Judged on the demo | Tested with questions employees actually ask |
| After handover | Nobody reviews logs; usage fades | Weekly log reviews and new material; it gets more accurate over time |
Bottom line: if nobody uses your AI, don't swap the AI first. Swap the question.
Further reading: if nobody inside your company can own the work of wiring AI into your processes, see Think BIG NEWS Issue 007, "What Is a Forward Deployed Engineer?", on how small businesses can fill that gap with an embedded-style approach.
Start with one task. Frozen scope. Delivery within 30 days.
(The following reflects Think BIG's own position and is separate from the neutral reporting above.)
We act as a company's remote FDE (embedded engineering) team, using TB FRAME, our four-stage method (assess, build, verify, maintain), to wire AI agents into your work: Start with one task: during assessment, we identify the one rule-based task employees actually do every day, then decide what the AI agent will and won't do. Frozen scope, delivery within 30 days: scope and acceptance questions are written down before work starts, and delivery happens within 30 days once scope is frozen and data is in place. Acceptance in your employees' words: we test against known answers before handover; answers cite sources, and the agent says "I don't know" when it can't find something. First year of maintenance: all three plans include first-year maintenance; after handover, we keep reviewing logs, adding material, and sending proactive alerts.
We don't promise sales results. We promise to wire AI into your actual work and let you see that it's running.
FAQ
- Q1: Why don't employees use AI after the company rolls it out?
- The most common reason isn't that the AI isn't smart enough. It's that nobody decided which task the AI was for, nobody taught employees how to use it, and it never became part of the daily workflow. Answers without sources and hard-to-find entry points also make employees give up after a try or two.
- Q2: What's the biggest barrier to AI adoption for small businesses in Taiwan?
- According to media reports on the 2025 Survey on AI Use Among Small and Medium Enterprises, commissioned by Taiwan's Small and Medium Enterprise and Startup Administration and conducted by the Industrial Technology Research Institute, the top challenge SMEs cite when considering AI is "no clear use case yet," followed by not understanding AI, high costs, talent shortages, and difficulty measuring results.
- Q3: Where should an AI rollout start?
- With one task: find work employees repeat every day, whose rules can be written down, and where a wrong answer is manageable, such as looking up product specs or internal policies. Get that task used every day, then expand to a second.
- Q4: How can you tell whether an AI rollout actually worked?
- Don't judge by the demo. Build acceptance questions in the words employees really use, and confirm that answers are correct, cite sources, and say "I don't know" when nothing is found. After handover, keep reviewing usage logs to see whether people use it and which questions get weak answers.
- Q5: Will employees avoid AI because they're worried about their jobs?
- They might, which is why communication matters. Start with the repetitive, time-consuming work employees find most frustrating, so the AI saves them time, rather than starting with work that depends on their professional judgment. When employees see AI lightening their own load, they're usually much more willing to use it.
This article combines industry reporting with Think BIG's first-hand practice. All cited surveys and reports are linked; figures reflect each source's published results, and the MIT research is a preliminary report cited through press coverage with its limitations noted. Companies are described rather than named in the text, with links in the sources; nothing here evaluates any company or product. Our first-hand experience is shared without client details.