Over the past two years, we have delivered more than 20 enterprise AI projects across manufacturing, trade, retail chains, legal services, short-form drama MCNs, knowledge commerce, and other industries. Looking back, failed projects almost always stumble at the same point: teams focus on “which tool to choose” instead of “which problem to solve first.”
A typical scenario is this - the company bought a bunch of AI tools and asked instructors to conduct two training sessions for all employees. The business teams were excited to try them during the training, then returned to their desks and never opened them again. Three months later, management asked "What exactly has AI brought?" and no one could answer.
It’s not a staff problem or a tool problem; it means the wrong use case was chosen.
Why “training first and finding a use case later” fails
Training addresses “knowing,” but business results come from “using.” The gap between the two includes data access, process changes, ownership, and a way to measure results.
We call this an implementation gap. Unless it is closed, more training only widens the distance between learning and use.
A control group from real delivery: When a traditional manufacturing company did AI training for the first time, everyone attended, but the training made little difference afterward; the second time, we changed the approach: instead of starting with tools, we first spent a week to break down the specific pain point of "Amoeba salary calculation takes 4 hours", and then taught targeted lessons. As a result, the accounting process was cut from four hours to five minutes—a 98% improvement. The same group of people, the same tools, the only difference is whether the use case is chosen correctly.
Six criteria for selecting a use case
We use a six-dimensional scoring table internally to screen use cases. Each item is scored from 1 to 5. We do not proceed with use cases that score below 20 in total:
| Dimension | Question to ask | What a high score looks like |
|---|---|---|
| Frequency | How often does this happen? | Performed daily or weekly |
| Time required | How long does it take each time? | More than 30 minutes per run |
| Clarity of rules | Can you write clearly "what counts as done"? | Clear acceptance criteria exist |
| Data availability | Where are the materials needed to do this? | Already digitized and accessible |
| Fault tolerance | Is it costly to make mistakes? | Errors can be caught through manual review |
| Responsible person | Who is responsible for the results? | A clear business owner exists |
After pulling out these six items and scoring them, we quickly filtered out those ideas that "sounded cool but couldn't be implemented." Instead, some seemingly inconspicuous use cases score highly – for example, cross-system order lookup, a salesperson has to move through several systems for a single lookup, taking hours on average. After we completed the changes, the query time fell to seconds, a 90% reduction. Another example is export-sales email responses: templates plus AI-assisted drafting improved response efficiency by 50%.
These use cases aren't sexy, but they take a real human toll on a daily basis.
A four-step implementation method
After selecting the use case, the advancement path can actually be standardized. What we settled on are these four steps:
- Map the process: Write the task as a sequence of steps from beginning to end, marking the input, output, and judgment rules of each step. Many companies get stuck at this step and find that they have not explained the process clearly.
- Divide the work: Divide the process into "what AI can do" and "what humans must do". The principle is to let AI handle repetition, retrieval, drafting, and comparison, let AI output the first draft, and let people make judgments and signatures.
- Define acceptance criteria: Agree on indicators before going online, such as time required, output volume, and error rate. Transformation without indicators will definitely turn into "feeling faster" in the end.
- Run for two weeks and then adjust: There will definitely be unexpected situations in the first two weeks. It is necessary to leave a window for adjustment instead of announcing the release immediately after going online.
A typical example from a legal service company: Lawyer’s letters originally took hours to produce. After we handed over the repetitive parts of document generation to AI, they could be generated in batches within minutes: document output rose by 80%, while administrative hours spent organizing evidence fell by 50%. Note that what decreased was administrative work, not a lawyer’s judgment – this is a healthy improvement.
Three things that must be done at the organizational level
In addition to technology and process, there are three things that determine whether the project can survive:
- Let business leaders be project owners, not the IT department. The implementation of AI is business transformation, not an information technology project.
- Prove the model in one department before replicating it elsewhere. A light-aesthetic beauty chain needed differentiated marketing materials for 2,000+ stores. Rather than launching everywhere at once, it first validated the approach with a regional pilot.
- Write "using AI" into daily actions instead of making it an extra task. For example, the generation links of daily reports, weekly reports, and business analysis are directly embedded into the original process. In a certain chain project, the regional manager's business report has been viewed in real time on the mobile terminal, instead of having to be sorted out by humans.
Three common questions
Q: Can we use AI if our company has a weak data foundation?
Yes, but the entry point needs to be narrower. When the data foundation is poor, give priority to scenarios where "a single point, a single document, and a single person can complete", such as contract-clause comparison, first-draft copywriting, and meeting-minute preparation. Wait until everyone has built up confidence before touching on cross-system matters.
Q: Should we begin with a private deployment?
It depends on data sensitivity. One bulk-commodities trading group had annual revenue of RMB 80 billion and could not allow business data to leave the company, so a private deployment was appropriate. But most small and medium-sized enterprises do not need it - it is more important to use it first than to set up the infrastructure first.
Q: How long will it take to see the results?
Our own experience is: a single-point use case can show measurable change within 2–4 weeks, and cross-functional process transformation often requires 2-3 months. If you still don’t see any indicator changes in half a year, you can basically judge that you have chosen the wrong scenario.
Summary
The usual order of enterprise AI adoption should be reversed: first find the specific problem that consumes people’s time every day; then discuss tools and training. Tools will be changed and models will be upgraded, but the method of "starting from business pain points and using indicators for acceptance" will not be outdated.
If you are not sure which use case to start with, you can start with the six-dimensional scoring table we compiled - first rate the three or five annoying things at hand, and the answer will usually emerge.