How to Avoid Turning AI Projects into Automated Amateur Operations?

Did your AI project succeed? Or is it about to fail spectacularly?

You must have built a bunch of workflows with n8n or Dify, or put together a RAG for an enterprise knowledge base, right? Whether you actually built an agentic harness yourself, or forked an open-source repo to pretend you built it in-house, you’ve built an AI assistant that can help clients generate all kinds of documents at a crazy rate, right?

It looks like everything is there, so why are your clients still dissatisfied with your work?

The documents AI produces, client employees say, are all empty verbiage, meaningless. You assume it doesn’t understand the business, so you keep helping the client import materials, work overtime every day cleaning data, update preset prompts version after version, and stuff in several sets of skills from unknown sources, demanding that it combine actual business context, conduct deep analysis, think strategically, and give actionable recommendations.

The result is still no good.

The materials are complete, the workflow looks like a spider web, and the prompts are complex enough, but for things the client could already do themselves, AI may not even reach half their level. For things the client never knew how to do, letting AI do them still won’t produce truly implementable results.

In the end, you’ve merely built an automated ramshackle operation; AI implementation achieved nothing.


Reversing cause and effect is the root of the problem.

Why did implementation turn out like this? Actually, after careful review, it’s not hard to find that many projects didn’t start from specific business problems, but from “we also need to use AI.” First build features, then look for needs, and finally treat what can be done as what the client needs.

Take enterprise knowledge bases. Many medium and large enterprises do have needs. But in businesses that depend on a small number of outstanding employees, what the enterprise truly wants to accumulate is high-value working methods. Why can only a few people do the same thing well? Can we retain successful experience so other employees can use it too, and avoid the business collapsing when an outstanding employee leaves?

When this need is translated into technology, it easily becomes importing data to make a Q&A bot, letting employees look up information in a different way. But even if they can find historical records, will employees know what to do? Most of the complete information of successful practices is not written into work documents. Even if you organize existing materials hundreds of times, a few outstanding employees won’t immediately grow out of them.

Therefore, many AI projects violate the common sense of doing business from the very beginning, reversing cause and effect.

If you really want to do AI implementation right, the first thing is to reverse the order. If an enterprise wants to reduce dependence on outstanding employees, we should first find a specific job that only outstanding employees can do well, cross-compare the shortcomings exposed by other employees when handling the same problem, and then decide what AI should help with.

Begin with the end in mind. Once the business problem is accurately identified, the next step is how to get AI to do this well.


Turning knowledge into tools is the solution.

Once the business problem is identified, the next step is to bring the relevant knowledge into actual work. This knowledge must both guide AI on how to do it and judge whether AI has done it correctly.

I discussed in “Why Is It Hard to Replicate a Sales Champion?” that collecting only the records left by outstanding employees easily replicates a “plastic champion.” You know what he did, but not why he did it.

In practice, even if outstanding employees are willing to cooperate honestly, they may not be able to fully explain their methods. At this point, we can first ask them to define the target outcome. Even if it is still vague at first, it can give AI direction when trying intermediate processes.

The analysis, judgment, and actions of outstanding employees can help AI find the right way to do things. Their evaluation of the results can point out where the problems are, allowing AI to continuously correct its path. Gradually organize this feedback into acceptance criteria, then keep revising them as new situations arise, and knowledge can become a usable tool through iteration after iteration.

This kind of collaboration is especially important when we face unfamiliar businesses. Even for a business that outstanding employees are not familiar with, their existing abilities and experience may still help define goals and identify problems in the results. As I said in “Using AI to Work, You Don’t Need to Become an Expert First,” I can’t write a voiceover script, but I can point out what’s wrong and then have AI adjust based on feedback. We may not be able to articulate the complete method first, but we can start from evaluating results, gradually establish standards, and let AI explore the intermediate process.


How do we clear the obstacles to implementation?

Whether outstanding employees are willing to participate, whether anyone adopts AI’s results, and whether the authority that should be delegated can be delegated all require organizational cooperation. A major obstacle in AI project implementation is that the boss agrees to do the project, but the departments do not actually support it.

For management, they will consider how their responsibilities, resources, and say in decision-making will change after the project succeeds. Outstanding employees will also wonder: if they contribute their experience and ways of working, will it make them more valuable in the company, or dig their own graves and let AI replace them?

So in the short term, we need to make management an enabler, working together around shared business goals, using AI to enhance their ability to solve problems, and making the project’s results part of their work results. Only then will they have the motivation to become a force for AI implementation rather than resistance.

At the same time, outstanding employees begin to take on the work of checking AI results and optimizing acceptance criteria, so the company must adjust their original workload and include these contributions in performance and compensation.

And in the long run, the company’s criteria for selecting managers must also change. In the past, a person who was good at assigning tasks, tracking progress, or office politics could become middle management. In the future, managers will also need to accept the results output by AI.

Bureaucratic managers who are detached from the front line and only know how to manage by spreadsheets and processes will find it hard to be competent for this job. If they lack the ability to judge good from bad, they may force AI and employees to cater to the wrong metrics together. This kind of management will spread the problem to the entire team.

Therefore, companies need to select people from the front line who have actually produced results and can discover and correct problems, involve them in management, and continuously maintain AI’s work and optimize business results.


AI implementation is a systematic project from the inside out.

Starting from specific business problems, turning knowledge into tools that can be used in actual work, then designing business acceptance methods, talent selection mechanisms, and re-orchestrating business processes—these things are all interconnected.

AI has changed the ways and methods of work, so organizations must reconsider collaboration between people, collaboration between people and AI, and redefine the responsibilities managers bear, as well as the company’s evaluation metrics and error-correction mechanisms.

Therefore, AI implementation is not only the implementation of technology and methods, but also the reconstruction of organization and management. Knowledge bases, AI customer service, and automated workflows are only part of it. Completing these features is still far from completing the transformation.

What AI implementation truly needs to deliver is a work system that can operate with the help of AI and continuously improve business results. Otherwise, no matter how complete the features are, it is just installing an automated engine on the original ramshackle outfit.

The faster you run, the worse you die.

Let's talk

Tell me what you think