15 Tips for Implementing a Successful Artificial Intelligence Project

According to many estimates, most AI projects fail (e.g.).? You can buck this trend and increase your odds of success by following a few simple rules.? Projects fail for technical reasons, for data reasons and for “people” reasons. Being prepared for all three will make your project run better and increase its likelihood of success.

1 Take care to understand the specific problem to be solved.

What exactly is the problem that you need to solve with AI? Check your assumptions.? Don’t make the problem harder than it needs to be.? What are the exact steps and processes needed to solve the problem?? For example, if the problem requires a distinction between one class of items and ten others, there may be no need to identify all ten classes, just one versus the rest.

Look for ways to simplify the problem and thereby make solving it easier.? Address the actual problem, not someone’s casual interpretation of what the problem is. If the problem statement is vague, try to get some clarity, but it may be necessary to prepare one or more demos to elicit what is actually needed.? Users are often experts in their work without being experts in the processes that enable them to do their work.

2 Examine how people dealt with the problem before AI.

Most business problems were being addressed before AI became a popular topic.? There may be, for example, manual processes where users copy data from one file to another.? Users may fill out multiple forms and then have difficulty reconciling the inevitable differences that crop up.? There may be published methods or even previous AI systems that were applied to the problems.? They may use templates to structure documents.? All of these may be useful sources.

3 Limit the scope of problems for the first version.nbsp; Better to do something useful, than to fail to do it all in the first attempt.

Solving the entire problem may tax the patience of the stakeholders.? Determine whether solving part of the problem may be useful.? Rolling out solutions in steps helps to hold the interest of stakeholders.

4 Prioritize simple problems first.

Organize the priorities of producing partial solutions by how easy they are to produce.? Often a small amount of effort can be useful to produce a large improvement in a problematic process.? Users will not judge the value of a partial solution by how technically challenging it was, but by how it affects their workload.? Stakeholders are usually more patient once they have seen some progress than if they have to wait to see any progress.

5 Recognize that coding is only a small portion of implementing an AI application.

As is commonly said, given a choice between better data and better algorithms, data usually wins.? There are many AI tools available today that can be used with only a few lines of code, but getting the data, vetting it, and getting it into the right form is where the real work typically is.

6 There is no easy button.

Even relatively simple AI problems tend to take more effort that stakeholders may assume.? There is no such thing as “just do …” Communicating effectively about the level of effort needed and timeline to achieve it are essential to maintaining stakeholder commitment.

Fitting the AI solution into the whole of an organization’s operations can also be a challenge. Do not under-estimate the amount of time needed for these auxiliary functions.

7 Show your work.

A simple demo can communicate your solution better than all of the documentation that you could write.? Users have their own ideas of how a process may operate, but most of what guides their expectation is past solutions and consumer applications. A demo of how a solution would work and specifically how it would be used can be a powerful tool to foster stakeholder interest and to solicit useful feedback.? A common comment on a demo of a new product is “that would be great, but can you make it do X?”? Many developers like to use wire-frames or drawings, but I don’t find these as effective as a demo of how it will affect the user.?

8 Be wary of “tools” that demo well but fail when confronted with real-life problems.

Demos are great at soliciting user feedback, but they can also be misleading. When your project must include outside tools, and most of them do, be at least a bit skeptical of the demos you see.? They often involve carefully crafted versions of specific situations known to succeed and they may conceal known or unknown weaknesses when applied to your problems.? Ask questions and make the best independent evaluation of the considered product as possible.

9 Pick the solution for the problem, not the other way around.

Customers are currently clamoring for products that include the latest AI technology.? Management may fear missing out if they do not deliver what the customers ask for.? From the point of view of a successful product, however, focus on the tools needed to solve the problem.? Not everything has to be done with GenAI.? Use it for the parts where it provides value, but do not limit your solution to what it can provide.? Hybrid solutions, including GenAI and other tools often work best. The goal of a successful product is to achieve the capabilities of the product, not to prove that one technology or another is sufficient for every need.

10 Be aware of the costs of each component for development and once deployed.

A product that costs more to develop and run than will be recoverable by revenue is a significant source of AI projects’ failure.? The current crop of GenAI models, for example, are very expensive to operate and may not support a business model.? Nevertheless, it may be better to develop a product that is effective, even if, in its current form, it is too expensive to operate.? Once the basic processes have been demoed it may be possible to find lower-cost means to implement them.? The latest GenAI models, for example, may require orders of magnitude more computing capacity but provide relatively small improvements relative to earlier, often cheaper models.

11 Plan for security and privacy.

A major reason that organizations shy away from using AI is their fear that they will lose control over their data.? The public interfaces of the GenAI models have been known to record the data submitted to them and to use those data for further training.? Commercial versions of the models protect user data according to the same standards that govern other cloud applications and so are much more secure than the publicly available tools.

Even within a product, exercise care to avoid data leakage from one user to another.? Access controls must be respected.? Fortunately, these controls can be designed into the earliest versions of an application and then carefully maintained through development.

12 Plan for failure and recovery.

Innovation involves experimentation, and some experiments fail.? Development may hit blockages.? The best, easiest to use, data may not be available.? Users may procrastinate providing feedback.?? Prototypes may not work as anticipated.? In short, stuff happens.? Be prepared to deal with it.

13 Identify the likely information needed to solve the problem and identify where it can be obtained.

Data are the lifeblood of any AI project.? The cleverest computational methods cannot succeed without necessary data.? Early in the project, determine as well as possible what data will be needed and make plans for obtaining them.? Be prepared to revise your plan as you learn more about what is available and as you see their value (or not) in addressing the problem you are solving.

14 Evaluate the sources of information for reliability, validity, and representativeness and be prepared to seek other sources and features.

Effective AI projects require valid, reliable, and representative data.? Even if you think that you have the data needed to solve a problem, it is possible that those data are not valid or not reliable.? Valid data measure what you need to measure.? Locker numbers are not valid estimates of student IQ, for example. Reliable data are consistent; if you ask the same question multiple times you will get the same answer from your data.? Unreliable data may, for example, be stored in different places, each of which gives a different value.? Representative data are data that cover the full scope of your problem.

15 Plan for change management.

Technically successful AI projects change business operations.? They affect people.? Plan for dealing with the expected changes.? Some people may feel threatened by the changes, others may welcome them, most will need support to take the best advantage of them.? Even technically successful projects fail if the people who are affected by them do not adapt and accept them. A “great” solution is worthless if no one uses it.? Be prepared to train and support potential users.