11 Secrets of Successful AI Projects

Artificial Intelligence (AI) is no longer science fiction. It is rapidly permeating all industries and has a profound impact on virtually every aspect of our existence. Whether you are an executive, leader or industry professional, understanding AI, its impact and the transformative potential for your organisation is of paramount importance. 

We currently live in a highly disruptive ‘new normal’. Although forward-thinking businesses are fast tracking AI into their digital strategy right now, the vast majority are in the ‘wait and see how this pandemic plays out’ category. Whichever camp you choose, the following checklist will help you assess the options and even select an AI provider. It will serve to reduce reputational risk and improve your chances of successful AI projects that deliver real value to your organisation. 

AI comes with its own risks, so we need to understand how to mitigate these before inviting it into our businesses. As digital technology becomes an ever more central part of every aspect of people’s lives, people should be able to trust it. Trustworthiness is also a prerequisite for its uptake. To help with this, the EU Commission for AI are committed to building an ‘Ecosystem of Excellence’. Their strategy involves attracting over €20 billion of investment in the EU per year in AI over the next decade. In short, AI is being backed on a global scale. Understanding its basic concepts is set to become a necessity. 

The European approach to AI will need to be underpinned by a strong focus on skills to fill competence shortages. Evidence shows that the skills gap is widening due to emerging technologies such as AI, and mass furloughing and unemployment will only accentuate this issue. Developing the skills necessary to work in AI and upskilling the workforce to become fit for the AI-led transformation will be a priority for many businesses and the wider economy. In my next article, I will explain the background and effects of the skills gap created by emerging technologies such as AI, and the specific in-demand skills for 2020/21 that follow. For now, we focus on the steps needed to select and deploy valuable AI. By following this checklist, you will develop your understanding of AI further and improve your capability to lead AI transformation programs.

Bear in mind you may not need to tick every box. Some of them may not be applicable to your program but you should score highly on at least 6 of the following.

1. Objective setting

Have a clear set of goals, measurable outcomes, requirements, assumptions and risk assessment, so these can be referred to as you go through the process. This is will ensure briefing and communication with suppliers is clear and aligned.

2. Your Data

Good AI needs good data, and large volumes of it. The application you are about to select may require a certain richness of content you may or may not possess. To find out if your data is sufficient, ask your vendor to review it. If it does not meet the minimum requirements, you may need to data cleanse or enrich your data. This can delay the go-live phase and increase the cost, as additional services will likely be required. 

3. Training Data

What data sets did the vendor use when testing and training their AI systems prior to your project? They will also need your data for training it further. Do they have accurate records regarding the data set used, including a description of the main characteristics and how the data set was selected? How did they address potential bias in the training data to date? The models will only be as good as the data used to train the system, so you should discuss this with any potential vendor.

4. Document everything

Keep record of everything noted in point 1 and all the steps taken in implementing the technology, including any key decisions made along the way. This will need ongoing updates during operational phases to check how learning is proceeding. You may need this documentation in the future. 

5. Bias

To avoid unintended consequences, it is essential to probe for any potential bias in the system. While AI-based products can act autonomously by perceiving their environment and without following a pre-determined set of instructions, their behaviour is largely defined and constrained by its developers. A good place to start here is by finding out how diverse the developer team is and the sources of training data. Furthermore, if the AI vendor will not share the detail requested in point 2, they may have used incomplete data for modelling which can result in bias. 

6. Access to the Vendor or System Integrator’s key players

The key players include, in the order of importance:

i. Project manager 

ii. Implementation team 

iii. Reference client(s)

iv. AI developers

v. Product owner 

A salesperson’s currency is stories, whereas, a vendor/SI project manager (PM) deals in reality. You might be surprised to hear that these two individuals rarely interact. As a result, the salesperson’s story can veer from reality. If left unchecked, sales can set unrealistic expectations about the tools they are proposing. To avoid any surprises, the buyer should have at least a one-hour showcase call with the assigned PM on the topics of ‘AI readiness’ and ‘lessons learned from previous AI projects’. If you want a successful AI project you need a PM with AI ‘scars’ on their back, meaning that you need someone who has travelled this path before and learned from the experience. To validate the supplier’s story, you should be connected with as many reference clients as possible. If the vendor refuses access to any of these key individuals (especially i, ii & iii, namely the PM, implementation team and reference client) they should explain why. Do not accept an answer such as “we are unable to grant access at this time. Our company process is to allocate resource after contracting”. This is a red flag. Good vendors should have no problem commencing the transition from sales to delivery once they are selected as finalists, or at preferred stage of the selection process. 

7. Transparency

It is important to inform the stakeholders that AI is being used for a decision or process, as well as potentially providing information on key variables or parameters. Provide meaningful and clear information about how the model makes decisions, and why that result matters. This may be difficult, depending on the opacity of the AI application, but it is becoming increasingly important and a prerequisite for the effective management of point 8.

8. Dedicated resource 

Assign trained ‘parents’ to monitor your AI app to prevent any unintended drift. You are about to adopt a ‘toddler’. With the right level of support and guidance they can grow into a responsible, well-functioning ‘adult’. Likewise, if you neglect your AI they will continue to make poor decisions and never really develop. AI needs dedicated ‘parents’ to ensure it completes its learning process. Much like humans, AI benefits from continuous evaluation instead of annual reviews. 

9. Governance

Having policies and procedures concerning when AI can be used, how the appropriate inputs to AI models should be determined, and who must be involved in final approvals of any applications.

10.Impact Analysis

It is necessary to determine when to conduct AI impact assessments before specific applications of AI decision-making are put into production—including, where appropriate, considering ways to minimise potentially negative impacts. It is also good practice to regularly review the impact your AI app is having versus the measurable objectives set out in point 1 and adjust as needed if results vary from expectations.


Make sure you have authorisation to use the proposed input data and a process to ensure that the data use is consistent with the authorisation received. This should include due diligence and sampling, to ensure that actual input data is consistent with what people think is (and what is not) going into the model. Where appropriate, provide privacy statements about how data is kept responsibly anonymous and non-identifiable. 

In conclusion:

By following this checklist, you will lay the foundations for sustainable, ethical, human centric AI programs. You will hopefully avoid negative media attention while significantly improving your product and the services offered. From experience, with any emerging technology program, the best approach is always to think big, start small and move fast. 

Visualising the big picture will help you and your team get through many of the frustrations associated with taking an unfamiliar path. By starting small you will limit the impact of errors. This considerably reduces the commercial cost, but also allows the program to fully develop by using less resource and time, not to mention the reputational risk, which is minimised if all the right steps are followed. Once you have a stable pilot program up and running, you can scale up quickly. Do it and maximise the important efficiency and productivity gains that your selected AI application can offer. 

Kevin Butler

Founder of Centigy