AI is non-trivial, so you can’t expect to just buy some software, build your model and get results out of the box. Superimposing AI over a dysfunctional system won’t work either. 

There are 4 phases to delivering an AI solution:

1.    Plan 

1.1.  Strategy

Having gone through the strategic goal setting process above, the organisation should have a quantified and time-bound target to pursue 

1.2.  Business Case

With the strategy clear, a case needs to be made to obtain resources. The business case needs to include ROI, but also what will be learned and how this learning will be applied to add value.

1.3.  Regulation

Some AI related activities are more regulated than others, depending on the industry and the nature of the application of AI In Europe, GDPR covers the use of personal data across all industries and scenarios.

1.4.  Change Management

AI is by definition a disruptive technology, so its implementation will require change management, including changes to organisation structures, job roles and extensive training.

1.5.  Program Governance

The scope of governing AI covers measurement & evaluation, ethics, control of autonomous systems, openness, privacy and security.

2.    Design

2.1.  Problem Definition

Arguably, the most important factor in any form of technical activity is to define the target problem correctly

2.2.  System Design

Designing AI systems covers broad architectural paradigms, infrastructure design, solution building blocks, data management, use of external cloud-based services, development methods, and machine learning environments. System designers need to understand what a proposed AI solution needs from other systems. They also need to understand what triggers code execution, interdependencies with other processes, and how AI-based predictive outputs will cause actions to be taken.

2.3.  Vendor Selection and Management

An organisation’s data security and compliance regulations could influence the choice of vendor or force an on-premises solution. Cloud-based solutions are usually more cost-effective in the short term but can add up over time, and add risks of ‘vendor lock-in’. A cloud-based vendor should be asked about encryption, security and data ownership. Whilst getting your data into a cloud-based solution may be easy, you should also consider how to get it out again.  

2.4.  Data Management

Leaders ensure that the organisation’s data and sources are good before embarking on an AI project.

2.5.  Control Framework

Controls should be built-in at the design phase. For physical systems, safety engineering must be part of the design process. In pure software systems, ‘red-line controls’ can be implemented - such as maximum transaction value or ‘behaviour inhibitors’ that overrides the algorithm when there is a risk of errors such as regulatory violation.

2.6.  Verification and Tuning

It’s essential to perform tests to verify that the right model has been used on the right data, and then use the results of these tests to tune the machine learning system.

3.    Implement

3.1.  Organisational Readiness

For your organisation to be ready, you need the following in place: strategic plan; problem identified; interdisciplinary team; data and data ecosystem; scale-out plan including leadership, organisation, infrastructure and training plans.

3.2.  Phase 1 Implementation

Options for deploying a Phase 1 implementation include: manual – take data files and run models on them using Python, or R; Virtual Machines with ‘sandboxed’ AI environment; batch/distributed; API; 3^rd party Cloud Services.

4.    Operate & Improve

4.1.  Scale

After Phase 1 has shown benefits, all that stands between the concept and profits is scale. Here, decisions have to be made about the use of nodes and whether to scale horizontally or vertically. Decisions need to be made about how to scale databases. If the organisation is going to run Neural Networks at scale, then decisions about processor types – GPU, IPU or CPU need to be taken.

4.2.  Security

According to Gartner, traditional prevent and detect approaches are inadequate and a shift to a “continuous response” stance is needed. Continuous monitoring of systems and behaviour is the only way to reliably deal with threats[i].

4.3.  Data Quality

A solid data quality foundation is critical for a successful AI project. Data has to be consistent, accurate, and complete. A data quality strategy and plan is a critical part of a successful implementation.[ii]

4.4.  Compliance

AI’s dependence on data means that applicable data privacy laws will need to be complied with. GDPR also demands that algorithms and decision-making processes are explainable. Other compliance considerations include potential future law enforcement access to data; IP protection; and broader ethical issues.

4.5.  Deliver and Refine

Whilst AI is extremely powerful and pervasive, it’s worth remembering that the core process is about delivering predictions. Whilst predictive capabilities can get ever closer to perfect with the right data, models and environments, we generally have to work with imperfect data and models. So, the end-point of an AI project is to continuously improve the data and the models.

[i] https://www.gartner.com/smarterwithgartner/build-adaptive-security-architecture-into-your-organization/

[ii] https://towardsdatascience.com/data-quality-in-the-era-of-a-i-d8e398a91bef