How do you measure your AI system's effectiveness?

In my experience, no single metric fully encapsulates the effectiveness of an AI system. When I worked on a project to predict customer churn, while accuracy seemed promising initially, the cost of false negatives was too high for the business. That's when I realised that sometimes business-specific metrics, or custom loss functions, offer a better lens through which to gauge performance. For instance, setting up a cost-benefit matrix that includes the financial implications of each prediction type can provide a more realistic measure of performance from a business perspective.

Defining AI applications' performance metrics can be complicated because of AI's probabilistic outcomes where algorithms rely on statistical models and are programmed for uncertainty. However, strategies exist and they mostly relate to improving success evaluation and making AI systems work better.

Effectively measuring an AI system's performance involves a multi-faceted approach. Start with well-defined metrics like accuracy, precision, or F1-score relevant to your application, whether classification or prediction. However, continue with these conventional metrics; consider data quality, model complexity, and fairness metrics to ensure alignment with business and ethical goals. You can use benchmarks and leaderboards to compare your system's performance against peers and state-of-the-art models. This can guide further improvements. Lastly, a user-centric approach is crucial. Collect feedback from end-users and stakeholders to gauge the system's usability and real-world impact, ensuring it meets diverse needs and expectations.

The ideal performance metrics for an AI system depend heavily on its intended use case and the potential costs of different types of errors. False positives and false negatives carry different risks and consequences across contexts. In medical diagnosis, for example, failing to detect a serious illness (a false negative) can be catastrophic. In cybersecurity, false alarms may interrupt valid users, so precision may be valued over recall. E-commerce recommendation systems often optimize for relevance, while autonomous vehicles require near-perfect object detection. Rather than using universal metrics, AI system designers should think critically about the specific needs and risk factors of the use case.

If we want to measure a model from a technical perspective, then accuracy, precision, recall, F1-score, ROC curve, AUC, MSE, MAE, RMSE, R-squared... are useful metrics. To get these metrics, we usually need a test set. Nevertheless, the way most people will interact with AI will be by using AI tools. They might not even know what a test set is, but their concerns will be related to tool cost-efficiency, implementation prerequisites, skills required, responding time, etc.

The voice of those who use or receive AI applications is equally very important. It helps to incorporate user's feedback and assessments when measuring AI applications' effectiveness. A more holistic view of effectiveness is achieved by balancing users' insights and objective performance metrics.

The problem we need to solve here is to be able to define metrix that help us without the users self-report to assess ease of use ... then we can still ask the users on top of that what they think. Ease of use design usually competes with emotional design and is a trade-off.

From a privacy perspective, I want evidence on the feedback loop and that an organisation is reducing harms such as bias. Whilst business metrics talk at at the org level, you also need user focused measures, that focus on user impact. For example, a method for users to provide regular feedback and regular evaluation of the AI system based on this feedback.

While quantitative metrics like revenue and customer satisfaction scores are crucial, they don't capture the whole picture. To truly understand the impact of your AI system, engage directly with the people using it. Surveys or interviews can reveal vital qualitative insights, such as missing features, that aren't captured through traditional metrics. This human-centric approach is especially crucial when your users perceive a gap in the system's capabilities, an aspect that would be impossible to measure without direct communication.

PRI, ROI, TCD and TCO are good measures to consider from the offset in order to justify the investment in any projects AI implementation. PRI - Productivity Return on Investment (compare to manual process) ROI - Return on Investment (consider people and profit) TCD - Total Cost of Digital (ask me about it) TCO - Total Cost of Ownership (ask me about it)

Ideally the impact metrics need to be in line with the incentives of the business. So the KPIs or lagging and leading indicators measured and their improvements need to be the shared. If those are not defined in the business (which happens a lot more often than we think) it is hard to show AI impact. This than need to be operationalized in sub-metrics that in turn need to be experimented on to see if they influence the final metric to the extend we had hoped.

In my experience, regardless of the evaluation method you choose, it's crucial to take into account the unique characteristics of your industry and the specific needs of your target audience. This helps ensure that you're focusing on the most relevant metrics. For instance, in the pharmaceutical industry where I work, it's not just about tracking standard CX metrics. Vital are those metrics that assess patient accessibility, safety, and the overall impact on patient well-being. In summary, I recommend selecting evaluation methods that align closely with your industry's standards and unique traits. Always keep your ultimate goal in mind, whether that's business improvement or better serving your customers and stakeholders.

One other way to measure the effectiveness of AI systems and models are process mining approach. As today's AI are pretty much narrow AI, process mining approach can effectively be used to ascertain how much productivity savings while using any AI systems, cost of the system vs users performance, faster responsiveness (e.g. while using co-pilot, how quickly you can generate business graphs, data analysis etc Excel). Another aspect is how is your user experience and engagement (AI time, similar to screen time). One critical aspects, most of the AI players are ignoring that parameter though, cost of compute vs carbon and energy cost.

In the food chain, it's essential for me, above all, to understand why AI is being used. In other words, what risks are we focusing on in that SPECIFIC food chain? Additionally, crucial for assessing the robustness of the AI/base system used is the number of samples (sample size) used to train the model, which, in scientific terms, is an algorithm.

The evaluation of AI systems has evolved significantly over time, primarily due to the proliferation of AI applications transitioning from proof of concept to production. As these applications become increasingly customer-facing, it’s no longer sufficient to rely solely on accuracy metrics as the gold standard for assessment. Instead, the scope of evaluation has broadened to encompass critical factors such as data security, privacy, governance, model explainability, and ultimately, customer success criteria. Attempting to distill this multifaceted evaluation into a single KPI is inadequate. Moreover, AI system enhancement isn’t just technical; it’s about organizational culture. Measuring and nurturing this cultural aspect is also crucial.

Measuring the effectiveness of an AI system is a complex task that requires careful consideration of various factors. It is important to use diverse datasets to ensure that the system is robust and can handle a wide range of inputs. Diverse datasets can help identify potential biases in the AI system and ensure that it is fair and accurate for all users.

So far this AI-generated article focuses on the technology and not the experience, minus a small shoutout for usability. AI is a tool to be wielded, not an end in of itself. If we're discussing effectiveness, not efficiency, we need metrics that tell us how well the tool performs at helping the user perform their task or reach their goal - CSAT tells you only part of the story. What is the task completion rate? For existing features, how much faster did someone complete a task with AI? - In testing, what is the rate of type one errors (false positives) and type two errors (false negatives)? How frequently does the system hallucinate? - Assess Product Market Fit. How disappointed would a user be if you took away the AI feature? UX matters.

Documentation and Transparency: Maintain comprehensive documentation of your AI system's design, data sources, algorithms, and evaluation processes. Transparency is crucial for accountability and auditability. Codelabs is a free, easy-to-use software for writing clean and neat documentations.

AI effectiveness is challenging to measure because of its non-deterministic or unpredictable nature. By non-deterministic, we mean that AI applications don't produce consistent, predictable outputs from the same inputs. This unpredictability has its roots in algorithms used in machine learning and deep learning, which rely on statistical models and complex neural networks.