How to Implement AI in Your Software Solution
Artificial intelligence (AI) is rapidly moving out of the laboratory and into business and consumer applications. As a result, we’re seeing a fundamental shift in how software is built and what it’s capable of doing.
Of course, the type of AI we’re talking about isn’t the artificial general intelligence of science fiction, where a robot or software program can do whatever a person can and more. Instead, we’re referring to artificial narrow intelligence. In other words, the use of AI in very specific contexts. It’s this type of AI that’s already being used to power some of today’s most successful technology businesses, including Google, Facebook, Amazon, Netflix, Uber and Airbnb.
AI is a rich and complex topic, and one that has evolved as a result of years of computer science research. Today, we have reached the point where AI can be applied within most businesses, provided that they have the right talent, and that they’re focused on applying it to the class of problems that AI is best suited to.
In recent years we’ve witnessed major progress in the field of AI. This is primarily thanks to advances in machine learning, one of its best-known and most widely talked about subfields. With machine learning, software algorithms learn new things from data without having to be specifically programmed to do so. In other words, machine learning can learn from and make predictions on data. The result is software that can tackle tasks that would otherwise be too complex to program manually.
This paper isn’t intended to help you understand the latest advancements in the field of AI. Rather, our goal is to help you recognize the problems to which artificial intelligence can be applied. We also want to help you develop a strategy for delivering capability and customer value that would otherwise be impractical, if not impossible, without AI. At its core, this is what we mean by the term applied artificial intelligence (applied AI).
How to Use the Principles
These principles will help you to get a grasp of the highly complex topic of AI, and can be a useful reference point when it comes time to develop and execute your AI strategy.
You can navigate using the menu bar on the left of your screen or by clicking the NEXT button at the end of each section.
The principles themselves fit into five related groupings as shown in the framework below. The purpose of grouping them in this way is to separate the principles into areas that can be discussed both individually and as part of the larger framework.
The Maturity Framework
1. Capability and Ingredients
Your organization has intrinsic company characteristics, assets or strategic intent that give you an advantage for executing on applied AI. You have a basic understanding of priorities and opportunities, and there are few or no specialist skills within your organization.
2. Readiness and Process Advantage
You have formalized company standards, practices, procedures and skills that give you an advantage for execution. You have partially documented your approach and priorities, and have some specialized machine learning skills within your organization.
3. Execution Advantage
You’re exploiting advantages with mature processes that capture outcomes. You’re also using data and insights to drive continuous improvement. Last, but not least, you have complete documentation of your approach, including roadmaps for implementation, and a complete set of appropriate skills to support the implementation of applied AI within your organization.
Start with the processes
Principle 1
Understand All Processes
Catalog the processes your software solution enables, any adjacent customer processes and relevant third-party processes.
Level ONE
Level TWO
LEVEL THREE
Opportunities identified in an
ad-hoc manner, primarily within your company’s solution.
Key business processes documented and prioritized.
All business processes prioritized across wider ecosystem; roadmap developed.
Applied AI represents an opportunity to intelligently automate or augment a wide range of business processes, including those your business doesn’t control today. To best direct your efforts, focus on understanding the ecosystem of business processes in which you operate. To do so, catalog all of the relevant ones.
Start by documenting the business processes that your existing software solution enables (e.g., the management of payments to suppliers). Next, consider areas of your customers’ business where they have processes adjacent to your area of expertise that are delivered using manual effort or custom software solutions (e.g., manual processing of an incoming invoice in an otherwise automated payments workflow). Finally, take note of adjacent business processes delivered by third parties, including partners and competitors (e.g., customer support software for managing inquiries from suppliers awaiting payment).
Don’t limit your thinking to improving existing processes. Instead, also consider how processes might be extended further if new types of predictions and automations were possible. Think, too, about opportunities to create new processes that aren’t currently possible with human effort and intelligence alone. For example, commercial loan approval has traditionally been a manual process with loan officers using simple models and limited data to make their decisions. New machine learning approaches now enable much richer decision-making that augments human judgment with models that can include dozens of variables while also incorporating historic data.
Document how predictions and decisions are currently made for each process. Are they made by people? Automated using business rules? Pay attention to potential information gaps within the ecosystem that limit current predictions or decision-making. Finally, identify and document any key performance indicators (KPIs) used to measure a particular business process.
Start with the processes
Principle 2
Prioritize the Most Valuable Processes
Identify and prioritize which business processes are best suited to be automated using applied AI.
Level ONE
Level TWO
LEVEL THREE
Prioritization of opportunities for automating processes started, but incomplete.
Some high-value opportunities for optimizing existing insights and automation have been prioritized and are part of the business solution.
Most opportunities to optimize existing automation and augmentation have been mapped out and some are in production. Some high-value opportunities for automation have been identified.
There are a variety of ways that artificial intelligence could be applied to your business. These range from simply improving the quality or efficiency of existing insights to augmenting human decision-making. Other common applications include fully automating business processes that previously could only be undertaken by people, as well as processes that weren’t possible before even with human insights and decision-making.
Not every opportunity is equally valuable. Some uses of artificial intelligence in your business will create outsized value compared to others. It’s therefore important to understand value, cost, risk and timelines to help identify and prioritize the best opportunities to focus on.
We recommend scoring each potential process using the following attributes. With each, consider the relative opportunity to create value, what the cost might be and what the risk of implementation is:Understanding. How well must the process be understood? Is the process well enough defined that a model could be trained to deliver human-like outcomes with sufficient quality?
- Automation. Does the opportunity represent a means to improve the quality of insights and decision-making in a process (augmentation),
or is it likely to enable full intelligent automation of the process?
Scale. Would automating this process remove constraints in the system, allowing more work to be done without an increase in human effort? - Quality. Is the required quality of the process outcomes understood? Are there specific legal or other governance requirements on the quality of outcomes? Would AI-driven automation meet, or even exceed, the quality of any current human or automated decisions
if they exist? - Data. Is there data being captured or could it be captured to allow the training of models with sufficiently high quality? Is there sufficient labeled data to support machine learning or a cost-effective means to create labels for the data?
- Control. Are you in control of a sufficient part of the process and do you have integrations into all decision endpoints so that you can have a high level of impact through automated actions?
The output of the exercise should be a list of candidate processes for automation and augmentation that have been prioritized (based on relative value, cost and risk) for further investigation.
Integration
Principle 3
Design Frictionless Integrations
Understand what a proposed solution needs from other systems, what triggers its execution, interdependencies with other processes, and how the predictions it makes will cause actions to be taken.
Level ONE
Level TWO
LEVEL THREE
Easiest integrations handled first. Some friction exists and simplification is possible. More integrations possible to increase AI utility.
Roadmap exists for major integrations. Care being taken to simplify use and reduce friction for a wide set of users. Implementation underway.
AI capabilities that deliver high levels of automation are easy to consume across the business. All processes have necessary integrations. Have found best way to present AI results to users. New integrations added systematically as needed.
As with all design processes, once the general objectives for the solution are understood, a good place to start is by considering the external aspects of the solution. That includes how it looks from the outside and how it interacts with the environment in which it operates. There will be plenty of time to deal with the details of data sources, machine learning models and quality. Initially, get as clear as possible on what the solution needs from other systems, what triggers its execution, and how the predictions it makes will cause actions to be taken. These are the integrations to be designed.
By frictionless, we mean that the objective for every integration is to be as automated, natural and complete as possible. Some questions that you may want to answer as part of the design to achieve the most frictionless interface include:
- Is the AI solution fully embedded within automated processes? If so, the AI predictions might be able to drive follow-on actions with no intervention or friction whatsoever.
- Does the optimal AI solution need to deliver a recommendation to a person? Then choose the most natural persona to simulate the expert, and the most natural interface the person would use.
- Is the AI meant to engage a person in a conversation? Think about the how, when and where of that conversation to ensure the most natural medium, language and timing to reduce friction. In some cases, the AI might provide feedback via visuals or other UX elements such as buttons.
- Are the steps that follow the AI prediction within your company’s control? If not, try to reduce AI friction through partnerships to extend the solution’s reach.
- Is the AI solution one that might be valuable to new users, systems or even third parties over time? Consider building interfaces through APIs to increase future flexibility and reduce friction.
At this stage, you can ignore the training process, the sources of data and the complexity of the models. They are all important, but the highest-leverage design decision for the AI solution is how it integrates into existing processes in a way that makes it most natural to use.
Integration
Principle 4
Integrate Human Judgment as Required
Look for opportunities to create new business value through trust.
Level ONE
Level TWO
LEVEL THREE
Little thought given to which processes need human judgment and how it will be used to continuously improve model predictions and actions.
Some thought given to which processes need humans in the loop. Some use of people to take actions in cases of predictions with low confidence. Some feedback incorporated into learning process and not all done in real time.
People integrated into select processes to add judgment to predictions to optimize outcomes and meet regulatory and user requirements. Higher levels of human judgment are iteratively replaced with more accurate predictions to increase levels of automation.
While in many cases the goal may be to complete process automation using applied AI, there are a number of reasons why that may not be possible. These include insufficient model accuracy, government regulations, and negative market or customer sentiment toward full automation. There may also be particular attributes of a business process that currently make full automation too costly or impossible.
Human involvement may include providing a level of oversight for automated predictions as well as providing feedback to improve performance. For example, when an AI solution is bootstrapped with little data, human decisions may accelerate model training to meet performance quality objectives. That human input can be captured either implicitly through product interaction data, including historic data provided in bulk to initially train the model, as well as through newly generated user interaction data. The system may also generate specific requests to multiple human trainers to classify a new piece of data and then incorporate that feedback into the model. Incorporation of this human interaction data is an important way of optimizing the system faster.
In-house experts such as financial analysts, compliance officers, nurses, attorneys and sales agents can also be an important source of judgment and training data. Having such domain expertise can be a competitive advantage. That’s particularly true when the impact of decisions is high, and the availability of experts in the market who can reliably make those decisions is low.
Additionally, people can serve as gatekeepers in situations where predictions may introduce disproportionate risk or when actions are irreversible. Regulation often plays a role in these domains and using AI to augment user decisions may be the only option.
For some processes, users may expect to be in control, thereby constraining the level of automation. Gaining the confidence of your users over time should allow you to increase levels of automation without human triggers.
Finally, there may be business processes in which the AI model provides a prediction, but where people must be involved to complete the action because the end-to-end process isn’t within your business’s control. For example, where the action resulting from an automated prediction in your business process is to be performed by a business partner.
Models
Principle 5
Understand Performance Objectives.
Look for opportunities to create new business value through trust.
Level ONE
Level TWO
LEVEL THREE
Incomplete and generic model performance objectives employed.
Model performance objectives and the business impact of model predictions are somewhat understood.
Rich user-based definition of model performance, including error tolerance, robustness and bias for every prediction. Expected performance levels are continuously validated with customers.
Given that achieving perfect performance is rarely possible, and even more difficult to prove, it’s essential to understand (and meet or exceed) the performance expectations for your market and application in terms of:
- Errors. False predictions originating from incorrect, incomplete or noisy training data,
or from programming errors in the model itself. - Robustness. Also known as model stability, robustness refers to how much the model is impacted by small changes in input.
- Bias. The unintended differential treatment of, or impact on, a specific group on the basis of common characteristics that are injected into an AI model through training data.
Performance expectations will most likely be industry- and even use case-specific. For example, in medical applications, there’s very little tolerance for errors such as false negatives that cause a condition to be missed. By contrast, marketing applications that recommend products to consumers are much more easily forgiven for mistakes.
It’s also important to consider the level of consistency with which the model must perform to meet expectations. Perfect performance is often not be required, but there will be some expectations of how well the model will perform with different inputs. Be aware, too, that some industries will have set rules, and penalties, to ensure that bias doesn’t enter into important decisions. For example, in financial services, loan decisions must not take certain factors such as race or gender into consideration.
Models
Principle 6
Start With Proven Modeling Techniques
Aim for the simplest model that can do the largest part of the job.
Level ONE
Level TWO
LEVEL THREE
Limited understanding of all possible machine learning techniques. Most simple techniques available are generally used. Acceptable fit to data and performance objectives.
Some process decomposition to achieve simplification as required as well as use of richer techniques. Experimentation used to optimize selection. Ongoing review of model performance and investigation of better alternatives. AI plan exists.
Expert understanding of the current ML landscape driving selection of the best possible techniques for the widest set of problems. Active in the research community, advancing the state of the art, and consistently testing results to optimize models. Rich AI plan.
The process of technique selection is not prescriptive itself. In the fast-changing world of artificial intelligence in general, and machine learning specifically, what couldn’t be done yesterday can be done today. And, what’s now possible may be faster and more accurate in the near future.
Generally, use proven methods. The resulting models are typically easier to implement, maintain and explain to customers. Using a proven method doesn’t necessarily sacrifice the richness or sophistication of the approach. Advanced machine learning methods are continuously being proven to work for different problems. Moreover, these methods are rapidly supported by commercial tools, open-source implementations, and educational resources that help reduce the complexity and cost of adoption.
The required degree of explainability of model predictions varies by problem and should be a consideration in technique selection. Users adopting a new AI solution often want to understand how the output of the model was generated. Regulation is also driving model explainability requirements. Models vary in the degree of explainability, and research is continuously improving the explainability of even the most advanced techniques. Understand your explainability requirements, and keep current on research work to ensure they can be met with selected techniques.
In some cases, business problems can be simplified by breaking them down into smaller issues. Different modeling techniques can be matched to each problem to get the best overall model and highest accuracy predictions. Multiple modeling techniques and resulting models can also be used for a single portion of the problem. This is referred to as ensemble learning.
To support optimal technique selection for all business solutions, hire experienced and curious machine learning scientists and engineers, and enable them to keep current on, and even advance, the state of the art. The machine learning scientist’s job also includes choosing and tuning relevant model parameters and testing trained models. Some of the machine learning scientist workflows may be automated over time with the maturing of AI techniques and frameworks.
Each trained model should be tested using carefully crafted test sets tailored for different business objectives. These will be used to ensure the documented performance objectives are met and to compare performance across different models for the same problem, or to support model tuning to improve the performance of a selected model. Different business problems require different levels of performance. Moreover, testing has a high cost and model tuning provides diminishing returns. As a result, only do sufficient testing to satisfy the objectives. Managing model quality must continue throughout the use of the models, since model results may change over time.
Document how each business process will be decomposed for modeling, the modeling techniques that will be used for each portion, and the data and skills required for training and testing the models. The plan should include future considerations of model and data improvement, the opportunity to increase automation and get to a best AI solution, and some guidance as to the improvement cycle timeline. Consider the cadence at which models will be updated.
Your leadership team should review, understand and champion those plans and ask questions about training, testing, meeting performance objectives, interpretability and the degree of automation. While AI models are complicated and constantly changing, the approach to their implementation and use should make business sense to all members of your company’s leadership team.
Data
Principle 7
Capture Relevant Data to Support Models
Develop any necessary plans and
infrastructure to capture relevant data to
support your machine learning models.
Level ONE
Level TWO
LEVEL THREE
Historical data captured for most processes, including most inputs, actions and outcomes, and some capture of context. Little to no understanding of how relevant the data is. Little to no infrastructure in place to match frequency.
Most relevant data inputs, context, decisions, actions and outcomes captured for some processes at the right frequency.
Full capture of relevant data inputs, context, decisions, actions and outcomes for all processes at the required frequency to match techniques and performance objectives. Appropriate use of data generation and simulation to amplify model training.
Depending on the type of problem you’re incorporating AI into, and the machine learning techniques used, you’ll need appropriate data to train and optimize the machine learning models. The quality of data is as important as your choice of machine learning techniques. You may already collect and have access to the data needed, although that data may require additional labeling. You might also need to acquire data from third parties. Given the performance objectives, the data characteristics will also inform the techniques you can apply.
Develop plans and infrastructure, as necessary, to capture relevant data to support your machine learning models. You may need to increase the frequency with which you collect data and instrument processes with new sensors to capture input, context, decision, action and outcome data end-to-end and in real time. Design the data-capturing pipeline to minimize data loss and, when using real-time data, use mechanisms to monitor data quality in real time.
The first step in solving any problem using ML techniques is to transform raw data to best represent the real-world entity. Identifying the best set of features for each task to maximize the utility of the model is called feature engineering and can have a large impact on model performance. One of the key advantages of deep neural networks for different applications is that these models can be trained through an end-to-end process to learn both input representation and parameters. Thus, task and domain-specific feature engineering may not always be required.
Finally, you can import subject matter expertise or leverage pre-trained models as starting points, or even derive data using data generation and simulation techniques, if necessary. These approaches can complement real data and can help address cold start issues, particularly in cases where capturing and leveraging sufficient amounts of in-process real data is a slow process.
In some cases, human judgment will be added to a AI process to monitor and tune the results of models. People’s reactions to AI-based predictions can also be biased based on their assumptions of higher or lower accuracy than reality, depending on the person. Further, human judgment may have the same underlying business process bias as the organization itself, making it difficult to detect a model error.
To minimize the risk of introducing bias, you first need to broadly consider potential biases in your organization’s processes, training data and human participation in training. Increase the diversity of data and participants, and use redundancy to reduce the potential impact of bias. Second, build assessments into your AI processes to detect incomplete or erroneous input data that might lead to bias and error. Finally, review model outcomes over time to ensure they continue to meet your business practice standards as model behavior can shift with new data.
Governance
Principle 8
Manage Quality for Error and Bias.
Effective execution to achieve and maintain the highest possible model utility while preventing the introduction of bias. Monitoring for errors throughout automated machine learning systems.
Level ONE
Level TWO
LEVEL THREE
Some awareness of the sources of bias and error, but no comprehensive approach to prevent and detect them.
Processes implemented to prevent the introduction of bias and errors and understand the remedies that can be used when they’re detected.
You create new value through leveraging your trust with customers.
Organizations must proactively recognize the potential to introduce bias into an AI system and design processes to avoid it. They must then automate the monitoring of system behavior at each stage of the AI process, and correct error and bias on detection. Most organizations will understand the impact of errors on their business, but many don’t yet consider the impact of bias.
While machine learning is often portrayed as an objective, fair and data-driven approach to decision-making, that can only be true if models are free of bias. Bias does not include the implementation of intended business decisions that, for example, may limit a company’s business activity to a certain geography, or certain market segment. However, there is risk when a business has experience in one portion of a market, and then uses its customer and transaction data in the context of a broader market. This exposes a blind spot, and can introduce bias into an AI model.
Bias and errors are related but different. However, the processes to avoid, detect and correct for either are similar. Often the problem in a model outcome will need to be investigated before it’s clear if the source of the problem was an error or an underlying bias originating from the input data. But, as with software development in general, errors are more difficult and expensive to find and repair the later they are found. Frequent review and early defect removal are possible and recommended in the development of AI solutions, and should result in better, more consistent and easier-to-interpret results.
Many types of input can be used to train machine learning models, including transactional data, which represents the real-time capture of business process execution. If the execution of the processes themselves have bias (hiring, promoting, lending, etc.), the model will likely be trained to reflect the bias, if no remedial steps are taken. In addition, some data may be labeled by employees, customers or a third party. If the rules they use to create the labels have bias, or the samples they are labeling are not fully representative of what the models may see at run-time, then the models will again reflect the bias.
Governance
Principle 9
Build Fault Tolerance as Part of Model Integration
Recognize that models will have errors, behave unexpectedly and sometimes fail.
Try to increase system reliability through fault tolerance.
Level ONE
Level TWO
LEVEL THREE
No specific fault-tolerance techniques implemented.
Some fault tolerance is implemented to prevent unwanted impacts from the most critical errors.
Fault tolerance is built into the system from the start. Uses a rich set of techniques to allow earliest possible detection and minimize the impact of errors in automated processes.
All systems built by people have errors that can result in unexpected behavior when they’re put into production. Non-AI systems have the benefit of providing generally consistent execution based on system inputs. That is, testing a non-AI system with a fixed set of inputs can largely eliminate errors prior to promotion to a production environment. A notable exception are errors that originate through timing — either sequences of events, time-of-day, or multi-process race conditions. These are a difficult kind of error to test for and eliminate, and most closely resemble those that can affect a learning AI system.
For AI-based systems, the behavior of the system can change over time as the model learns from new data, including human interventions. Even if you expand your training and testing data to include disaster scenarios, eliminating all problems in advance of the system’s use in a production environment is unlikely. In addition, increased automation of an AI system may allow an error to be introduced and go undetected for some time, affecting the quality of the actions, and reinforcing and potentially increasing the degree of error over time. To prevent this, all traditional fault detection techniques for downstream processes must be in place when actions are initiated by an AI system.
Fault tolerance must be designed and built into AI-based systems. Techniques such as rules-based circuit breakers, and human and AI-based monitoring and audits, check-and-balance redundancy, mandatory human review, recognizing data shifts, kill switches and least-harm actions are all viable and must be used selectively to provide results that sufficiently reduce the consequences of error.
Governance
Principle 10
Defend Your Legal and Ethical Stance
Consider external factors such as legislation and customer sentiment, and be open and transparent.
Level ONE
Level TWO
LEVEL THREE
Meeting legal requirements on information rights and privacy. Little understanding of the implications of automation/AI around social bias, repeatability, interpretability and future of work.
Adopting a market leadership position on information rights and privacy. Some understanding of market expectations on AI and social bias, repeatability, interpretability and future of work.
Adopting a market leadership position on all ethical and legal issues. Arguments for business process automation are supported with facts.
AI and the data associated with it will continually raise ethical and legal issues. The ethical and legal investments you make will be relative to the domain space and market expectations, the problem, and the data. Develop a strong understanding of the overall benefits of AI in your problem domain. Benefits can include achieving massive efficiency gains and freeing up people to work on more engaging, high-value activities, reducing risks by bringing consistency into a process, or helping people solve the otherwise unsolvable. Frame ethical and legal concerns in the context of the overall benefits and support your arguments with facts.
First, you will have to accommodate external factors, such as legislation, that introduce restrictions on the information you gather and store. An example of such legislation is the General Data Protection Regulation (GDPR) in the European Union. Your customers may also restrict you from using data via information rights agreements. Have a decision framework for whether you need some specific data points and how the respective data is being used to support your AI strategy and create value.
In addition, have policies in place to trim data that is no longer useful both on the time and attribute dimensions. Finally, incorporate privacy-preserving techniques that provide guarantees, such as differential privacy, where possible. This should enable you to proactively address legislation requirements, be transparent with your customers, give them insights into what data you’re recording and how they benefit from it, and gain their trust in the process.
Second, regulation can introduce restrictions on your AI-driven decisions and actions, to ensure fairness, but also reduce the risk introduced by automation in certain domains. Consider the consequences of an individual error when automating decision-making and have an explanation for why your AI-driven solution is better than the state of the art. If you start with augmenting experts because of regulatory restrictions or user requirements, how you explain your solution performance relative to the best alternative can allow more automation in the process, by influencing your customers and even regulation in the long run.
Third, acknowledge that increased automation will shift workloads and impact jobs, as most new technology waves prior to AI have. In this case, however, the effects may be more pronounced as AI solutions directly replace some human activities, as AI solutions evolve from augmentation to automation. Understand the impact and the issues surrounding these transitions.
Finally, it’s important that you show transparency and openness and adopt a market leadership position about ethical and legal issues. With the use of AI, there is a responsibility to understand it and explain it to customers and educate the market. Talk at conferences and publish articles on privacy, fairness, error, interpretability of AI decisions and the overall changes automation brings to your market.
Conclusion
Adopting the principles of applied AI will be an evolutionary process for any organization. And while the starting point for every business will vary, all can work toward higher levels of adoption of each of the 10 principles of applied AI over time. The best companies will hire smart, curious, capable machine learning scientists and engineers, and allow them to keep current on emerging machine learning capabilities. Those machine learning scientists and engineers should have business skills, or be partnered with people who do, so that applied AI projects take customer needs and sensitivities and business objectives, among other considerations, into account.
All the principles presented in this paper are important for the creation of the best AI solutions. However, don’t consider them as a playbook for implementation. Your own implementation will likely start with understanding the processes, but you will iterate in your thinking about integration, models and data. Managing quality and bias, and dealing with governance issues, is something to be mindful of throughout the entire process.
AI is an exciting technology and one that promises to fundamentally disrupt the way that you do business. By applying AI to your organization using the principles outlined in this paper, you’ll be able to derive the greatest competitive advantage, better serve your customers, and stand out as true leaders in your field.