The AI Playbook: How to develop an AI strategy
Before you invest time and money in AI, you need an AI strategy. We offer a blueprint for success.
👉️ Download the AI Playbook, your guide to developing and deploying AI.
AI is a powerful tool. But before you invest time and money in the technology, you need a strategy to guide its use. Without an AI strategy, AI will become an additional cost that fails to deliver a return on investment. Below, from our AI Playbook we describe how to: identify appropriate use cases for AI; select your first AI initiative; explore deployment strategies; anticipate timescales; predict required budget; and establish the cultural buy-in necessary for success.
At our AI Masterclass, Richard Potter (Founder & CEO, Peak) explored the Playbook and shared his own best practices for AI Strategy. Tune in!
Recognise AI’s potential for value creation
“Always focus on the problem you’re using AI to solve.” Tim Sadler, Tessian
AI is a powerful set of techniques offering companies tangible cost savings and increased revenue. Further, adoption of AI is ‘crossing the chasm’, from innovators and early adopters to the early mainstream. While you should not attempt to add AI to your initiatives for the sake of doing so, and should be mindful of its limitations, you risk losing competitive advantage if you fail to explore what AI can offer. Approach AI based on its transformational potential.
To engage effectively with AI, separate AI myths from reality:
Identify appropriate problems
AI can be effective at solving problems — but it’s important to begin with a clear problem in mind. Broad considerations are insufficient. When creating a list of potential AI initiatives, develop a precise definition of a problem you wish to address. Do you have a problem whose solution will add value within the business or to customers? Can the problem be solved using AI? AI is particularly effective in four problem domains: assignment; grouping; generation and forecasting.
All businesses will have challenges of the types above — and therefore problems to which AI can be usefully applied. The table below provides examples of popular AI use cases.
There are many ways to identify and evaluate potential AI projects, including:
- Network: to familiarise yourself with AI and its use cases, engage with your professional network and AI communities on LinkedIn and Meetup.com (on Meetup.com, communities can establish informal gatherings and there is a thriving AI community). Many community events are free. Ask an attendee for a coffee and you will find a useful sounding board for your questions and ideas. Informal advice is valuable; you can discuss whether AI might be suited to your use cases, why, and how to turn your idea into an initiative.
“Find someone who is already using AI and bounce your ideas off them. Work out if your idea is possible. Have that conversation before even thinking about a consultant.” (Miguel Martinez, Chief DataScientist, Signal).
- Conferences: seek inspiration, talk with experts and understand industry best practises through conferences. Conferences tend to be high-level executive briefings, sales pitches or presentations of academic research. If you are early in your AI journey, prioritise events with multiple tracks for less experienced practitioners, or a mixture of levels so you receive an overview. Useful conferences will provide access to companies with successful AI solutions, which you can talk to for advice and collaboration. The cost of conference attendance varies from several hundred pounds to several thousand. Familiarise yourself with sessions before you book to ensure a return on investment.
Prioritise projects according to value and viability
Once you have ideas for AI projects, beyond assessing the relative value of each to your company, determine the most viable by addressing the following questions. As well as enabling you to choose a feasible project, the answers will help you define project parameters and objectives.
- Problem: Does the project fall within the definition of assignment, grouping, generation or forecasting? If you cannot clearly define the type of problem, it may be a viable undertaking but is unlikely to be an ideal first AI project for your company.
- Definition: Can you state the problem clearly and concisely? If not, you will lack a clear definition of the system’s purpose and will struggle to select and employ appropriate AI techniques.
- Outcomes: Can you define the levels of accuracy and speed the system must achieve to be successful? Avoid initiatives that lack these measures. If converting an existing manual process, do you know the accuracy and speed of the current workflow? If you are undertaking a new initiative for your company, define what will be deemed a successful outcome.
- Data: Do you have sufficient data to train and test a system? Without adequate, high quality data to train your system your initiative will fail. If you are choosing between a range of otherwise viable projects, select the engagement supported by the greatest quantity of high-quality data.
“Without adequate, high quality data to train your system, your initiative will fail.”
It can be challenging to assess data suitability. Typically, data must be:
- Representative: Data you use to train your AI model should reflect the data you will feed your system in its live environment. If the data differs significantly, results will be poor even if the accuracy of your system during training is high.
- Diverse: Even rare situations should be captured in available training data. Without diverse data, your system may not generalise effectively. Overall accuracy may be high, but your model will fail (misclassify, wrongly correlate or poorly predict) in less frequent situations.
- Balanced: A biased data set produces a biased system. Does your data have inherent bias? For example, are you analysing CVs for suitability to a role and most candidates are of the same gender? Liaise with individuals in your organisation who understand your data and can advise on its inherent bias.
- Exhaustive: All relevant variables must be included in the available data. For assignment and grouping problems, missing variables will lead to oversimplified results (unwarranted correlations). In other problem domains, you may be unable to derive utility from your system.
- Sufficient: While a smaller volume of high-quality data is preferable to extensive, poor-quality data, the volume of data you can acquire must be sufficient to train your algorithm well. For assignment problems, useful results frequently begin to emerge after approximately 1,000 examples for each output label. Some problems require more or fewer. For forecasting problems, you may require data spanning at least double the duration of the periodicity of the item forecasted. For grouping and generation challenges, typically output improves with data volume but again 1,000 examples are frequently a minimum. Typically, the more complex the challenge, the more data points you will require.
In Chapter 3 of the Playbook, we explain how to develop a full data strategy to support your AI initiatives.
Timescales will extend non-linearly with accuracy
Timescales for AI initiatives can be less certain than for traditional software development. AI systems cannot predictably be developed once, tested and then deployed. Typically, multiple cycles of training are required to identify a suitable combination of data, network architecture and ‘hyper parameters’ (the variables that define how a system learns). These dynamics will vary according to domain, the nature of the problem and the data available. Accordingly, it can be challenging to predict or automate AI initiatives unless they are similar to projects you have previously undertaken.
While timescales will vary according to the problem you are addressing, the resources you have committed and the buy-in you have achieved, you can frequently develop a prototype within three months. It may take days to develop a first version of a system that offers 50% accuracy, weeks to improve the system to 80% accuracy, months to achieve 95% and much longer for additional incremental improvements.
For straightforward problems, expect a similar progression but over shorter timescales. For particularly challenging problems, which require extensive data to describe the problem or new techniques to solve it, this timeline may extend significantly.
“Solving really hard problems using AI takes time and depth. It follows a different curve” (Fabio Kuhn, Vortexa).
Align your budget with your goals and deployment strategy
The budget you require for your AI initiatives will depend upon multiple factors including:
- the nature, complexity and domain-specificity of the projects you undertake;
- available, and preferred, development strategies (use of third-party services versus an in-house development team);
- availability, quality and consistency of relevant data;
- a well-considered starting point;
- regulatory and ethical considerations to be addressed.
Some challenges can be addressed with a readily-available third-party application programming interface (API). Others may be solved with a single pass of data through an existing, public domain network architecture. Others still will require extensive research and multiple iterations of training and adjustment to meet success conditions.
“Costs will vary according to the development strategy you select.”
The following strategies offer progressively greater functionality and uniqueness in return for increased spend:
- Third-party APIs: If another company has already solved your business problem, and you need only call the counterparty’s service via an API to receive a result, prices can start as low as several hundred pounds. Using third-party APIs is the fastest way to deploy AI in your company and requires minimal time from your existing development team.
- Bespoke third-party services: To obviate the need for your own AI team, you can engage third-parties to develop and train your AI models. You will need to gather and prepare your own data and have a broad overview of the process of creating models. You are unlikely to require a budget of more than a few thousand pounds for training and running costs, plus the cost of an individual — ideally a data expert already in your business — with an understanding of AI to manage the process.
- A small, in-house team: A dedicated in-house AI team is likely to cost at least £250,000 to £500,000 per year, even for a small team. Whether you seek to repurpose publicly- available models, or solve unique problems, you will need to pay for: two to four individuals; the hardware they require to train and run their models and potentially extra hires for productionising the resulting system.
- A large, in-house team: An extensive team, recruited to solve problems at the edge of research, will require a multi- million-pound investment in personnel and hardware. This investment may yield a unique AI offering. It should only be considered as a first step, however, if your challenge cannot be solved with existing AI techniques and solutions, if you have access to unique data, and if you face significant restrictions on your ability to pass data to third parties.
We describe, in detail, the advantages and disadvantages of different development strategies in Chapter 4 of the Playbook.
You may wish to develop a proof-of-concept, using your existing development team and third-party APIs or paid services, before creating a budgetary proposal. Most companies then start with the small, dedicated AI team.
Seek sponsorship from senior executives
Support from senior management in your organisation will be important for new AI initiatives to succeed. Your company may have a Board that strongly favours adopting AI; that sees AI as over-hyped and irrelevant; or has a healthy scepticism and seeks validation of benefits before assigning extensive resources. To build support within your company, define the focus of your first AI initiative and then set realistic goals. Your system will not, and need not, offer 100% accuracy. If it can save effort, even if results require human verification, you can deliver increased efficiency.
You can then present to senior management a plan that includes:
- a statement of the problem your AI will solve;
- a summary of outputs and benefits for the company;
- details of the nature and volume of data required;
- a viable approach with realistic timescales.
Leaders may be reluctant to invest in technology they do not understand. To achieve buy-in, it may be necessary to educate senior management regarding the benefits of AI while setting realistic expectations regarding timescales and results.
Anticipate and mitigate cultural concerns
When deploying AI, anticipate the potential for cultural resistance. For many in your team, AI will be unfamiliar. Some employees will see their workflows change. Many employees are concerned about the impact of AI on their job security.
Frequently, AI will enhance an individual’s role by delivering what is termed ‘Augmented Intelligence’. AI can bring new capabilities to an employee’s workflow or free a human operator from repetitive, lower value tasks so he or she can focus on higher value aspects of their role.
Address concerns proactively by highlighting the ways in which AI will support individuals’ goals and workflows — and enable your team to redirect their time to the most engaging aspects of their roles.
“We go through a change management program to educate the workforce. We explain that AI takes care of repetitive tasks so people can focus on bigger things” (Dmitry Aksenov, DigitalGenius).
Address non-traditional security considerations
AI systems can be attacked in non-traditional ways. If a classification or grouping system is given an input beyond the scope of the labels on which it has been trained, it may assign the closest label it has even if the label bears little relation to the input. Causes of confusion, more broadly, may be exploited. Malicious individuals have manipulated system inputs to obtain a particular result, or to disrupt the normal practise of AI systems (for example, by spraying obscure road markings to confuse autonomous vehicles).
Protect against malicious activity via thorough system testing and exception handling, undertaken from the perspective of an individual deliberately attempting to undermine or exploit your system.
Your long-term strategy should incorporate evolution and extension
When your first project is underway, anticipate the longer- term aspects of your AI strategy. Then “obsess about capabilities to make your vision come true over five to ten years” (Timo Boldt, Gousto). Your long term AI strategy should consider:
- Maintenance: To maintain your system’s intelligence, regularly test results against live data to ensure results continue to meet or exceed your acceptance criteria. Set aside budget for future updates and retraining and monitor for performance degradation. Chapter 5 of the Playbook provides a blueprint for maintaining AI systems effectively.
- Data: Monitor changes in your data over time. As your business grows or changes focus, data fields (including time series data, languages and product characteristics) will evolve and expand. Retraining your system regularly should be a component of your long-term AI strategy. To develop a comprehensive data strategy for AI, see Chapter 3 of the Playbook.
“Remember that AI is a capability, not a product. It’s always improving” (David Benigson, Signal).”
- Algorithms: AI techniques are developing rapidly; what you create today may be less accurate and slower than systems you develop in 12 months’ time using the same data. Ensure a member of your team understands advances being made in AI and can advise on when to apply them to your use cases.
- Scaling: A plan to leverage your existing AI systems by extending their deployment to additional business units and geographies.
- New initiatives: A roadmap of new use cases for AI within your organisation to deliver increased cost savings, greater revenue or both.
- Legislation: Developments in AI are being monitored by legislative authorities (see Chapter 6 of the Playbook). Develop a strategy to comply with new legislation as it emerges.
💡 Interested in AI? Sign up for our blog posts.
