An approach to AI Requirements in an existing set-up

AI is a new, profound technology frontier like nothing before — including fire or electricity.

Unfortunately, it’s easy to be disappointed with the overhyped benefits that AI claims to deliver. This is especially true with the overpromising done by those with little knowledge about on your business-specific data, process, technology or the expectations of your end-users and clients.

One of the root causes of this disconnect is a lack of guidance in writing AI requirements.

Here is an approach to capturing AI requirements in an existing data-centric operations (Of course, this acknowledges that a different approach will be needed for AI projects in a new set-up).

Discover

A good place to begin an AI journey is with the assumption that all existing processes and sub-processes within your enterprise are likely candidates for automation and your operating model is ripe for re-imagination.

Start by establishing a small and diverse multi-disciplinary team (“MDT”) involving, at a minimum, your experts from the following: 

Technology

Operations

Data 

Potential technology partners

The first set of tasks for this seed (“AI Prototype Factory”) MDT is capturing all workflows, human decisions, data variations, and time-consuming activities to identify the opportunities for improvements — either through process simplification, standardisation, optimisation, technology interventions, or combinations of these.

No idea is small or trivial. 

In fact, the simpler a process or quicker a human decision is, the more likely it’s a good candidate for AI ML automation. 

Focus on the scenarios, for example, how the processes work, steps in a workflow, how human decisions are made, the volume and human effort involved, and end-user expectations. Explore how different business-specific data variations influence the above points.

Here’s what a sample AI Requirement output looks like:

Based on available data quantity and quality as well as the promised AI capabilities, hypothesise your confidence in meeting end-user goals, accuracies, ease of implementation, cost and time of implementation, and scalability. Prioritise use cases based on this hypothesis and potential ROI.

Analyse your processes for prioritised use cases and business-specific data with the intent to identify problems or opportunities that are most likely to be strong candidates for AI automation (“AI Viable”).

Prototype

The primary purpose of Prototyping is a critical evaluation of the promised AI technologies and the anticipated outcomes with real-world data. 

Through a series of Show & Tell (S&T) sessions, validate AI technical solutions and the results achieved against the hypothesis. 

Results must be supported by “Evidence-of-Thinking”— an explanation about what and how AI algorithms or techniques will be applied, details of third-party APIs as well as libraries and products that will be part of the proposed solution. 

Based on S&T experiments, experiences, and AI approaches, explore opportunities for process re-imagination and technology transformation. The levers above shown on their own and together enable the re-imagination and re-wiring of the existing operational processes, IT systems, business rules, commercial models, and end-customer engagement. This design process is crucial for the successful adoption of AI within the enterprise.

At the end of the Prototype phase, the authorised MDT chair makes a go/no-go decision, which should be based on the S&T results, a revised hypothesis, indicative costs, timelines for the next phases, and potential ROI. 

If the results achieved deviate beyond an acceptable gap, deprioritise or remove the use case from the pipeline. Notably, this is still a success for the Prototyping phase. It validated the hypothesis against real-world data and the current set-up then avoided costly next steps.

For the approved use-cases, the MDT must acknowledge that all members understand and agree to the newly proposed hypothesis, committing to interdependencies and other expectations in a timely manner.

Iterate

One of the earliest use cases for AI automation was auto-matching two advertisement videos broadcast by national television over a time period. Our AI Prototype Factory MDT overzealously accepted and promised a MVP with high accuracy performance.

As shown below, with each interaction between multi-disciplinary teams, we discovered new aspects that the MVP must adhere to. While the AI Prototype Factory MDT was at unease and questioned “why were all specifics not shared earlier”, the client MDT was resolute that “it was a common sense in their business domain”.

In addition to the above requirement-centric challenges, the existing operational and IT set-up included its set of requirements. The IT department wanted to refresh existing systems. The Operations Department wanted to adopt simplified and optimised workflows. Commercial teams wanted new capabilities to serve pent-up client demands.

For the above reasons, it is not possible to write down all specifics of use-case and data variations at the outset of an AI project. We should accept that the requirements will be discovered iteratively. The delivery, commercial, and procurement teams’ expectations should align to that reality.

This all points to the importance of and need to anticipate iteration.

Another set of requirements that are critical, complex, and ever-growing is for Responsible AI. This extremely important set of requirements for organisations and the society at large must be part of every AI project. Below I’ve illustrated the various requirements (knowledge components) that will grow in complexity over time. Some of these are critical to consider at the outset, for example, AI risk assessment for bias.

During the go/no-go decision making stage, even though “AI Viability” and business case is high, the MDT should also evaluate “AI Suitability” considering the Responsible AI matrix.

Re-imagine

It’s clear that a diverse multi-disciplinary team is crucial to the success of any AI project. However, how should we keep all parties informed, involved, and influenced by each others’ work? For our playbook, I drew inspiration from teams designing rockets, where different teams work on many interdependent components in parallel and separately.

Learning from this example, establish an “Integration” forum comprising of different groups from within the organization and external teams. Through biweekly meetings, share individual progress and all the deliverables that influence each other’s work, for example the AI Requirement Matrix, Evidence of Thinking, or definitions of micro-services.

The AI Prototype Factory MDT should publish and fervently update the AI Requirement Matrix and externally consumable micro-services. Similarly, an IT team undertaking any major tech refresh should post assumed automation capabilities and re-imagined process details. Meanwhile, Operations leadership should champion knowledge collaboration, continuously engaging the Commercial teams for process and commercial re-imagination opportunities.

As all member groups’ work influences others, constant communication of changes and acknowledgement of them is crucial.

For full realisation of the potential benefits from AI programs IT, Operations, and Commercials team should use this opportunity for a holistic re-think of existing process and technology, rules of client engagement, and commercial agreement.

AI programs provide a strategic opportunity to re-wire your organisation with a new operating model and compelling market propositions.