Getting AI-Ready Is Not About AI

It is about institutional knowledge and complexity. Depending on the specific industry and function, there are prerequisites to work through in your own contexts before celebrating the efficiency boost from a full-blown AI project. Assuming you have an initial list of processes or workflows in scope, a thorough assessment of whether they are ready for the implementation is still necessary.

Data

Take client information as an example. There may be multiple systems holding data about the same client. Is each system up-to-date? If multiple systems store "client industry" as an attribute, are they all consistent, and if not, what's the next step? If you find your systems or files missing data or carrying duplicates, it's time to consider how data is (and should be) distributed, and who owns the missing pieces. If any client preference lives in unstructured data forms such as emails or PDFs (no technology team likes them!), should it be added to a system, or should the extraction of it become a process for AI?

Process

Compare the procedure as written with the procedure as run, and note any inconsistency between the two. What are the sources of inconsistencies? Are there any unwritten rules or norms understood by the team but not captured? For workflows that are judgement-heavy, is the decision-making process documented? If it's important for your colleague to know about who to reach out to when additional input is needed, it's important for AI to know too.

Org Chart

The organizational chart shape both of the above, and ultimately how to productionize the prototype and scale it.

Think from a long-term perspective: how can you design a system so that AI can plan, execute, test, and review as autonomously as possible, while it only pauses when a step requires a human judgement call? Design it to prevent human beings from becoming the bottleneck. You may have working groups, pre-meetings and pre-pre-meetings to align internally and with stakeholders, or work may be completed but can only be reviewed a week later. If AI is waiting for working group decisions or reviews, you've uncovered the bottleneck. Decide what should be reviewed by AI and what should be reviewed by humans; it may be a dynamic decision to be updated periodically. I expect that how responsibilities are divided between human and AI, and between different teams, will change significantly.

Crucial topics such as access control and sensitive data protection are discussed more and more, and they should be. Demos are run in a controlled environment with a clear boundary and clean data, while production environment is much trickier. For instance, AI may need real-time authentication to access information on a need-to-know basis. AI governance topics need to be thought through before productionizing a prototype.

Context

Context is the critical concept not to be missed. Context is often in unstructured data scattered in every corner. Your key stakeholders' views and the concerns underlying them, a meeting three months ago that decided an approach was not a viable option, an industry convention that is not written down: all of these can be no-brainers to you while AI is similar to your new colleague who learns fast but has no prior institutional knowledge. The question to answer is how to feed AI with high-quality context so it can do the right thing and follow best practices. Depending on the number of legacy systems there are, it may be worth considering whether a centralized context layer for AI should be used, and, if so, how centralized it should be.

Where to Start

All of these look like daunting tasks; I can imagine the amount of work involved if every issue is investigated in an organization. The assessment for AI-ready itself is a valuable opportunity to clean up the house and determine next steps, while individual tasks definitely require sizing and prioritization. After all, not every process needs AI: many processes require a deterministic outcome only from rule-based engines, and natural language is not a concern. Those processes need something we are quite familiar with: automation. For the processes you scope in for AI implementation, a phased approach may be suitable. Remember to look for quick wins for well-defined problems that can generate the most benefit relative to effort.