Most AI projects fail!
Truth.
Sometimes the failure is obvious and quick. The developed models are never put into operation. Sometimes that is for technical and efficacy reasons. Sometimes it is because of lost interest in the organization. Sometimes key players move on before operationalization is complete.
These are the usual scapegoats. However, they often cover the true reasons for the failure.
Many AI projects fail because the actions they recommend are not delivered to the business user in a usable format, or business users do not trust the AI recommendations. (Let me use the term recommendations for the delivery of AI to business users. Because in many AI projects, that is what is being surfaced to business users: recommendations to use inside of an existing business process. I know that other outcomes of AI models are possible.)
Integration to Business Processes – Big Problems
I have been lucky enough in my career to be part of many analytical projects – both historical and predictive. As time goes on, more and more are Artificial Intelligence related. I go back to a time when AI was not a common term, and you were more likely to hear data mining, predictive analytics, machine learning, and other terms used. The AI technology ecosystem has changed almost as much as the marketing terminology.
The biggest technology change is the ability to surface AI recommendations into the workflow of business operations.
This is where the potential of AI projects is dying today. Getting the right AI recommendation into the right person’s hands at the right moment is how an AI project is successful. Getting that person to trust the recommendation is another thread we need to discuss too.
Background
Not so long ago, the most common deliverable for an AI project was a report. Way back, this report came in the form of printed paper in big binders. Less way back in time, recommendations came in the form of spreadsheets with a document written to aid understanding.
These deliverables were often accompanied by a presentation or executive training session.
Rarely did that understanding filter to business users even if those people used the recommendations.
Then came Business Intelligence (BI) systems, which could report the AI recommendations from a database to the business users. This was supposed to be the quick cure that operationalizes AI into workflows and business processes. With current technology now the same process can happen using program calls to the model. Same, or at least, very similar results as far as the business user is concerned.
However, this has usually not worked well. Almost no one constantly goes into their organization’s BI system to look at reports while trying to get their work done (analysts excepted). Most organization’s BI systems are designed to aggregate data to make it easy for people to look at high-level summaries.
AI recommendations are targeted and precise for individual actions. That is not exactly the opposite of how people use BI, but it is very different.
Even when they see AI recommendations on their screen in reports, it is not easy for the business user to make the immediate connection to the task they are performing at that moment.
The Last Mile
The Last Mile of AI refers to how to effectively get specific action recommendations in front of business users at the time in a process the information needs to be used.
It also includes the work needed to be done to make the business user trust the AI recommendations.
Without trust that what is being recommended is a good or best action, business users will revert to old ways and gut instinct.
Let me re-phrase that for clarity. The Last Mile of AI is a part of the AI operationalization process. It includes how the AI recommendations are surfaced during the normal course of how a business user is performing their job. The Last Mile of AI includes education and resource materials that make business users comfortable with the AI recommendation.
Combined, this is how the AI recommendation is put into action by the business user.
Terminology
The term Last Mile of AI is borrowed from telecom providers getting the last segment of internet connection to the home. It also seems to be gaining popularity for the delivery of online orders too. In general, most people intuitively get that the last mile is potentially a missing connection to the point of need.
There are a few other AI practitioners who are also using the term. It isn’t pervasive yet, but it does seem to be growing. Almost everyone I speak to gets this idea and sees how it makes a big difference in AI project operationalization effectiveness.
Field of Dreams
Too many times, the data science team does not work directly with the business user community. This is just a comment on how things are usually not a critique of specific people or job types.
Usually, an executive works with the data science team to define a problem, sketch the desired outcome, and push the AI project forward. Often this leads to technically adequate AI models that then need to be operationalized.
The data science team are not specialists in the business process affected. Too often, operationalization is done through the BI system because that is the easiest point of delivery. BI, as we covered above, is not usually the right vehicle to deliver targeted, specific AI recommendations into the flow of a business process. As the many failed AI projects in the world attest, building an AI model is easier than getting people to use it.
Integration to Business Processes – the Goal
Working specific recommendations into screens the business user accesses during the processes affected is the best way to deliver AI recommendations. This does require a multi-disciplined approach and way more involvement from different parts of the organization than just dumping AI model results in a database for reporting or standing up an API for programmatic interaction.
The best integration is into the regular business system the business user works in daily. That could be an ERP, CRM, web store administration, HRIS, or any other common business process application.
The integration needs to be inserted into the right screen. Adding additional screens in the menu with titles like “Insights” or “Recommendations” is marginally effective and not really much different than creating reports in a BI system. Adding a field or table onto a screen the business user must-see during the business process is the best method,
Linking actions from recommendations presented directly are even better. There are so many potentials; it is hard to construct examples understandable across all industries but let me try.
A simple one is a screen recommending contacting a customer because they are predicted to buy soon. A link in the customer details section could suggest the best way to contact this customer is by phone or by email. The link then initiates the action when clicked.
Again, that is an elementary example. Actual uses are going to be specific to each business and combination of software used for business processes.
Trust Factor
Determining the best location to put the recommendation is only part of the Last Mile of AI.
Getting business users to trust the recommendation is the other part.
Organizations must make business users comfortable using AI recommendations without question or delay.
Part of the process is a basic education in how the AI model works, including what data it is looking at, how accurate it is in training and testing phases, and the update schedule for model enhancement.
Business users do not need to become data analysts, much less statisticians. They do need to be treated like rational actors in the process who have a stake in the outcome of an AI project.
Enough background on the model needs to be presented, so business users have a sense of trust. Otherwise, they will revert back to what they were doing before the AI model existed.
Last Mile of AI – Takeaway
The Last Mile of AI has two components
- proper location of the AI recommendation
- a program building trust in the business users
Failure to do both means your AI project will almost certainly fail to achieve its objectives.
Of course, this is predicated on having an AI model that is effective. A good AI model is only the starting point of operationalization.
The Last Mile of AI dictates success or failure.
