Artificial Intelligence (AI) comes into play to augment and guide people through operating complex processes, consuming the exponentially growing volume of information and extracting actionable insights from the noise. While there’s been an acceleration in the adoption of specialized AI technologies, we’re seeing most of this come from organizations with access to differentiated talent and significant funding.

With AI technologies and training datasets both increasingly available in open source, here’s a blueprint to start your company’s journey toward operating an enterprise-grade AI function:

Identify an AI-worthy problem

First and foremost: Don’t acquire a hammer and then look for nails.

AI implementations aim at applying humanlike reasoning to solve tasks. In enterprise settings, they promise to step away from decades of linear thinking with complete coverage of all processing rules and approach solving problems in a more abstract manner. AI models are slow learners but extremely fast operators, finding patterns across diverse, complex ecosystems.

Start with a problem that seems repetitive, has a low to average level of ambiguity on its execution and has a concrete yet slightly variable outcome. For example, inserting a cognitive assistant in the form of a chatbot between your customers and your customer representatives can free up your reps to focus on other priorities while the chatbot fields simple questions.

Additionally, baseline the KPIs that will demonstrate the validity of your implementation: How many customers say the chatbot answered their questions, and how much time is freed up for your representatives’ other priorities as a result? Ultimately, each implementation has a cost, and you need to demonstrate a longer-term value generation. Using an FTE proxy to calculate savings is a good start — top-line growth might come later.

Select a technology and (a lot of) data, and prove the concept

Once your use case is identified, and you’ve formalized enough assumptions to establish a direction, it’s time to demonstrate feasibility.

Depending on your budget and expertise, you could try a click-and-play solution offered by any major cloud vendor (Google, AWS, Microsoft or IBM), adopt a containerized tutorial on long-term or short-term memory assembled for a chatbot to your use case or partner with an AI technology vendor to handle the development.

From there, you will need data. Favor a data set that critically represents your day-to-day operations, and prepare it such that it has 1) enough variability to identify distinct patterns, 2) enough volume to disprove those patterns and 3) ideal tags to tie output from the chatbot to expected outcomes. Many AI-flavored projects fail to start because they have a good problem to solve but not enough data to work to its core.

Once you have this data set, confirm that you have assembled the right models and AI techniques and that you are addressing your problem the right way.

Test and stabilize your prototype

After demonstrating the feasibility of your idea, it’s time to bring in testers.

With a stable model performing well in a controlled environment, assemble a team of internal testers. This team can run more unstructured, unexpected tests that represent real end users, allowing rapid-fire iterations and strengthening the models to be the core of your AI agent and grow its intelligence.

Next, compare the traditional approach outputs and the outputs from the AI agent. Where the tests are inconclusive — and the models seem too unstable with variability of inputs — you will have to decide if the problem worked in isolation but cannot be solved for real enterprise context or if you need to improve the models you’ve assembled. Your data set may be too small, too homogeneous or too specific to a sub-strain of events to support the real-life execution.

Once your AI agent is better or on par with your staff, you can start working with your colleagues by managing infrastructure and corporate portals to integrate your new digital agent within your enterprise environment.

Time to go live

As you approach production, establish a parallel run between your traditional operations and the AI-enabled ones. Similar to A/B testing, start by exposing your chatbot to a select population of customers that won’t hurt your business in case of failure (i.e., customers you know to be receptive to a new digital engagement paradigm).

As the bot succeeds, it reinforces your model selection; as it fails, it creates an opportunity to actively gather feedback for an end user and/or passively gather feedback from the interactions to inform the next release.

Once your AI agent reaches an acceptable level of stability and performance, you can reduce the human oversight and focus on creating an automated feedback loop with a machine learning module. Every touch point between the output of the AI agent and an end user becomes a teaching/learning opportunity.

Scaling across the enterprise

You’ve done it — your chatbot is in production, and you’ve generated enterprise value. The next question: Should you pursue an enterprise AI function, or is your organization not quite ready?

More than likely, you are ready, and this success should create a thirst for more.

There are countless statistical models and dozens of ways AI can be adopted and deployed to create enterprise value. The adoption blueprint in itself is all about incrementally increasing the breadth of effort and reliability of the technology.

When sharing your learnings with your organization, you will have to expand your influence and collaboration to sustain momentum:

  • Create an AI strategy and business case.
  • Acquire AI expertise.
  • Create an information management function to feed individual cases.
  • Communicate and celebrate, internally and externally, to gain sponsorship and attract more talent.
  • Have strict business continuity procedures in place to regain control over failing AI agents in mission-critical activities.
  • Continuously monitor and curate the behavior of your AI agents — they might deviate from the baseline and require calibration.

That’s the blueprint to start your AI adoption journey: five steps that will take you from dabbling with AI to establishing a market-differentiating and enterprise-value-creating function.

Categories: Blog


Leave a Reply

Avatar placeholder

Your email address will not be published. Required fields are marked *