Felicia Chen
Felicia Chen
Updated: July 1, 2021

Building Blocks Every Enterprise Needs for a Natural Language Automation Strategy

Building Blocks Every Enterprise Needs for a Natural Language Automation Strategy

Enterprises are adding more natural language projects and terms each day. This creates more pressure on organizations, which makes it difficult to act strategically and results in cognitive debt. “All tactics and no strategy is expensive and limits intelligence,” says Anthony Mullen, Senior Director Analyst at Gartner.

During the Gartner 2021 Application Innovation & Business Solutions Summit, Mullen led the “Strategic Roadmap to the Language-Enabled Enterprise” session. To help companies avoid this cognitive debt, Mullen laid out the steps and components that are needed to assemble a strategy for automating how machines understand, process, and generate natural language.

Takeaways

  • “To be strategic, we have to get to grips with vendor/technology strategies, orchestrating and scaling the virtual + human workforce, managing the end-to-end data pipeline and information architecture.”
  • “Unbundle the confusion of technologies and vendors into three core competencies to develop your strategy: information architecture, application engineering, and AI and algorithms.”
  • “You don’t need to model the whole universe of language: general knowledge can be free, data common across industries can be bought, industry-specific and company-specific data should be developed and controlled.”

Understand vendor and technology strategies

There’s been a rapid increase in the number of vendors and technologies. Technology has also evolved, changing how we automate experiences and language.

Markets are also consolidating, giving way to language platforms. Historically, natural language solutions have been tied to a specific modality, such as email, speech, or search. However, you don’t need different approaches to modeling the language and topics that are used for each modality. The models that you develop should be able to be used across multiple modalities. Now, we’re starting to see multimodal language solutions, where a single platform can handle multiple modalities by leveraging the common components and data.

Natural language automation requires sociotechnical engineering

You don’t just need technical engineering for natural language automation. You also need sociotechnical engineering. Sociotechnical engineering combines technical elements, such as data, algorithms, and infrastructure, with social elements, like domain experts, processes, and culture.

Gartner found that when you introduce natural language automation, about a third of the staff will be happy to engage with it, a third can be convinced, and a third will try to break the technology because they view it as a threat. While it’s true that automation has taken away some tasks from humans, it’s also given them new tasks, roles, and responsibilities.

Three new roles have been created due to natural language automation:

  • Exception handler: When the machine can’t complete a task, the task will be passed to the exception handler to complete.
  • Trainer: Humans are trainers of these automated systems. When machines can’t complete the task, you don’t want to just hand the task over to the exception handler to complete. You also want to train the system so it improves.
  • Quality control: Organizations need to insert humans throughout the pipeline to do quality control. This role checks the language output of the automation and compares it against human output.

To succeed with sociotechnical engineering, it’s critical to have an interface that humans can use to give feedback to the AI. For example, indicating which outputs are correct or incorrect. Having a user interface also moves you toward explainable AI, where humans can question the AI and look at the rationale behind the AI’s answer. If you don’t have these interfaces and your vendor can’t provide them for you, you’ll need to design them yourself.

Sociotechnical engineering can’t just happen on a per project basis. It needs to be ongoing, and it’ll take a variety of parties and skill sets to accomplish. You’ll need to work with vendors to get your data in shape and bring in third-party proprietary data, NLT scientists to develop an information architecture, and expert staff members to refine your model.

Develop your strategy around three core competencies

In putting together your natural language automation strategy, there are three core competencies that you want to address:

  • Information architecture: The concepts and objects that are the model for your organization are also the foundation for developing language automation. Build concept and object models for your industry and adjacent models. Keep in mind, you don’t need to do this all at once. Avoid the cognitive debt by working across different business units and joining up your efforts. Everything else outside of this can be bought or integrated.
  • Application engineering: Designing, integrating, and orchestrating natural language automation requires a balance between building and buying. There aren’t enough developers to build everything yourself, and by only working with vendors, it’ll cost a lot. Involve as many staff and in-house experts as possible in the process. Empower them by giving them the tools to build language automation and use external parties to complement them.
  • AI and algorithms: Advances in AI are still happening, so expect rapid change over the next few years. Design your systems with this in mind by making sure you can pick and choose from a variety of models and engines. Only build your own models, algorithms, and engines if you need differential performance.

“Realize your strategy with a three-tier buy/build architecture”

Mullen says “90% of the chatbots live today will be discarded by the end of 2023.” That’s a lot of money that’ll be lost. To avoid this, you want to choose the best-of-breed technologies to develop with, have flexible design tools that can be used to create experiences across any modality, and build and refine your information architecture.

In bringing your strategy to life, “You don’t need to model the whole universe of language: general knowledge can be free, data common across industries can be bought, industry-specific and company-specific data should be developed and controlled.”

If you’re ready to bring natural language to your voice channels and automate customer service, see how an enterprise company realized tremendous cost savings with Replicant Voice in six weeks.

Download The Total Economic Impact of Replicant Voice study

Gartner, ‘Application Innovation & Business Solutions Summit – Americas’, Presentation (Strategic Roadmap to the Language-Enabled Enterprise), Anthony Mullen, May 26-27, 2021