![](https://d15shllkswkct0.cloudfront.net/wp-content/blogs.dir/1/files/2024/08/a-futuristic-and-dynamic-scene-of-robots-and-gener-tdizzzqsSRaABsGNNNYWWA-xVGNANXESQGFEyF-bEV1xg.jpeg)
![](https://d15shllkswkct0.cloudfront.net/wp-content/blogs.dir/1/files/2024/08/a-futuristic-and-dynamic-scene-of-robots-and-gener-tdizzzqsSRaABsGNNNYWWA-xVGNANXESQGFEyF-bEV1xg.jpeg)
The rise of robotic process automation in the late 2010s heralded a new era of knowledge work automation. RPA bots could emulate human tasks, completing simple workflows via scripted interactions with user interfaces or application programming interfaces.
Though RPA can save organizations money by replacing repetitive human activity with less expensive bots, it suffered from two core limitations: Bots were brittle and added to the organization’s technical debt.
In response, RPA vendors added AI to their bots, crafting a new “cognitive RPA” market – but it wasn’t enough to address these core challenges.
Gartner responded by positing a new “hyperautomation” category, shifting the spotlight off RPA to combinations of technologies that might address the limitations of RPA.
Then along came large language models that support generative artificial intelligence, disrupting RPA, hyperautomation and the rest of the knowledge work landscape.
It seems that every software vendor must have a gen AI strategy these days, and vendors in the automation segment are no different. Today, however, vendors are moving past early conversational interfaces to deeper applications of this transformative technology.
I spoke with eight of the most innovative vendors in this market – a mere fraction of the crazy number of disruptive forces in the market today, but nevertheless representative of the progress vendors are making in addressing the difficult problems of knowledge work automation. (* Disclosure below.)
Here’s what I learned:
The term “agent” refers to an autonomous piece of software that achieves specific business goals independent of other software in its environment. Just how autonomous they are and what they actually do, however, depends upon whom you ask.
The agents from Agents.inc (AGENTS HQ GmbH), for example, connect data sources, interact via prompts with various AI models, and can generate reports, alerts, and interactive dashboards. Agents.inc is focusing on market intelligence use cases for now, although it is capable of far more.
Tines (Tines Security Services Ltd.), in contrast, focuses primarily on cybersecurity-related tasks. With Tines, building automations is so simple that security analysts have the option of building automations proactively, or if the situation warrants, on the spot to deal with a new issue.
For its part, Leena.ai Inc. focuses on service management workflows with the goal of eliminating the need for support tickets.
These choices of use cases are more likely to represent low-hanging fruit opportunities than any technical limitation with the respective platforms. Expect more general-purpose AI agents as the technologies mature.
One of the reasons why RPA bots can be so brittle is because they are sensitive to changes in the meaning, format and availability of data in various fields. Any change to the data in any underlying system can break the bots.
Gen AI can be particularly useful for addressing such data integration challenges. Bardeen Inc., for example, leverages millions of workflows as training data so its AI can automatically figure out what fields a knowledge worker needs to map and their respective data formats. As with other gen AI applications, this capability improves with time.
Agents.inc uses agents to prepare data and then combines AI with knowledge graphs before feeding LLMs, resolving semantic issues and eliminating most “hallucinations” – where gen AI will make erroneous guesses in the absence of relevant data.
The basic idea here is that most data integrations follow familiar patterns. If someone prompts an agent to take data out of Salesforce and put it into HubSpot, for example, then the agent can make a very good guess as to the specifics of the individual fields based upon large quantities of historical data.
Existing offerings of AI-supported automation typically provide “next best action” recommendations – essentially an autocomplete for building workflows, where the AI would guess the next step in a workflow in progress based upon its historic experience with workflows.
Today’s automation offerings offer this capability as well but take this principle one step further. For example, Boomi LP offers a gen AI-based conversational design tool for building both automations and integrations, leveraging training data containing more than 300 million examples.
Bardeen offers several prebuilt automations, and then enables users to build new automations via a conversational interface. To reduce the chance of hallucinations, the platform converts the conversational specification of the workflow to a domain-specific language representation.
This approach turns the gen AI-based understanding of intent into a deterministic workflow for improved repeatability, with the addition of a verification layer that further reduces hallucinations.
More common, however, are low-code, AI-supported workflow construction tools. With Tines, for example, analysts build automations on a visual storyboard by connecting seven basic actions. Analysts can copy and paste third-party APIs into Tines for interactions with any third-party tool or app.
SmythOS (INK Content Inc.) also offers a low-code platform, both for creating AI agents as well as creating workflows leveraging its application orchestration framework.
With SmythOS, users can spin up their own agents and workflows, where each agent implements some LLM-based task. The platform enables users to connect the actions of different LLMs into individual, often complex workflows.
Sema4.ai Inc., however, takes a different approach, focusing on document-centric automations rather than workflow-centric ones.
In Sema4.ai, organizations create runbooks using natural language descriptions of the behavior they require, with the help of an AI assistant. These runbooks, in turn, act as prompts for the AI to build the agents.
Runbooks describe the specific business needs for a document-centric task, for example, invoice reconciliation. Semantic document intelligence within the platform then leverages business rules that are encapsulated in documents for validation.
RPA adds to organizations’ technical debt because it does nothing to solve the problem of inflexible and obsolete legacy business logic and its representation in code. Instead, RPA adds an additional layer of brittle software that organizations must now maintain – increasing the technical debt for the organization overall.
Most of the vendors in this article leverage gen AI to reduce the maintenance burden of their respective AI agents. Updated LLMs and continually improving data sets will lead to agents that get better at what they do automatically – rather than requiring the expensive, largely manual maintenance that burdens RPA bots.
Dealing with the legacy code problem, however, is a different challenge. When no API exists and a custom-built connector isn’t cost-effective, most automation vendors default to an RPA bot as the only remaining alternative.
The one exception among the vendors I interviewed is Beezlabs (Beez Innovation Labs Pvt. Ltd.).
Via Advanced Business Application Programming or ABAP, SAP SE’s proprietary programming language, Beezlabs can interact with SAP directly without either a dependency on the user interface or requiring any customization of the SAP application code itself. The company, in fact, is unabashedly SAP-specific, leveraging its deep expertise with ABAP.
Beezlabs also leverages the open-source business process orchestrator from Camunda (Camunda Services GmbH) to connect SAP tasks with other platforms, for example, Excel and Salesforce. With Beezlabs, organizations can run ABAP bots in parallel on the vendor’s cloud-native Kubernetes infrastructure. These ABAP bots interact with SAP on an internal coding level, rather than via the user interface as RPA bots do.
For SAP customers prioritizing an SAP “clean core” strategy (avoiding hard-to-maintain customizations), the Beezlabs offering provides the best of both worlds: AI agents that run natively on SAP without requiring any customization of SAP itself.
Gen AI-based automation technologies like the ones in this article are disrupting many markets. By improving resilience and lowering technical debt, gen AI agents are rapidly displacing RPA bots – but there’s more to the story.
Earlier generations of low-code/no-code automation platform vendors leveraged visual metaphors (aka “boxes and lines”) to provide workflow creation capabilities. Today, natural language interfaces are becoming the norm, transforming the entire low-code/no-code market.
AI agents also play an important role in organizations’ cloud-native strategies, as each agent can run statelessly in containers. As a result, each platform has the ability to scale them automatically, deploying as many identical agents as necessary to address any situation.
The platform, in turn, maintains the state of each workflow, while providing gen AI-based workflow creation and management capabilities.
The writing is also on the wall for hyperautomation. Rather than requiring a smattering of different tools, AI agent technology represents a convergence.
New gen AI-based technologies are gradually supplanting not only RPA, but also business process automation, low-code/no-code platforms, rules engines, data integration technologies and more.
Stay tuned for plenty more disruption as this convergence trend plays out.
Jason Bloomberg is founder and president of Intellyx, which advises business leaders and technology vendors on their digital transformation strategies. He wrote this article for SiliconANGLE. (* Disclosure: SAP and Tines are Intellyx customers. Camunda and Sema4.ai are former Intellyx customers.)
THANK YOU