Rise of the Agents

Technology trends always arrive in waves. A disruptor releases a revolutionary new feature, which everyone then copies. We've seen this with Generative AI, which has become embedded within every application imaginable, including many that don't need it. Now, we're seeing it again with the next wave of Generative AI. A couple of weeks ago, all the leading CRM providers announced the introduction of AI agents into their platforms. Who says innovation is slowing down?

Of course, none of Microsoft, HubSpot or Salesforce have been the first to market with AI agents. They're an evolution of a concept first introduced by OpenAI at the start of the year with their GPTs app store. GPTs allowed developers to train their own AI models on the ChatGPT platform, which could then be used to automate routine tasks with Generative AI. Developing this capability for in-house enterprise use has always been an obvious next step.

Cost vs Benefit

Now, six months later, it is customer conference season at HubSpot and Salesforce. Both vendors headlined their keynotes with a set of impressive AI agent demos intended to demonstrate the power and ingenuity of AI automation. These demos are important because agents are a critical step in addressing the main question that has held back generative AI over the last eighteen months: namely, what do we use it for? Generative AI was supposed to deliver substantial cost efficiencies and time savings that simply haven't been seen to date outside of a few highly specialised use cases, such as meeting notes and review summaries.

Investing in AI only makes sense from a financial perspective if it can be used for enterprise automation. That's not happened so far because tech firms are incurring unsustainable losses that need a pay-off sooner rather than later. In order to achieve profitability, AI agents need to deliver cost savings commensurate with the resource requirements of a high-end LLM model. That needs end users to identify use cases for automation that cannot be solved by existing low-code integration platforms. Those definitely exist but are probably not as extensive as advocates would like. We're still many years away from a world where business users are routinely creating new AI agents to automate their day-to-day responsibilities.

Use Cases

At recent conferences, Gartner have been warning about the delayed adoption and low ROI of Generative AI, issues motivated in part by the immaturity of the technology. The initial hype around chatbots died down because marketers realised that the technology couldn't be trusted to tell the truth. Air Canada even lost a court case over a false compensation policy hallucinated by their customer service chatbot. Within marketing, there have been some notable examples of AI being used to generate bespoke content at scale. Brands such as Klarna and Juniper have touted the case savings and increased flexibility of AI generated content when compared to content created by creative agencies. However, this only works for businesses with a highly centralised marketing engine that goes to market with a high volume of campaigns.

So far, specialist AI models have struggled amid competition from ChatGPT. There are still far too many AI workflows where business users copy/paste the input into a general purpose AI chatbot, rather than integrating a dedicated AI model for the task. The benefits of bespoke AI models have not outweighed the mindshare and ease of use advantage enjoyed by ChatGPT. As a result, IT departments have highlighted Generative AI as a major threat to enterprise security. Salesforce hope to tackle this problem by leveraging Slack as the user interface for AI agents. That way, users interact with AI agents within a platform they already use to communicate with human colleagues.

Self-Service Automation?

AI agents do solve another critical challenge necessary for Generative AI to reach mass adoption. The types of human-AI work partnership pitched by futurists require customised AI to become more accessible and easier to customise. Prompt engineering is arcane enough. Fully training a new AI model is generally seen as a task for developers, even if it's built on top of a major LLM. A code-free method for training AI is needed, in the same way that the likes of Zapier and Tray.io have democratised enterprise integration. AI agent builders deliver this. Much like low code automation tools, agents expand the range of users who can automate business processes to include operations teams as well as development teams. Previous waves of self-service automation have substantially accelerated digital transformation initiatives by allowing developers to focus on more complex integration scenarios that code-free integrations can't handle. Basic scenarios can then be managed within business teams.

If there's one thing we can learn from the demos, it's that AI can automate 75% of many routine tasks. Any business process which follows a defined procedure with clear guidelines around decision-making can be automated using AI or other automation technologies. However, this has been true for some time. Even today, automation is not limited by technology. Instead, poor alignment and a lack of process are the key barriers to adoption within many organisations. Automation is only useful if it delivers an output that businesses want. To automate an existing business process, there needs to be clear business requirements and a defined procedure to follow. AI doesn't change any of these issues.

Generative AI does expand the range of business inputs that can be automated due to its ability to interpret free text in a way that wasn't possible previously. The question is about how to deal with scenarios and inputs that the model isn't trained to handle. After all, no business process is foolproof, and even manual processes have exceptions that weren't considered in the process design. Increasingly powerful models are reducing the frequency of those exceptions but will never be able to eliminate them entirely. Human workers can make a judgement call based on experience, and companies are generally willing to accept that. It will be a long time before AI agents are trusted to make decisions without human supervision.