Revolution When: The Limitations of Agents

AI will soon revolutionise the world economy. Some people think it already has. Certainly, Generative AI has had a significant impact on the public debate. It's had less of an impact on business operations, although most analysts think that's only a matter of time. Predictions tend to underestimate the amount of time it takes to adopt new technologies. For a technology as immature as Generative AI, this is especially true.

For all the hype, analysts are guilty of overestimating the number of situations where AI supported decision-making is actually needed. In many enterprises, the biggest barrier to automation tends to be the lack of a defined process. Too many workflows are ad-hoc or highly manual. Compliance barriers and internal politics are also a major problem, which limits the workflows that executives are willing to automate. AI can't help with any of these issues. Where agents would be transformative is in overcoming technical barriers to system integration. However, they're simply not capable enough to do that at the moment.

Screen Agents

Just last week, OpenAI launched the first version of their Operator agent. This followed similar demos from other AI companies last year. Operator gives ChatGPT the ability to interact with web pages, allowing Large Language Models to fill in web forms and SaaS applications. At the moment, its accuracy is limited. It struggles to complete many basic end user activities. However, I can think of several legacy applications that would benefit from automation but don't have the API access to support direct integration. Operator would be useful for those apps.

The trouble is that Operator is a workaround rather than a long term solution. It will always be more reliable to directly integrate two apps together rather than relying on AI to fill in a web interface designed for human use. In that respect, Operator is the AI version of an Excel macro. It works as a quick fix but isn't ultimately scalable and opens up a bunch of security loopholes. If companies want to properly automate their business processes then investing in new technology is a better solution.

Content at Scale

Instead, the most important capability introduced by agents is the ability to automatically generate content at scale. Generative AI works best when rewriting existing content for a highly specific audience. A fully trained Large Language Model can tailor content to the recipient far more effectively than any human copywriter. However, the human option is still better when writing new copy for a mass audience. That's a bigger issue than many AI vendors realise. In B2B, personalised messaging is generally the task of sales rather than marketing. After all, marketers are primarily writing for a broad audience. One to one communication is the responsibility of sales.

As such, where AI agents do get adopted in 2025, they won't be used by marketers. It will instead be to help sales and customer service teams provide better content to customers. We're already seeing that with Generative AI chatbots. For marketing use cases, manual prompts submitted through a ChatGPT style web interface are sufficient. Marketing content generation workflows don't need to be automated, because volumes are lower and the content being generated is bespoke to the intended campaign pitch.

Cost vs Benefit

All this means the big winners of the AI agent boom will not be your traditional CRM or marketing automation platforms. It will be the tools sales reps already use for collaboration or content generation. Thankfully, both Microsoft and Salesforce appear to have realised this. Slack and Office are being touted as the main way of interacting with AI agents, rather than through specialist applications.

Ultimately though, cost is the single most important factor determining whether a ChatGPT Pro subscriber might choose to deploy Operator or another AI agent. After all, agents are hardly cheap to run - although the cost will come down in the long run. For now, consumption based pricing is inevitable. Agentforce is already priced this way, as are custom agents built using Microsoft CoPilot.

Recent comments from Sam Altman indicate that usage caps will even be coming to the most expensive of AI services. That's an issue for consumers, but not for businesses. The concept of consumption pricing is familiar to CFOs - cloud computing services are billed this way. It will slow down corporate adoption of AI in the short term though. Eventually, a tipping point will be reached. At that time, widespread adoption of corporate AI will become unavoidable. We're a long way from that milestone.