Main story – What should businesses build with agentic AI?

Phil has had the opportunity to slice and dice agentic AI across a large number of vendors. Now it converges that into a burning question: what to build? Phil:

My first takeaway is that all this talk about agentic AI actually boils down to something pretty simple once you strip away the marketing hype and its smoke and mirrors. Think of this new generation of agents as a more flexible user interface that leverages existing systems and data. Previous generations of chatbots and agents could only execute highly structured instructions in a predetermined manner.

A good summary of why Gen AI agents are different. But there’s more: unstructured data plays a key role.

But generative AI means these new agents are much better both at extracting meaning from unstructured information such as conversational interactions or document collections, and at mapping possible actions to anticipated outcomes. Whereas before you had to carefully map out each step you wanted an agent to take, now you can simply tell the agent: “Follow the policies and processes written in this set of documents and choose the actions that will produce the outcome “. desired result with a given data set. Then come back with me for approval before continuing.

Kudos to Phil for that last line: “come back with me for approval,” which many vendors have been happily pass into silence. Generative AI agents are more adaptable and flexible than their predecessors, but they are also less precise, especially when ready to use. Human supervision will sometimes be necessary. Publishers have mixed feelings on this point: they want to reassure the human worker, but they are aware that the human loop impacts ROI and productivity gains. We’ll see. For now, the fact is that Agent workflow design is important and will vary based on accuracy and outlier tolerance per use case. And, as Phil points out, data quality is an important factor:

But there is a big problem. The agent’s ability to respond to this request is highly dependent on the accuracy and robustness of the data and automations in the underlying system. This is what Benioff is getting at in his criticism of Microsoft Copilot. He believes that Salesforce has a much more robust object model than Microsoft for making sense of information and performing actions in its applications. We’ll hear more from Microsoft at its Ignite conference in a few weeks.

Yes, prepare your popcorn. Phil met virtually all of my agent qualification criteria, including:

• appropriate design and human supervision
• quality of pre-existing data and reliable automations that the agent can invoke

As for data quality, it can also take the form of a RAG and/or knowledge graph architecture:

Companies like Asana and Atlassian make a similar argument based on their creation of proprietary work charts that map the different entities and relationships managed by their applications. They’ve already done the hard work necessary to structure all the information in their systems, giving their agents a head start in making sense of it.

Despite my criticism of agent hype, I like the ability to “orchestrate” pre-existing automations into new sequences, stitching together different tasks, perhaps even on the fly. So, Phil, what should we build? He cites examples from recent clients, the so-called “low-hanging fruit” that every software company seems to have. In the case of Salesforce:

Salesforce showed off some Agentforce customer case studies, with educational publisher Wiley and restaurant reservation service Opentable both freeing up significant resources in their customer service teams with automated agents. But these are examples of high-volume call centers where many incoming issues will be very similar and therefore very likely to be automated effectively. Complex B2B sales will likely prove more difficult and require more sophisticated orchestration.

Phil cautions: Using pre-existing deterministic automations gives AI agents a head start in reliably completing tasks, but when we do that, we miss an opportunity to rethink processes . He quotes a panelist from the Atlassian event:

We need to question the way we work today. We must not just use this AI to equip an existing process with artificial intelligence, without thinking about it first, is this how we want to work tomorrow?

A valid question that we shouldn’t lose sight of – and one that many startups will likely take as a rallying cry. Let’s see how they do.

Diginomica choice – my top stories on diginomica this week

Supplier analysis, diginomica style. Income Reports to Note:

Other major supplier stories:

• “I can’t imagine a better way to live my life” – Larry Ellison explains why Oracle is on a mission to transform healthcare – We didn’t hear much about this topic at CloudWorld’s keynote, but Oracle is now back on its ambitious healthcare pitches. Stuart is on the case, with more context from Safra Catz as well: Why isn’t access to healthcare as easy as shopping? It will soon, says Oracle’s Safra Catz. I view health care data problems as difficult to the point of intractability. Still, I can think of many less pressing issues that vendors have raised – so here you go.
• IFS Unleashed ’24 – “You need an AI business plan,” says IFS CEO – Phil reviews the IFS Industrial AI approach and explains why on-premises systems have some limitations in their pace of AI adoption. (Yes, there is AI readiness – what it looks like and the tools/strategies to get there – should be a major priority for businesses over the next year).
• Leading orchestration to the next level – Camunda swings the baton – Martin explores another angle on end-to-end process orchestration, with AI in the mix.
• Happy Holidays to Agent Technology for Retailers! Salesforce’s Michael Affronti Makes His Point – Stuart on Salesforce’s AI agent plans for retailers and how they could change the holiday season. This should be interesting…

A new batch of customer use cases from Diginomica:

Jon’s handbag – Chris asks the tough (but right) questions again in Opinion – When it comes to AI policy, why is the UK desperate to appease authoritarian providers? Mark Chillingworth reports from a major industry forum in Get Ready for AI Talks and the AI ​​Factory (two big takeaways: the need for AI talent and the vital role of consistent metadata for digital and AI services). Finally, Stuart returns to one of his old favorites leaves hobby: analyzing Mark Zuckerberg. But this time, with a surprising twist:

Regular readers will know that I’m not inherently a big fan of what Meta CEO Mark Zuckerberg has had to say over the years, but credit where credit is due – while short-termists of Wall Street are stepping up their “show us.” “Money is screaming around AI infrastructure spending, Zuckerberg is having none of it.”

Indeed. And:

A message that more sellers could usefully reinforce without apologizing to Wall Street.

The best of the business web

My top seven

Puffs

I’m not sure about this potential rebranding:

I guess I should have seen this one coming:

Another product that could use a little more of the “smarts” that Microsoft Windows can spare:

If you find a #ensw piece that qualifies for hits and misses – right or wrong way – let me know in the comments form Clive this is (almost) always the case. Most of the articles on Enterprise’s successes and failures are selected in my selection @jonerpnewsfeed.