Enterprise hits and misses - LLMs need context, enterprise events need analysis, and AWS ruins Monday

Lead story - knowledge graphs, digital twins, and better AI context

This week, we got a closer look at how knowledge graphs can shake things up with AI and digital twins, and provide better accuracy and context. In Why better knowledge graphs are essential for trustworthy autonomy. Harness CEO Jyoti Bansal explains, George makes the case: 

In software delivery, the knowledge graph guides specialist agents in breaking tasks into subtasks that can be trained and tested for reliability, enabling the safe automation of various aspects of complex software delivery processes.

Knowledge graphs can also be "vectorized," drawing on the best of both formats, by making knowledge graphs semantically richer, while retaining the structured relationships of the graph. Graphs-per-customer allow for: 

Bansal stresses that they have focused on building these mappings per customer, and these are not shared across enterprises. This ensures data privacy and security, while also allowing the model to form a more accurate representation of each organization's unique environment, policies, and practices. This model is continuously enriched with new data about deployments, builds, security scans, testing infrastructure, and databases.

As George says, this brings agents closer to enterprise-grade expectations: 

This enables AI agents to conduct semantic searches using natural language queries. It also helps ground agents using advanced retrieval augmented generation (RAG) in specific, relevant facts to improve accuracy and reduce hallucinations.

Yes - though I prefer the word "guide" - I would rather save the word "ground" for a later day, when LLM agents have a deeper world logic embedded within. "Ground," to me, implies something more foundational than a context window the agent can choose to utilize, or, in some cases, not - despite the instructions given. But for now, that's semantics, because the approach George describes is a winner. He adds: 

When most people discuss agentic AI these days, the underlying premise seems to be that agents will be able to perform certain tasks or processes more autonomously. Although the agents themselves get the most press, the underlying world model that supports them will prove essential in progressing to higher levels of autonomy. 

When RAG first came out, it felt like a bandaid. But now, as we move into context engineering as a discipline, it seems more robust. These techniques are still, to some extent, a workaround for LLM limitations. But they also take us away from scaling obsessions, and into a more fruitful realm, where smaller models and creative tool/database assembly show promise for enterprise-grade AI expectations. I don't think this is the path to so-called AGI, but it's certainly a path to better enterprise results. 

And as for digital twins, this factors into the architecture also, by creating a "dynamic living digital twin of an organization's entire software development and delivery ecosystem." For more on digital twins, and the opportunities/challenges of a "digital twin for people," check George's Could employee digital twins streamline collaboration? Viven co-founder Ashutosh Garg thinks so. 

Diginomica picks - my top stories on diginomica this week

• AI doesn't replace customer service teams, it empowers them to do things differently, says Zendesk CTO - Phil picks a big talking point from Zendesk's user event, and if you ask me, it's the right one.
• Macy's adds AI to its so-called 'Bold New Chapter'. But will it make a difference? - Stuart rides the Macy's rollercoaster again. But are we going up, or down?
• Cath posted two inspiring stories to commemorate Women in Tech Week, including: How Asia Solnyshkina went from Showbiz Journalist to software development agency CEO. She did the same for Ada Lovelace Day. My pick:  Ada Lovelace Day – five ways to survive and thrive as a woman-of-color in tech.

Vendor analysis, diginomica style. Here's my three top choices from our vendor coverage:

• ServiceNow World Forum - agentic AI needs the right conversation - Mark Chillingworth pushes the agentic conversation forward, with customer views from ServiceNow's London event.
• How UiPath's CTO is building the unsexy infrastructure that makes AI actually work - Alyx on what's next for UiPath's agentic agenda.
• How two organizations are using RISE with SAP as a platform for business transformation - Mark Samuels with a two-fer of SAP RISE use cases. Biggest lesson? Treat this move as a business transformation, not a tech migration...

Dreamforce and Oracle AI World are in the books - and we published a slew of content across keynotes, news analysis, exec interviews and customer use cases. I can't possibly summarize all of that here, but I'll pick one top quote from each show. Meantime, check the Dreamforce 2025 and Oracle AI World 2025 collections for whatever angle suits your fancy. My Dreamforce pull quote: 

People still crave for human interaction. I'm not saying that AI is going to replace humans because now we can spend that time in creating more human contacts where they need it. If you have 10,000, 200,000 leads, the agents can try to convert from cold to warm to hot and in that fashion and the scientific team can spend more time innovating. [Veenu Aishwarya, CEO of Aum Lifetech, from Dreamforce 2025 - the importance of human intelligence in the agentic mix, according to Aum Lifetech - by Stuart Lauchlan]

 My Oracle AI World pull quote: 

The key enterprise pain point Ellison pointed to: powerful AI models—ChatGPT, Claude, Grok, Gemini, Llama—are trained on publicly available data from the internet. That's great for general knowledge and use, but it’s not particularly helpful for solving your specific business problems. Your competitive advantage doesn't come from public data. It comes from your proprietary customer data, your operational data, your industry-specific knowledge... What's needed is contextual intelligence—the ability to provide AI models with enterprise-specific context without compromising data security or requiring massive retraining efforts. [Oracle AI World 25 - Ellison's AI vision for Oracle - robot surgeons meet enterprise data reality - by Derek du Preez].

A few more vendor picks, without the quotables:

• Intercom announces the latest evolution of its AI Customer Service Agent. Just getting started... - Barb
• SuiteWorld 25 - Packer Fastener tightens up inventory control with AI - Sarah
• Canvas 25 – Miro notches up 100 million users as it launches AI collaboration tools - Chris
• Team ’25 Europe – the unglamorous work of wiring trust into Atlassian’s AI strategy - Alyx

Jon's grab bag - What does it really mean to 'rewire how work gets done'? A Wrike epiphany - Ian attended Wrike's recent event, and found a key message via a guest speaker. Meanwhile, Katy has a caution from Cisco on AI infrastructure debt: Cisco is warning of AI Infrastructure debt. Here’s why - and what it means for enterprise buyers. 

Finally, Brian laid down the satirical smack in Something for the weekend - when your resume isn’t working, you need BlatherGPT!!! "BlatherGPT users can rest easy knowing their resume will contain all of today’s hot new skills whether you have them or not!!! It’s incredible!"

Best of the enterprise web

My top seven

• Welcome to the context chorus: There’s no AI without context - Speaking of context... As I rack up future chiropractor bills slog my way through flight delay purgatory on the way to Constellation's Connected Enterprise event, it seems apropos that Constellation's Larry Dignan would land the top spot this week: "Exner said LLMs have to be grounded with the right data and context to be accurate. Agentic AI only ups the ante. "Now that an AI agent is performing an action, it's doing a task and potentially doing it badly," said Exner. "The consequence of not grounding that LLM or that agent, not giving it the right context could be disruptive and damaging." Yeah, I'm not going to trash talk on the use of 'grounding' again this week.... But maybe next time.
• Agentic AI’s OODA Loop Problem - Schneier on Security has another LLM security hot potato for agentic teams everywhere (and the execs who want to hurry up that AI project because...). How about this for a wake up call: "Systems that allow AI to use large corpora of content, such as Retrieval Augmented Generation can ingest poisoned documents. Tool-calling application programming interfaces can execute untrusted code... That means that fixing AI hallucination is insufficient because even if the AI accurately interprets its inputs and produces corresponding output, it can be fully corrupt."
• Deploying agentic AI with safety and security: A playbook for technology leaders - McKinsey is pretty jazzed about agentic AI, but credit for balancing that with risk factors in this playbook.
• SAP Connect 2025: Unpacking CX, AI, and Does Cinderella Finally Get to Dance? - CX analyst and long time SAP watcher Thomas Wieberneit brings those two themes together in his SAP Connect analysis.
• Anthropic, IBM and the Future of the Enterprise AI Market – RedMonk's Steve O'Grady asks why IBM partnered with Anthropic, given IBM's own significant AI investments - and he answers.
• ChatGPT sales are flat — so Altman offers OpenAI adult chat – Yeah, arguably this could go in the whiffs section, but I thought this line, credited to Jack Klompus on Bluesky, was the sharpiest/scariest quote of the week: "The entire economy now rests on a porn pivot by a piracy company." Oh geez let's hope not...
• The Oracle AI World 2025 podcast review - as per tradition, Brian Sommer and I claimed an unused meeting room, and hashed out Oracle's AI vision versus customer realities (I did the same thing with Josh Greenbaum at the end of SAP Connect: SAP Connect 2025 - the end-to-end podcast review). And that wraps my Vegas residency...

Whiffs

So, AWS ruined everyone's Monday... should we be glad or bothered it was a trivial reason (DNS?) compared to a coordinated hostile hack? 

Microsoft continues its hot streak in the whiffs section of late... 

Not the best AI- generated advice I have seen lately: 

See you next time... If you find an #ensw piece that qualifies for hits and misses - in a good or bad way - let me know in the comments as Clive (almost) always does. Most Enterprise hits and misses articles are selected from my curated @jonerpnewsfeed.