Getting real about agentic AI, from protocols to data platforms - more inbox adventures with AI vendors

Last time around, I revealed how valiant PR pros braved my inbox filters, to deliver lessons on getting RAG right.

But now that we have fresh ideas on RAG (Retrieval Augmented Generation), it's time to move into agentic AI - and how the focus on context engineering can move us to better AI results. What is context engineering? In the shortest version: 

The art of providing all the context for the task to be plausibly solvable by the LLM.

Or: 

Context engineering is the delicate art and science of filling the context window with just the right information for the next step. 

You can see the shift: now we are moving from prompt engineering and even RAG informational context, into a broader canvas: providing the AI agent with whatever it needs to do its job. That raises plenty of new questions, eh? 

Once again, PR firms valiantly tried to charm my inbox on these topics - with mixed results. But occasionally, we got somewhere. 

Pitch #1: KNIME makes the case for agentic protocols

For me, RAG-related pitches raise flags about LLM's ignoring context - in favor of the LLM's own training data. With agentic pitches, that problem remains, but now we have a new one: the compound error problem. 

AI agents are not easy to deliver in a reliable manner; that problem builds up across end-to-end workflows, particularly with multi-agent handoffs (see: the study, Why Do Multi-Agent Systems Fail). If you get it right, the use cases are interesting - but that's not a low bar, at least not yet.

Without further ado, I heard from KNIME, an "end to end data science platform." The pitch: 

Hi Jon, the AI world is shifting toward interconnected and interoperable agent ecosystems, but many orgs are stuck with fragmented architectures, siloed data, and custom integrations. New standards like the Model Context Protocol — now backed by players like Anthropic, Microsoft, and Google— signal a new approach where agents can plug into each other without breaking everything else. Michael Berthold, CEO and co-founder of data analytics platform KNIME, argues that standardization (e.g., MCP) is the tipping point for making enterprise AI scalable and trustworthy. Interested in hearing more? 

Agentic protocols are the right way forward. MCP has already proven itself, but more in the context of calling third party tools. While you can technically  "call" other agents with MCP,  Google's A2A protocol is dedicated to that challenge. I view agent-to-agent as important, but not ready for prime time yet. The best way to build trust is with live AI projects that are within tech scope.

My response: "While I welcome MCP, I don't agree it's the tipping point to making AI trustworthy. It helps agents talk to each other, but that's also a can of worms as agents are fallible. MCP doesn't fix that." I also sent KNIME my last piece on agentic workflow evaluation:

I'm trying to get deeper into the issues of accuracy, inference, context, and getting the use cases right so that customers can have success. A big part of it is how to maximize the interaction between probablistic (LLM agents) and deterministic solutions (e.g. automated workflows etc).

At the time, I had deep dives coming out on SAP's foundation model, and an autonomous agents for finance. The PR publicist for KNIME took this and ran with it:

On your note about probabilistic + deterministic interaction - that’s actually where KNIME has been seeing traction lately. Data teams are using visual workflows to encode repeatable, rules-based logic that agents can call as modular tools (striking that balance between flexibility and control). Not a silver bullet, but certainly a way to bring structure to otherwise ambiguous agent behaviors. Would love to keep in touch as your SAP and finance coverage takes shape, especially if there’s a fit for KNIME’s approach in the context of explainability or evaluation.

If helpful, I can send over a short guide from KNIME we’ve been using to frame how some data teams are approaching this probabilistic/deterministic balance in agent systems. Not for coverage, just a quick look at how this is being applied practically.

KNIME has plenty of resources (and videos) for building effective AI agents. One of my big talking points is "AI Readiness": there is much that organizations can/should have in place,  if they want effective AI agents. KNIME has good concepts pertaining to "Transitioning Your Data Team to AI Agent Readiness."

The best part is organizations already possess much of what’s needed to develop AI agents. Centralized data teams - analytics, data science, or business intelligence — have long been building the infrastructure, expertise, and data assets necessary to enable agentic solutions. Rather than creating entirely new departments or starting from scratch, businesses can use existing analytics tools, pipelines, and the extensive expertise their data specialists already hold.

No one has an appetite for mult-year data cleansing projects. But: you don't need a massive data makeover to get going. Via KNIME: 

You don’t need to overhaul your infrastructure or rewrite your processes to benefit from agents. The key is to start small. Identify tasks that are repetitive, structured, and already require data or decision-making. These are perfect candidates.

Pitch #2: better processes lead to better AI - a CX example from SupportNinja 

One of the best pitches I received had nothing to do with AI directly - and yet, in terms of AI readiness, it had everything to do with AI. On a recent YouTube video review of AI news, I stumbled on this apt quote: 

The most successful AI implementations happen at companies that already have excellent operational discipline and mature IT processes. So you've got to fix your operation basics, and that isn't sexy, but it's a  prerequisite for making meaningful AI deployments.

On the CX front, I recently heard from SupportNinja about their mystery shopper report: 

Following up on my note from Friday – let me know if this would be of interest! A top CX leader mystery-shopped top ecommerce brands—and found bots are blowing it right when trust matters most.They shopped from brands ranging from emerging athletic apparel brands to the biggest luxury brands, finding that:

• Even luxury brands lost the sale at the support stage
• One bot took 11 minutes to reply
• Only one small brand got it right

I bolded that last line for a reason: what did this smaller brand get right? Does this brand have clues to offer on how to fuse tech into better processes? 

Soon, I got my hands on the full 2025 CX Outsourcing Report (free with sign up).  Here's what I learned about that upstart brand: 

This brand sent prompt, clear return emails anchored in a sustainability narrative. They didn’t require any additional paperwork and the UPS “Happy Returns” integration made drop-off simple and intuitive. The refund confirmation included exact timing and amounts, reinforcing transparency and trust. And the refund hit as soon as we dropped our shirt off at the Happy Returns Bar.

Why did this brand stand out? 

Compared to other retailers that offered more vague return confirmations, this brand delivered a pro-active, confidence-building post-purchase experience.

Which leads to a potent lesson: 

Our biggest learnings? Premium pricing (or a global footprint) doesn't guarantee a premium customer experience.

The enemy of a "premium customers experience"? The friction of the seemingly mundane: 

The most consistent friction appeared in the unglamorous moments: a failed checkout, a vague return policy, a bot that took too long to reply. These gaps erode trust quietly but quickly.

Now we get our first AI clue: this brand is clearly using AI-powered bots as part of its winning process. But even if I hadn't mentioned a bot, doesn't this company sound like it has tackled its data silos? Otherwise, this brand couldn't deliver that kind of superior experience. That's the ironic thing about today's AI - agentic, machine learning or what have you. It acts as an accelerant, making good processes better - or bad/mechanistic process even worse: 

The smallest brand in our research delivered some of the strongest signals of trust and care. From a smooth, bot-enabled return to immediate refund confirmation, this retailer proved that thoughtful automation, clear communication, and brand-aligned tone can rival (and sometimes surpass) the polish of much larger retailers.

So: we have smart automation - but well-integrated into an overall solid experience. By contrast: 

Meanwhile, even established and luxury brands struggled to meet their own standards. Delayed replies, inconsistent tone, and unclear policies reminded us that CX excellence requires consistency and follow-through.

If companies think AI can fix their unclear policies or brand shortcomings, they are wrong (yes, AI might be able to help with reply delays and tone issues, but not without an underlying data structure, including consistent messaging content on brand narratives. That's what AI's "consistent" tone needs to derive from). 

My take - time to learn more about deterministic architectures (and AI agents)

My inbox doesn't stop - the hits keeping coming. A few weeks ago, Nekuda managed to push my buttons very effectively: 

This week, a new company—backed by Visa and Amex Ventures, with a round led by Madrona—is emerging from stealth to solve one of the biggest infrastructure gaps in AI-driven commerce: enabling agents to actually complete transactions. While AI agents are getting better at discovery and intent, they still hit a wall at checkout. This team is building the rails that let agents transact autonomously—handling secure credentials, authorization logic, and trust signaling across the payments stack.

Agents autonomously handling commerce transactions? Yes, that pushes enough buttons to make my entire office light up. 

I sent back some firecrackers, including: "There is no such thing as 'Building rails' because probabilistic LLMs make their own decisions at inference time." However, Nekuda came back with some details on an intriguing hybrid architecture - one that includes deterministic components, including a software-based authorization engine. Once again, smart responses to my barbs worked out: we're talking in a couple of weeks. Stay tuned...