Enterprise hits and misses - Bayesian uncertainty makes for better AI, tokenomics gets demystified, and Meta's AI Avocado isn't ripe yet

Lead story - How Bayesian-inspired uncertainty management could shape the future of trustworthy AI 

If/when we ever get to truly trustworthy AI - especially in high stakes fields like healthcare - managing uncertainty will be a core part. And: few scribes write about the edges of enterprise AI better than George. So what's at stake here George? 

Bayesian approaches and foundation models that treat uncertainty as a first-class primitive are helping to mitigate a critical gap in more trustworthy and efficient health care. This stands in stark contrast to the Large Language Model (LLM)-based statistical approaches driving current AI hype and investment. This has important implications for other industries bumping up against the limits of LLMs in handling edge cases and exceptions.

How can this play out in medical settings? George:

The Bayesian specifics came later through an iterative process of asking better questions until the answers became compelling. Once you strip away the mathematics, the Bayesian approach informs a willingness to sit with uncertainty and to frame better questions. Babina says sparked enthusiasm for improving the process she was already familiar with in medical research.

He shares the discoveries of Concr CEO Irina Babina:

In her view, the LLM paradigm runs into a structural problem trying to make sense of this. LLMs struggle when there are questions or uncertainties about the inputs. This can be dangerous in oncology because the data are structurally incomplete, and the stakes of overconfidence can be severe.

Agreed - so what are the benefits of a Bayesian approach?

Babina found three compelling properties of Bayesian approaches: explainability, flexibility, and computational efficiency. With explainability, Bayesian models can tell you not only what they predict but why. It provides a framework for understanding which components in a model contributed to which aspect of the prediction and with what degree of confidence. This also helped researchers trace causal chains through complex, interconnected biological systems. This was critical because scientists don’t fully understand biology. 

I'll add one more: better identification of causal factors, which has implications for decisions across many fields, but in particular, for medical diagnosis and treatment. LLMs struggle with causality, but with Bayesian approaches, you can literally map out dependencies between variables, predicting cause-and-effect, not just correlation (LLMs). How? As Google Gemini helpfully surfaced as I head to deadline, by incorporating expert domain knowledge, counterfactual reasoning (altering causal factors), and "uncovering potential causal structures from observational data."

Causal inference is one of the hardest AI problems. Even humans (that's us) cannot necessarily determine causal factors that contributed to outcomes, even when drawing on randomized control trials or lived experience. It's clear Bayesian Networks have a role to play here, and it's not necessarily in conflict with LLMs - though the two clearly have different strengths. For example: LLMs can be of use generating causal graphs and Bayesian Networks, albeit with plenty of human output review. I look forward to seeing where George takes this research next... I have some ideas, but I'm not issuing any spoilers here. Watch this space. 

Diginomica picks - my top stories on diginomica this week

• CIO Chris Ortega automates Lake Michigan Credit Union - Mark Chillingworth has fresh profiles from the diginomica network: "Ortega’s clarity on the right tools for the right business requirements is right on the money. Also see: PureGym CIO Andy Caddy keeps the sector fit for growth.
• Why AI augmentation and up-skilling are the secrets to boosting staff productivity - Cath parses a notable new report: "Employees matter. Change management matters. Up-skilling matters. And if they aren’t at the heart of your AI implementation, businesses will never achieve the productivity benefits most of them crave."
• What I’d say to me back then - Digital.ai’s Patricia Prince-Taggart says to get into every room, even if it’s uncomfortable being there - Madeline continues her signature series. How's this for Monday am career advice: "I would say being uncomfortable in the room doesn't mean you don't belong." Yes...

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

• Adobe loses its veteran CEO as Wall Street's nervous AI ticks are again on show despite record revenues reported by the firm - Stuart has the latest on the hot seat SaaS experience, a thankless but fascinating time for software executives: "For now, what [Narayen] calls the 'AI-first sort of book of business' has tripled, but there’s a lot more to do, he argued: 'That should be our next billion-dollar business'. But getting there will be someone else's job."
• The cheaper code becomes, the more orchestration is worth, argues Daniel Dines. Yes, he would say that, but it's a thesis that just gave UiPath its first profitable full year! - Alyx on UiPath's profitability milestone: "UiPath closes fiscal year 2026 on a high, with co-Founder and CEO Daniel Dines making a bold case that falling AI development costs are good for the automation market, not a threat to it."
• Oracle turns in a strong Q3 with time left over to debunk the so-called 'SaaSpocalypse' - Stuart on the vendor that put the breaks on the SaaS market spiral. He quotes Oracle President Mike Sicilia on the SaaSpocalypse: "I don't agree with that at all. I do think that AI tools and their coding capabilities would be a threat if we weren't adopting them, but we are and very rapidly."

Standout vendor use cases - we had some standout use cases from vendors this week: 

• From call trees to agentic AI - how Kingfisher is winning back employee time for customers - Derek has a fresh ServiceNow use case:  "McKenna's advice to any ServiceNow customer starting this journey: I would 100 per cent have looked at data and knowledge first."
• Winter event highlight - how Tech Electronics turned cloud ERP into a force for organizational change (and AI readiness). An Acumatica story - I filed a memorable customer story from the road: "Don't force tools on your employees? Chandak is aligned with my own stump speech. I don't care if 'AI First' is a religion now. When you shift from imposing tool use on employees to a culture of experimentation, I like your chances."
• How Create Music Group strikes the right note by placing data at the heart of the music-streaming process - "The move to Airflow 3 will help the organization to ensure that human-in-the-loop operators and real-time orchestration keep its data platform refreshed."
• When agents meet contact centers - two Salesforce Agentforce early adopters communicate the benefits they're seeing - Stuart shares two early-but-interesting Agentforce use case. Promising quote from Savant: "I think the technology is going to empower any employee to be one who can drive revenue..."

A few more vendor picks, without the quotables:

• Genesys takes a deliberate path to autonomous CX with large action models - Alyx
• Burp! Tech Mahindra, AI and the metabolic rate of change - Katy
• Planful's Analyst Assistant tackles Finance workflow friction - but the GENCFO Trends Report 2026 shows the real bottleneck is cultural -  Alyx
• What's next for NetSuite? CEO Evan Goldberg fields the questions at SuiteConnect NYC - Brian

Jon's grab bag - Sarah adds to our use case binge with 'Ello, ello' - how one US communications services firm is connecting to an AI-enabled future for Mid-west communities. Stuart keeps up with the high stakes AI drama in With multi-billions of dollars of revenue at risk, Anthropic sends in the lawyers against Trump 2.0's ban. What next? 

Are the copyright wars entering a different phase? Stuart explores in Shift happens! Could AI vendor sentiment be changing to meet publisher pragmatism halfway when it comes to content theft? My take: unless the LLM vendors want to implode their own models, they should start recognizing that funding fresh human-generated content is part of their game. Finally, Madeline blew a gasket (in a good/necessary way) in Something for the weekend – yes, we do need a radical plan to close the gender gap in tech; the latest 'big idea' from the UK isn't it!

Best of the enterprise web

My top seven

• Where’s tokenomics for the rest of us? - Constellation's Larry Dignan demystifies tokenomics: "Tokens are a cost center for most enterprises and CIOs are looking for on-prem deployments, smaller models, cheaper compute and more AI inference efficiency."
• Sonar Data Reveals Critical "Verification Gap" in AI Coding: 96% Don’t Fully Trust Output, Yet Only 48% Verify It - There are some fascinating debates on AI coding going on right now. I missed this notable AI developer report in January, but it's worth a look now as well: "This explosion in code volume has not necessarily delivered the expected improvements in efficiency. Instead, the study reveals that the surge in output has created a new bottleneck at the verification stage of software development, with more work now required to review code—leading to urgent new challenges regarding the reliability and security of deployed software." Yes, that's a good summary of the contrast between code output speed and downstream risk.
• Rethinking enterprise architecture for the agentic era - Nice to read some sober agentic AI views: "When it comes to integrating agentic AI, the incremental approach is rooted in pragmatism. It sees agentic AI not as a wholesale substitute for existing systems but as a layer of augmentation that can supercharge what already works. Just as the adoption of microservices enabled agile software development without dismantling the enterprise core, the first wave of agentic AI will sit atop legacy systems to extend existing capabilities." My only add: you could certainly have a parallel sandbox track of aggressive experimentation and app building.
• Meta wants its goggles back - Yeah, not the shiniest of news weeks for Meta. In another sign that scaling models is a game of diminishing returns,  Meta pushes AI model 'Avocado' rollout to May or later, NYT reports - because of performance disappointments. (Avocados are tasty but they do have pits). And: Meta planning sweeping layoffs as AI costs mount. Looks like the "year of efficiency" isn't over yet (that was supposed to be 2023). Note: Meta officially denies the layoffs report so we'll see...
• Multi-media picks - After a blistering closing keynote, I cornered Constellation's Esteban Kolsky for some friendly podcast sparring with lav mics: Hashing out a provocative enterprise AI keynote - live at CRM Playaz IRL with Esteban Kolksy - from operationalizing AI to Kolsky's issue with context graph hype. And: an easy-to-follow/balanced YouTube roundup on AI code reviews: Are AI Agents Going to Replace Human Code Reviewers? 

Whiffs

This sounds awesome: 

This, not really:

Frank Scavo asked me if Meta would be changing its name again, with the Metaverse play officially on life support. My reply: 

Looks like someone needs a breather... 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.