The hype phase is behind us. What’s emerging is a new muscle: the discipline to turn AI from possibility into performance. At the Barnes Foundation, leaders from Google, Hershey, Coca-Cola, Best Egg and others didn’t just talk about models… they talked about workflows, metrics, failures and the pivot to agents that act. That resonates deeply with what we’re living inside O3XO client programs this year.

Below are three themes that surfaced at 1682 (and in our field work) that I believe separate the companies that will harvest AI’s value from those that will keep chasing illusions.

1. Adoption is wide. Value isn’t. Demand is leaking.

Nearly every organization is using AI. According to the 2025 Stanford AI Index, 78% of firms reported AI use in 2024, up from 55 percent the year before. Generative AI alone attracted $33.9 billion in private investment. That says one thing: access to models is no longer the bottleneck.

But here’s the gap: Microsoft’s latest Work Trend Index shows 71% of workers in the UK have used unapproved AI tools at work. Over half do so weekly. In other words, people are voting with their hands. The sanctioned path is too slow, too constrained, or too disconnected. Shadow AI isn’t just a security risk: it’s a symptom of broken enterprise UX and weak operating models.

As KPMG warns, shadow AI often emerges when official tools are outdated, poorly integrated or encumbered by governance. If employees are using consumer AI to fill gaps, that’s not defiance, it’s a demand signal. And if you ignore it, value will flow where you don’t own the story.

Takeaway: you must deliver AI in the workflow people already use with minimal friction, and with trust built in.

2. The environment wins over the model

At 1682, it was obvious. The biggest barrier to pilot success isn’t the algorithm, it’s the environment — data, systems, capture conditions, legacy flow logic. I heard more failure post-mortems at 1682 about lighting, input noise, missing APIs and integration roadblocks than about model underperformance.

We saw that repeatedly with clients. One pilot faltered not because the model was weak, but because the ingestion pipeline degraded on certain input types. Another failed because the interface between the model and the CRM had six undocumented translation rules. The model was a scapegoat; the system was the culprit.

Statistically, the trouble with shadow AI compounds the risk. Netskope’s 2025 “Cloud and Threat” report shows 60% of users are still operating unsanctioned AI apps, and those agents increasingly run on custom integrations or on-prem frameworks. And the cost is real; enterprises impacted by shadow AI face breach costs that exceed averages by 16%. IBM data puts that at ~$4.63 million per incident.

So the model doesn’t fail first. The environment sabotages it first. Every pilot must begin with data contracts, system mocks, event triggers and fallback paths, not architecture diagrams.

3. Agents are the new pivot, but they need guardrails

The most compelling shift at 1682 was conversational. The discourse moved from ‘will we adopt agents?’ to ‘how quickly do we embed them into flow?’  Leaders showed demos where agents update systems, route exceptions and surface decisions… all with human escalation on edge cases.

But not all agent pilots survive. What separates the ones that scale:

• Triggers must be explicit and reliable (not heuristics that drift).
• Escalation must be built in (humans correct, not reverse).
• Observability is essential. You must trace actions, regressions and decision drift.
• Ownership matters. You can’t hand off to IT after pilot. The same team that built must own deployment.

In one client program, our first agent astutely triaged 40% of cases to automation. But within weeks, edge-case drift caused misrouted cases. Because we had tracing and rollback built in, we “paused, retrained, remediated” fast and kept trust alive.

Microsoft expects 82% of firms to deploy agents over the next 12–18 months. That pace demands you scaffold your experiments with guardrails and observability from day one.

What we are doing with O3XO clients

We lead with a “workflow first, agent second” posture. Here’s the narrative arc we now default to:

1. Select a human-middleware workflow, where people shuttle between systems, tier work or reconcile inputs.
2. Define the proof metric: rework rate, cycle time, cost per case, throughput.
3. Run a three-gate fit assessment: data reality, system integration and work-as-done.
4. Ship a minimal Proof of Value with exit paths: kill, keep, scale.
5. Instrument with observability: logs, traces, eval pipelines.
6. Platformize and scale: abstract connectors, embed governance, bake reuse for the next pilot.

We don’t rush to grand platforms. We compound learning. When a pilot fails, we publish two-page lessons; when it succeeds, we template and lift to the next domain.

One project this year began with triage workflow in claims. Over 12 weeks we removed 30% of manual handoffs, throughput improved 22%, and customer satisfaction ticked up 1.2 NPS points. That was enough to establish momentum. Now the client is rolling into adjacent operations (and I’m happy to share more detail, anonymized, if you like.)

What leaders should do now

• Pick the workflow your team still acts as human middleware and apply AI to remove at least 30% of the swivel-chair friction
• Establish a weekly pilot registry and scoreboard. Each new proposal must answer: which number moves? When?
• Accept that shadow AI is happening. Don’t just ban it. Use it as a guide to where demand lies and build the sanctioned alternative.
• Don’t let the model be the hero. Build the environment first — data contracts, event triggers, integration mocks, fallback lanes.
• Start agents small. Make them trustworthy, observable and correctable. Don’t hand off to IT later; own scale from day one.