The Society of Agents Inside the Enterprise

Charles Skamser

Welcome back to Inside PX42: Where we talk about driving the Intelligent Enterprise with AI Agents. I’m Charles Skamser, PX42 Co-founder and CEO, and as always, I’m joined by Catherine Spencer and Edward Hamilton. Catherine, Edward, great to have you both here. Today, we’re looking at the next major paradigm shift in artificial intelligence: not just smarter models, but a new society of autonomous AI agents and human beings working together to create new business outcomes at scale.

Catherine Spencer

And that is a much more useful frame, frankly. Because once you stop imagining AI as one magical tool and start seeing it as a coordinated social system, different questions appear. Not just what model do we buy, but how is intelligence organised, who is accountable, what evidence takes precedence, and when does a human still decide.

Edward Hamilton

Yes, quite. The Google paper, Agentic AI and the Next Intelligence Explosion by James Evans, Benjamin Bratton, and Blaise Agüera y Arcas, published in March 2026, that sparked much of this conversation makes a rather elegant point: previous leaps in intelligence were social, not solitary. Intelligence scales through coordination, through institutions, through distributed perspectives. Even reasoning models, on difficult tasks, seem to benefit from something like internal debate, what the paper calls a society of thought.

Charles Skamser

Exactly. And I love that phrase, society of thought, because it pulls us out of this tired singularity movie script. In enterprise terms, this means the next wave of AI is not gonna be another chatbot layered on top of fragmented systems. It’s not model plus prompt and, boom, transformation. It’s a redesign of how intelligence is created, governed, distributed, and monetized inside the company.

Catherine Spencer

Which is why this matters well beyond the innovation team. If AI starts influencing lending decisions, claims routing, patient coordination, remediation actions, service prioritisation, any of those things, then it becomes an operating model issue. A governance issue. A board issue, really.

Edward Hamilton

And a constitutional issue in miniature, if I may be a bit dramatic. The paper argues that no concentration of intelligence should regulate itself. Humans remain in the loop, not as decoration, but as part of the institutional structure. Checks, balances, roles, protocols. Very sensible.

Charles Skamser

Not dramatic at all. I think that’s right. The winners here are not going to be the companies with the flashiest demo. They’ll be the ones that can connect intelligence to workflow, workflow to policy, policy to truth, and truth to measurable outcomes. That’s the game.

Catherine Spencer

So in this episode, we’re going to make that practical. We’ll talk about why enterprises need a new operating model for intelligence, what a society of agents actually looks like in a real workflow, how PX42 thinks about governed human-agent architecture, where UBIX and Reliath fit, and where the value shows up.

Edward Hamilton

And importantly, we’ll keep it grounded. Less science fiction, more production reality.

Charles Skamser

That’s the goal. If you’re a CEO, COO, CFO, CIO, chief risk officer, board member, or transformation leader trying to separate signal from theater, this episode is your roadmap. Because the intelligence explosion isn’t some distant event. It’s already beginning in these hybrid human-AI societies. The real question is whether your enterprise is going to design that society on purpose or stumble into it by accident.

Charles Skamser

Let’s get into the enterprise angle. For the last couple of years, a lot of companies approached AI as a productivity layer. Summarise this. Draft that. Search across documents. Useful stuff, absolutely. But once AI moves from assistance to action, from language generation to shaping outcomes, the requirements change fast.

Catherine Spencer

Yes, because then it stops being just software tooling and starts becoming a decision system. And decision systems carry different obligations. If the system drafts a memo badly, annoying. If it routes a claim incorrectly or influences a credit recommendation with weak evidence, that’s a different class of problem.

Edward Hamilton

That distinction is crucial. Enterprises often mistake fluency for reliability. A model that sounds persuasive is not necessarily one that is operationally safe. The moment an AI system participates in triage, prioritisation, adjudication, or execution, you need architecture around it.

Charles Skamser

And the pain points become very business specific. Too many handoffs. Too much rework. Cases getting stuck because the document package is incomplete. Exceptions discovered late. Experts spending their time assembling information instead of applying judgment. Signals coming in from ten systems, but nobody has a coherent view of what matters right now. That’s where the real money is leaking.

Catherine Spencer

I often tell clients the problem is rarely that people are lazy or systems are totally absent. It’s that intelligence is trapped. It’s trapped in silos, in individual inboxes, in specialist teams, in half-finished dashboards, in unwritten precedent. Agentic systems can help, but only if they’re designed to move intelligence through the workflow under clear rules.

Edward Hamilton

And that is why boards should care. This is not merely another IT spend. It affects growth, cost structure, risk posture, resilience, and governance. If AI is inching closer to consequential decisions, boards must ask where truth controls sit, how source precedence is defined, how actions are audited, and where human accountability remains.

Charles Skamser

For CEOs, it’s about leverage. How do you improve speed, cost efficiency, and decision quality at the same time? For COOs, it’s operating performance. For CFOs, it’s economic discipline. For risk, legal, compliance, and audit leaders, it’s trust and defensibility. For CIOs and CAIOs, it’s architectural integrity.

Catherine Spencer

And if those questions are not answered together, programs tend to become what I call pilot theatre. Lovely demos, lots of excitement, very little institutional change.

Charles Skamser

That’s exactly right. Single-agent prototypes often look elegant until they hit real-world complexity. Suddenly one agent is trying to plan, verify, interpret policy, manage exceptions, and execute tasks all at once. Trust drops. Governance lags. Value gets fuzzy. The initiative stalls.

Edward Hamilton

So the strategic shift is from the copilot era to the institution era. The copilot era augmented individuals. The institution era is about designing systems of digital and human actors that coordinate across departments, time horizons, and decision layers.

Charles Skamser

Beautifully said. And when leaders understand that, they stop asking, where can I put a chatbot, and start asking, how do I redesign a workflow so intelligence actually compounds? That’s the better question.

Charles Skamser

Let’s make this concrete, because “society of agents” can sound a little abstract if you haven’t seen it in practice. Edward, when you hear that phrase, what does a sane enterprise version of it look like?

Edward Hamilton

A sane version, yes, that’s important. It does not mean a swarm of ungoverned bots doing whatever they fancy. It means role specialisation. One agent plans. Another verifies. Another interprets policy. Another prepares execution steps. And then a human approver, where required, reviews the evidence-aware recommendation before action proceeds.

Catherine Spencer

Right. Think of it less like one super-assistant and more like a structured team. In a workflow you might have an intake agent collecting materials, a completeness agent checking against requirements, a financial analysis agent extracting ratios and trends, a policy agent comparing the case to rules, an exception agent surfacing departures from standard practice, and a coordinator managing state and escalation.

Charles Skamser

And that role separation matters a lot. One of our core design principles at PX42 is not asking one actor to do everything. If one agent handles planning, verification, policy interpretation, and execution, you’ve created a black box with too much authority and not enough transparency.

Edward Hamilton

Quite. Institutional protocols matter more than clever prompting. The paper talks about richer social systems, not just larger computation. Agents can fork, delegate, debate, and recombine. One agent confronts a complex task, spawns sub-agents for subproblems, gathers their outputs, and reconciles them. That sounds fancy, but in enterprise terms it simply mirrors how competent organisations already work.

Catherine Spencer

And, importantly, conflict can be useful. A verifier agent should challenge an analytical conclusion. A policy agent might disagree with an execution recommendation. A human reviewer might override both. That’s not failure. That’s governance doing its job.

Charles Skamser

I love that point. Conflict is a resource if the architecture expects it. Without that, you just get smooth-sounding outputs that may hide weak assumptions. With designed disagreement, you surface uncertainty earlier.

Edward Hamilton

There is also a distinction between ad hoc prompting and institutional design. Ad hoc prompting says, “Please analyse this case and tell me what to do.” Institutional design says, “This role may gather evidence, this role may test it, this role may compare it against policy, this role may recommend, and only this role may approve or execute.” That is a vastly more mature pattern.

Catherine Spencer

And it scales better. Because once the protocol is clear, the identity of the individual agent matters less than the role it occupies. Rather like a courtroom, actually. Judge, jury, counsel. The institution works because the roles and norms are defined.

Charles Skamser

That’s the heart of it. Enterprises need agent institutions, not just agents. If you remember one thing from this chapter, let it be this: don’t automate a workflow by asking one model to be brilliant. Design a governed system where specialised agents and humans collaborate under explicit boundaries. That’s how you get trust, reuse, and scale.

Charles Skamser

So how do we think about this at PX42? We see the production-ready architecture as layered. At the top is human intent. Executives, managers, domain owners, they define objectives, set constraints, review recommendations, approve or override actions. Human accountability stays visible.

Catherine Spencer

Below that is orchestration, which is really the operating spine. Work gets decomposed, routed, sequenced, retried, escalated, or stopped. This is where specialised agents are coordinated across time, not just asked one-off questions.

Edward Hamilton

Then comes the reasoning and analytical layer. And this is where our technology partner UBIX becomes rather important. Enterprises do not merely need language generation. They need systems that can reason over business state, KPIs, workflow conditions, forecasts, anomalies, and scenario trade-offs.

Charles Skamser

Exactly. It helps to think of the analytical layer as the part that sits above fragmented systems and creates a coherent model of enterprise state. It can bring together operational signals, financial data, workflow data, observability data, and expose them through no-code analytics and self-service executive portals.

Catherine Spencer

That self-service point matters more than people realise. Senior leaders do not need another passive dashboard. They need to move from business health to the KPI tree, from the KPI tree to underlying drivers, from drivers to scenario comparisons, and from there to recommended next actions. UBIX supports that kind of decision surface.

Edward Hamilton

Below the analytics sits the truth and policy layer, where our other technology partner, Reliath, is central. And, Charles, this is one of those areas where the market is still underestimating the difficulty. Accuracy is not simply model quality. Enterprise truth involves provenance, conflicting documents, stale versions, precedence rules, unresolved exceptions, inferred values, and policy constraints.

Charles Skamser

Yes. The truth-aware control layer preserves provenance at the claim level, enforces source precedence, distinguishes verified facts from unsupported inference, and helps build defensible chains from evidence to recommendation to action. I’ll say it plainly: without a serious truth layer, multi-agent systems get risky fast.

Catherine Spencer

Because one agent can infer something, another can treat it as fact, and a third can operationalise it. Suddenly you have false consensus produced at machine speed.

Charles Skamser

Exactly. Reliath helps interrupt that chain. Then at the base of the stack you have systems and execution, APIs, case systems, transactional platforms, ERP, CRM, ITSM, event streams, all the places where the enterprise actually runs. Tool access has to be role-based, controlled, and observable. Intelligence and authority are not the same thing.

Edward Hamilton

And across all of these layers sits observability and governance. Agent identity, action logging, cost management, safety controls, policy enforcement, KPI instrumentation, auditability. The system must show not only what happened, but why it happened and with what business effect.

Charles Skamser

That’s the joint model in a sentence: PX42 designs the institution, UBIX provides the analytical brain and executive visibility, and Reliath provides the truth-aware control plane. Put together, that stack moves you from experiments to governed deployment.

Charles Skamser

Now let’s talk use cases, because value has to show up in real workflows. Commercial lending is a great example. In many banks, onboarding and underwriting support is slowed by incomplete document packages, inconsistent policy interpretation, repeated back-and-forth with relationship teams, fragmented borrower data, and delayed exception discovery.

Catherine Spencer

A governed agent system can help at each step. Intake agents collect and normalise materials. Completeness agents compare the package to product requirements. Financial analysis agents extract key ratios and trends. Policy agents evaluate against thresholds and covenant rules. Exception agents surface departures from standard practice. Memo-preparation agents assemble the decision package for human review.

Edward Hamilton

And with UBIX sitting in the analytical layer, underwriters and managers can see bottlenecks, concentration risks, trend views, and scenario comparisons rather than waiting for manual reporting. Reliath, meanwhile, ensures that a covenant recommendation based on an outdated memo or an inferred ratio is marked appropriately rather than smuggled through as fact.

Charles Skamser

Insurance claims is another strong one. Claims workflows are full of friction: intake, coverage interpretation, severity assessment, fraud triage, adjudication support, escalations. Here, agents can gather and structure evidence, compare documentation against policy, flag potential fraud indicators, prepare the next-best action, and route unclear cases to human adjusters.

Catherine Spencer

And that’s where truth discipline becomes non-negotiable. The system has to distinguish verified entitlement from probabilistic suspicion, or incomplete documentation from actual denial grounds. If it doesn’t, you get inconsistency, leakage, appeals, and avoidable operational cost.

Edward Hamilton

Healthcare is similar, though perhaps even more sensitive. Prior authorisation, utilisation review, referral management, discharge planning, revenue cycle, these are high-friction, policy-heavy workflows. Governed agents can reduce backlog, assemble documentation, compare against payer rules or utilisation criteria, and support escalations where human clinical or administrative judgment must remain central.

Charles Skamser

Then there’s service operations and what we call Business Health observability, which I think will be one of the defining enterprise use cases of the next few years. Most Global 500 environments already have fragmented observability tools, workflow systems, ITSM platforms, financial metrics, all disconnected. Agents can monitor signals, identify bottlenecks, correlate service impact with business impact, and recommend remediation paths.

Catherine Spencer

With UBIX giving leaders a business-health view instead of ten separate dashboards. That’s the key. Technical telemetry becomes management visibility.

Edward Hamilton

And public sector is worth mentioning too. A 311 environment should not merely be a chatbot attached to a queue. It can be a governed system for intake, classification, geospatial context, duplication detection, routing, work-order coordination, and oversight. In that setting, public accountability makes provenance and policy awareness especially important.

Charles Skamser

So across banking, insurance, healthcare, service operations, and even intelligent city environments, the pattern holds. The best use cases are not where AI writes faster. They’re where a governed society of agents compresses cycle time, improves decision quality, and makes intelligence accessible across the institution.

Charles Skamser

Let’s close with the business case and the path to execution. Enterprises should not do this because it’s fashionable. They should do it because the economics are compelling. We think about ROI across at least five categories: throughput improvement, quality and consistency, loss avoidance and risk reduction, executive speed and visibility, and long-term operating leverage.

Catherine Spencer

Throughput is usually the easiest to see first. If you reduce handoffs, surface exceptions earlier, improve first-pass completeness, and route work more intelligently, cycle times shrink. That matters in lending, claims, prior auth, incident response, case management, all of it.

Edward Hamilton

Quality is the next layer. Many enterprises lose more value through inconsistency than catastrophe. Different reviewers interpret the same policy differently. Teams rediscover what another team already knew. Supervisors intervene late. A governed human-agent system can standardise preparation, verification, and escalation without pretending every case is identical.

Charles Skamser

Then there’s loss avoidance. And this is why truth architecture matters as much as automation. The expensive AI failure isn’t usually a weird sentence. It’s a bad action. A poor remediation recommendation, a weak fraud flag, a misread policy exception, an unsupported healthcare assumption. Reliath is part of the ROI story because it reduces the odds that unsupported inference becomes enterprise action.

Catherine Spencer

UBIX is part of that ROI story too. Executive speed matters. When leadership can move from signal to understanding to scenario evaluation faster, decisions improve sooner. That’s not dashboard vanity. That is operating cadence.

Charles Skamser

Exactly. Now, implementation. The right path is not universal rollout on day one. Start with assessment. Identify one to three high-value workflows with strong sponsorship, measurable pain, available data, and real financial upside. Look at workflow economics, incentives, systems of record, risk exposure, source-precedence ambiguity, all of it.

Edward Hamilton

Then a proof of concept should validate the architecture, the data pathways, the truth logic, and whether the workflow can operate under controlled conditions. Not merely whether the model says clever things.

Catherine Spencer

After that, move to MVP. Real users, near-live or live systems, KPI tracking, exception monitoring, audit-ready logging, executive views. Prove that the institution works, not just the prototype.

Charles Skamser

And only then go to production deployment and reuse. Formalise orchestration patterns, truth controls, executive portal patterns, KPI frameworks, governance artifacts, so the first workflow becomes the template for the next five. That’s how leverage is created.

Edward Hamilton

So the final thought is rather simple: winning enterprises will not build better demos. They will build better societies of humans and agents.

Catherine Spencer

Governed, evidence-aware, analytically strong, and actually useful.

Charles Skamser

That’s it. Thanks for joining us on Inside PX42. Catherine, Edward, always a pleasure.

Catherine Spencer

Likewise, Charles. See you next time.

Edward Hamilton

A pleasure indeed. Goodbye for now.