There is a new concept emerging at the frontier of enterprise AI, and it is worth naming clearly: Loop Engineering.
Loop Engineering is the discipline of designing the feedback loop between human judgment and agent execution — the architecture that determines where AI acts autonomously, where it routes to a human, and how every decision made in that loop makes the system smarter over time. It is not AI strategy in the abstract. It is the concrete operational design of how a business runs when agents handle the volume and humans hold the judgment.
Nick Milani wrote a great piece about this from a software development perspective — The Autonomous Software Factory — and I'd encourage every business leader to read it. Nick describes precisely what is happening to software engineering right now: AI agents writing code, running tests, deploying to production through the night, while humans set direction and approve at key gates. His perspective is sharp, and the maturity model he outlines maps almost perfectly onto what is happening across business operations.
The software factory is being built for engineering teams. The business factory — built on the same Loop Engineering principles — is what comes next for every other function in the enterprise.
The Three Waves of Enterprise AI
Wave 1: Human with AI Assistant
The first wave is where almost every enterprise started. Analysts ask ChatGPT to summarize a contract. Executives use Claude to draft a communication. Teams use Gemini to research a question they would have spent time Googling.
We still get asked regularly whether Rollio can help draft an email. And yes — we can. But the more interesting question is whether that email needs to exist at all.
In many cases, the process that would have generated that email — a supplier notification, a payment confirmation, an exception alert — can be handled entirely between agents. Agent-to-agent communication. In some cases, agent-to-agent negotiation. No human in the transaction at all. That is a fundamentally different category of thinking than "draft my email for me" — and it points directly toward what the frontier firm looks like.
In Wave 1, that insight is not yet visible. The AI is a tool. The human does all the work.
Wave 2: Human-Agent Teams
The second wave is where most enterprises operate today. AI now executes — but only when a human triggers it. Someone opens the tool, prompts the action, reviews the output, and decides what happens next. Copilot processing your inbox. Claude reviewing a document. Gemini assisting with a task. Real productivity gains follow: teams move faster, mechanical work per person decreases, junior staff punch above their weight.
But the AI waits to be asked. Every workflow still starts and ends with a human prompt. The structure of operations — who touches what, in what sequence, through what approval chain — is identical to the pre-AI version. The team is more productive, but not more scalable.
Most organizations that believe they've completed their AI transformation are in Wave 2. They've installed better tools. They haven't changed how the business runs.
Wave 3: The Frontier Firm
This is the stage Nick describes as the Software Factory — and its business equivalent is what we call the Frontier Firm.
AI agents read incoming exceptions, analyze them in full context, determine the right path, route to the appropriate decision-maker or resolve autonomously, execute the resulting action back into the system of record, and document every decision in an auditable trail. Humans set the rules. Agents run the process. The business operates continuously, at a scale no human team can match.
The loop — agent execution generating signal, signal improving the system, the system executing better — is the moat.
What This Looks Like in Practice
Quality & Claims: A Global Manufacturing Customer
One of our manufacturing customers managed quality claims the way most manufacturers do: emails to track down context, separate systems to find history, no single place where claim status and decisions were visible to everyone involved. The process was slow, siloed, and dependent on whoever happened to be available.
The frontier firm model deploys an AI agent across three tiers:
Tier 1 — Auto-resolve. The exception falls within policy parameters. The agent validates, executes the resolution, and logs the decision. No human involved.
Tier 2 — Expert review. The exception requires judgment. The agent assembles all relevant context — claim history, product specifications, prior decisions on similar cases — and routes it to the right specialist with a recommended action.
Tier 3 — Multi-team orchestration. The exception requires multiple teams. The agent briefs all relevant parties simultaneously with full case context, facilitates the joint decision, and executes the resolution.
Institutional knowledge that lived in individual inboxes is now centralized, transparent, and auditable. Claims resolve in hours, not days. The process runs 24/7.
Procurement: Supplier Price Changes at a Major US Utilities Company
For procurement teams managing supplier catalogs with hundreds of thousands of line items, price changes are one of the most operationally painful processes in enterprise operations. The traditional workflow: submissions land in a tracking spreadsheet, buyers manually review each line, approval workflows begin, negotiations happen over email, and then — after all of that — every accepted change gets re-entered into the ERP line by line. Teams spend the majority of their time moving data from one place to another. Not analyzing. Not negotiating. Data entry.
The frontier firm model routes each price change based on deviation and context: auto-approve within tolerance, Commodity Manager for strategic category items, Buyer or Buyer+Manager for deviations within authority ranges, and Buyer+Supplier facilitated negotiation for contested items. Each party receives the context relevant to their role. When a resolution is reached, the agent executes the agreed price directly into the ERP. No re-entry step. The spreadsheet disappears because the agent is now the process. Buyers focus on negotiation and strategy.
Why Wave 2 Is Not the Destination
You cannot scale a Wave 2 operation without scaling headcount proportionally. If your business doubles, you need roughly twice as many people working with AI tools. You have improved individual productivity — you have not changed the ratio of humans to work volume. Operational cost still scales with business volume.
A Wave 3 operation breaks that relationship. When exceptions double, the agent pipeline processes them. When price changes spike at quarter-end, the procurement agent works through the queue. Headcount scales with strategic complexity, not transaction volume.
Wave 2 does not build institutional intelligence. Satya Nadella's concept of token capital names the difference precisely. Token capital is the proprietary AI capability a firm builds through its own workflows, data, and embedded institutional knowledge — the accumulation of every agent execution, every routed decision, every logged outcome. Wave 2 companies use the same models, the same tools, and the same generic automations as their competitors. Nothing compounds. No moat forms. Wave 3 companies build token capital. The gap grows every week.
What Happens to the People
This is the question most business leaders do not ask clearly enough during AI transformation — and it deserves a direct answer.
The frontier firm does not eliminate roles. It reorganizes them.
Operations and process managers move from managing queues and escalations to governing the agent layer. Their operational expertise becomes the policy layer — they define which decisions agents resolve autonomously, where escalation triggers, and what the human gate looks like. This is higher-leverage work. The people who understand the process deeply become the architects of how the agent runs it.
Finance and AP/AR teams shift from exception chasing to exception governance. When agents handle routing, matching, and reconciliation, the finance team focuses on pattern-level decisions: refining approval thresholds, reviewing anomalies, and handling the genuinely complex cases where business judgment has real commercial value.
Procurement and buying teams move from data movement to decision-making. When agents handle line-item routing, ERP re-entry, and approval tracking, buyers apply their expertise where it actually matters: supplier relationships, negotiation strategy, and category planning.
Compliance, risk, and audit become both more important and better supported. More important because agents acting at enterprise scale generate compliance exposure at enterprise scale. Better supported because every agent action generates a complete, automatic audit trail — often more thorough than manually documented processes. The compliance function shifts from chasing documentation to governing the rules that produce it automatically.
IT and business systems leaders find the integration layer becoming strategic. Agents that can read, contextualize, and write back to systems of record are only as good as the data infrastructure beneath them. The leaders who understand both the business process and the agent architecture become more central to operations, not less.
Executive and operational leadership face a different kind of challenge: managing a hybrid organization of humans and agents requires a different operating model. Capacity planning looks different when your operation includes agent pipelines. Leaders who understand where human judgment gates belong will define how the frontier firm operates. Leaders who don't will find their AI strategy is still Wave 2 regardless of the technology they've deployed.
The through-line: the mechanical layer moves to agents. The judgment layer moves up the stack. The people who thrive are the ones who understood the process deeply enough to define how the agent runs it.
What the Transition Requires
Moving from Wave 2 to Wave 3 is not primarily a technology problem. The technology exists. It is an architecture and operating model problem.
Embedded business rules, not prompt-based ones. Rules in prompts are brittle — they drift, they cannot be audited, they do not compound. Rules embedded in an execution layer are durable: every decision applies them consistently, every exception sharpens them, and every audit can trace back to them.
Contextualized business data, not just data access. An agent that can read your ERP is not the same as an agent that understands it. Knowing a field value is different from knowing what it means in the context of a supplier relationship, an approval history, and an open dispute. The context layer is what makes agents capable of judgment-adjacent routing rather than simple rule-matching.
Defined human gates, not humans in every loop. The human-agent boundary needs to be explicit. Which decisions auto-execute? Which require expert review? Which require multi-team orchestration? Answering these questions per process is the design work of Loop Engineering. Without it, you end up with agents that require human confirmation at every step — which is slower than Wave 2, not faster.
Agent-level governance and auditability. When multiple agents are executing across processes simultaneously, you need to know which agent did what, when, and why. Not just for compliance — but for continuous improvement. When an agent resolves an exception incorrectly, the trace needs to identify which rule drove that decision, which agent applied it, and what data it was working from. Governance of the agent layer is not a compliance checkbox — it is the operational infrastructure that lets you trust, improve, and scale the system over time.
Frequently Asked Questions
Q: Is the frontier firm model practical for mid-market businesses, or only large enterprises? The operational patterns — three-tier routing, autonomous resolution, context-aware escalation — are process-agnostic. The complexity of the deployment scales with the complexity of the business. A mid-market procurement team can deploy a Wave 3 model for price change management as effectively as a large enterprise. The architecture is the same; the scope is smaller.
Q: How do we know which processes to automate first? Start with the highest-volume mechanical process — the one where your team spends the most time routing, re-entering data, or waiting for information that lives in another system. The test: if a new hire could execute the decision after reading a one-page policy document, an agent can execute it automatically. Start there, measure the before and after, and expand from that foundation.
Q: What happens when the agent makes the wrong call? Every agent execution generates an auditable record: what it saw, what decision it made, what action it took. When a wrong call happens, the trace is immediate. The policy layer that drove the decision can be corrected precisely — because the rule is embedded in the execution layer, not reconstructed from a prompt. Wrong calls also become the signal that sharpens the system over time.
Q: How does Loop Engineering differ from standard process automation? Traditional process automation executes fixed rules on structured data. Loop Engineering adds three things: context awareness across unstructured data, dynamic routing based on exception analysis, and an audit trail that generates continuous signal for improvement. The loop is what makes it compound. Traditional automation runs the same process repeatedly. Loop Engineering makes the process smarter with every run.
Nick closed his piece with a question worth adopting directly: the window in which you can help design the factory — rather than show up after it is already running — is finite.
The same is true for business leaders. The business factory is being built right now. The discipline of Loop Engineering — designing the feedback loop between human judgment and agent execution — is the work that separates Wave 2 from Wave 3.
Human-led. Agent operated.
See how Rollio deploys this model across enterprise operations → or talk to our team about your process.
