The Three Phases of Software
David Friedberg articulated something on this week’s All-In Podcast that every enterprise leader should internalize. He described the evolution of software in three distinct phases:
Phase one: worker productivity enhancement. Software helps people do their work better. Spreadsheets, word processors, databases, CRMs — tools that make humans more efficient at tasks they were already doing.
Phase two: work completion. Agents that complete the work people used to do. Not just assisting, but autonomously executing — drafting documents, processing transactions, resolving service tickets, analyzing data sets.
Phase three: the work humans can’t do. AI that tackles problems too complex, too data-intensive, or too cross-functional for any human team to address. Synthesizing insights across millions of documents in real time. Running continuous multi-system optimization. Executing workflows that span a dozen platforms and require simultaneous coordination.
We’re at the threshold of phase three. And most enterprises are still stuck in the early stages of phase one.
The Records Trap
Enterprises have spent the last thirty years building systems of record. CRMs, ERPs, EMRs, HRIS platforms, data warehouses, document management systems — collectively representing trillions of dollars in investment.
These systems serve a critical function: they remember. They store customer interactions, financial transactions, patient records, inventory levels, employee data, and operational metrics. They are the institutional memory of the enterprise.
But memory without intelligence is just storage.
Ask your CRM how to handle a complex multi-stakeholder deal that mirrors one your team closed successfully three years ago across a different region. It can’t tell you — even though all the data is there. Ask your EMR to synthesize a patient’s full treatment history across multiple providers and recommend a care pathway consistent with the latest clinical guidelines. It won’t — even though it holds the records.
These systems record what happened. They don’t understand what it means, and they certainly can’t decide what to do next.
Copilots: A Half-Step Forward
The first wave of AI in enterprise software was the copilot model. Embed a language model inside the application. Let users ask questions in natural language. Generate summaries, draft responses, surface analytics.
This was genuinely useful — and genuinely limited.
Copilots are reactive. They respond to prompts. They operate within a single application’s context. They scale with human throughput: one person asking questions, one copilot answering. The bottleneck is still the human in the loop.
More fundamentally, copilots don’t act. They suggest. They draft. They recommend. But someone still has to review, approve, copy-paste, switch applications, and execute across systems. The operational complexity remains. The human is still the integration layer.
For enterprises trying to transform — not just optimize — this isn’t enough.
The Agentic Shift
True agentic AI represents a qualitative break from the copilot model. Agents don’t wait for prompts. They execute workflows. They coordinate across system boundaries. They make decisions within governance frameworks. They operate with varying degrees of autonomy depending on the risk and policy context.
Consider the difference:
Copilot approach to claims processing: An analyst opens a claim, asks the AI to summarize relevant policy language, manually cross-references member eligibility, checks provider network status in a separate system, and makes a determination. The AI helped with one step. The analyst still did the work.
Agentic approach: The system ingests the claim, automatically pulls member eligibility from the enrollment system, verifies provider network status, checks medical necessity against clinical guidelines, cross-references fraud detection patterns, applies the relevant policy rules, and either processes the claim automatically (for straightforward cases within policy bounds) or presents a fully prepared recommendation with complete evidence trail for the adjudicator (for complex or edge cases).
The first approach makes humans slightly faster. The second transforms the operation.
The Autonomy Ladder
The shift from copilot to autonomous agent isn’t binary — it’s a progression. Enterprises need a framework for managing this transition:
Level 1 — Assist. The system retrieves relevant information. Humans make all decisions and take all actions. This is where most enterprise AI lives today.
Level 2 — Draft. The system generates recommendations, draft responses, or proposed actions. Humans review and approve before execution. Think auto-drafted emails, suggested claim determinations, or pre-populated service orders.
Level 3 — Approve. The system prepares fully formed actions — complete with evidence, reasoning, and policy justification. Humans approve or reject with a single action. The cognitive load shifts from creation to judgment.
Level 4 — Low-risk autopilot. The system executes autonomously within tight, well-defined policy boundaries. Straightforward claims under a threshold. Routine service ticket resolutions matching documented procedures. Standard compliance checks against clear criteria. Humans are notified but don’t need to intervene.
Level 5 — Policy-bounded autonomy. The system operates with broad autonomy within explicit governance frameworks. Complex multi-step workflows, cross-system coordination, and adaptive decision-making — all within auditable, enforceable policy constraints.
The key insight: this ladder isn’t about replacing humans. It’s about expanding what the enterprise can accomplish. At each level, the organization’s capacity grows — not by adding headcount, but by extending the reach of governed intelligence.
Doing the Impossible
This is where Friedberg’s third phase becomes transformative. When you combine semantic understanding across an enterprise’s full knowledge base with autonomous agents that can execute governed workflows across systems, you don’t just do existing work faster. You unlock entirely new categories of work.
A pharmaceutical company doesn’t just accelerate drug safety reviews. It runs continuous, real-time surveillance across millions of adverse event reports, published literature, and clinical trial data simultaneously — something no human team could sustain.
A healthcare payer doesn’t just process claims faster. It identifies emerging fraud patterns across its entire book of business in real time, correlating provider billing behavior, member utilization patterns, and network anomalies that would take human analysts months to surface.
A field service organization doesn’t just resolve issues quicker. It predicts equipment failures by reasoning across service history, sensor data, environmental conditions, and manufacturer specifications — preventing downtime before it occurs.
These aren’t incremental improvements. They’re capabilities that simply didn’t exist before. The sum of today’s software market, as Friedberg suggested, could grow four to ten times in five years — but the value will concentrate in platforms that enable these new categories of work.
Building Systems of Action
The transition from systems of record to systems of action requires three foundational capabilities:
Semantic fusion fabric. Not flat document retrieval, but structured, ontology-grounded understanding of your domain. Knowledge organized with provenance, temporal context, and semantic relationships. A fabric that compounds over time as the organization’s knowledge deepens.
Governed execution. Typed action registries, policy gates, idempotent operations, safe write-backs with verification, and rollback patterns. Every action traceable. Every decision auditable. Every workflow operating within explicit governance constraints.
Progressive delegation. The ability to start at Level 1 and systematically advance to higher autonomy as trust is established, policies are validated, and governance frameworks prove reliable. Not a leap of faith — a measured progression.
The enterprises that build these foundations now won’t just be faster. They’ll be capable of things their competitors cannot do at all. That’s the real value migration — not from one software vendor to another, but from systems that remember to systems that act.