Tasks Accelerate

I · The Curve We Just Watched

Three years ago, AI in software engineering meant tab complete. GitHub Copilot suggested the next line. Engineers shrugged, accepted maybe a third of the suggestions, and moved on. It was useful. It was not transformative.

At Google Cloud Next 2026 last month, Sundar Pichai disclosed that 75% of all new code at Google is now AI-generated and approved by engineers, up from 50% last fall and roughly 25% in October 2024. In twelve months, the number tripled. In thirty-six months, software engineering went from a discipline where AI helped with autocomplete to one where AI writes most of the first draft.

That curve compounded faster than almost any enterprise technology shift in living memory. Faster than cloud. Faster than mobile. Faster than the original SaaS wave.

Now every CIO is being asked the same question by their board: when does this happen to the rest of the work?

The honest answer is: not on this curve.

II · The Temptation

The temptation is to extrapolate. If 2023's autocomplete reached roughly one-third acceptance and Google is at 75% generation in 2026, the obvious move is to assume enterprise agents for sales, support, finance, and operations are on a similar arc. Three years to workflow automation. Two more to autonomous operations. By 2030, the back office runs itself.

The pitch is everywhere. Salesforce. ServiceNow. SAP. Workday. Adobe. Every major ISV has shipped agents in the last twelve months. The hyperscalers are selling agentic platforms. Boards want a date.

The answer requires unpacking why code accelerated the way it did. Because the properties that made code generation compound are not the properties most enterprise work has.

III · The Verifiability Gap

Code has one property most knowledge work doesn't: it tells you when it's wrong.

A unit test passes or fails. A build compiles or breaks. A type checks or it doesn't. A function returns the expected value or it raises. Engineers operate inside one of the most verifiable feedback loops in any profession. The system tells them, within seconds, whether their work is correct.

This is what made the curve compound. Models generated. Engineers reviewed. Tests caught the wrong answers. The wrong answers became training data. The right answers became the new floor. Repeat, every few months, for three years.

When an agent drafts an email to a customer, what's the test? When it reconciles a financial entry, what's the type check? When it triages a support ticket and resolves it, what's the assertion that says correct?

The right answer depends on the customer's history, the regulatory environment, the brand voice, the policy of the day, the relationship the account manager is trying to repair. The feedback loop runs in churn, NPS, audit findings, revenue impact. Sometimes the feedback never arrives. The customer just quietly leaves. The regulator just quietly files a note.

This is the verifiability gap. It shapes how the agentic curve will move.

IV · Five Levels of Autonomy

The industry is starting to use a Society of Automotive Engineers analog for agentic adoption. Five levels, from assisted to fully autonomous, with humans handing over progressively more of the loop.

Most enterprises with AI in production are at L1. A few are doing real L2. L3 is rare and usually inside a narrow domain: IT service desk, commodity support, expense reimbursement. L4 is essentially aspirational in regulated industries. L5 is, for now, marketing copy.

The reason this matters: each level requires governance, integration, and trust architecture that the level below doesn't. The jump from L2 to L3 isn't an incremental model upgrade. It's a re-architecture of who owns the workflow, who owns the data, who owns the customer relationship when something breaks. That re-architecture takes years, not months. And it doesn't compound the way code generation did, because the bottleneck isn't the model.

V · Everyone Has Agents Now

The ISV response over the last twelve months has been to ship agents inside the existing product. Agentforce. Joule. Now Assist. Workday's AI Agents. Adobe's GenStudio.

This solves a real problem. Customers wanted AI inside the systems they already use. But it creates a new one. If every ISV has an agent, and agents will increasingly call each other across systems, where does the work actually live?

The current answer at most enterprises is: nowhere coherent. Sales agents inside Salesforce. Support agents inside ServiceNow. Finance agents inside SAP. Marketing agents inside Adobe. HR agents inside Workday. Each one trained on its system's data, blind to the others, governed by its vendor's policy layer.

The architectural question follows: who owns the customer journey when six vendors are each running an agent inside it?

The next move, already underway, is headless primitives. Strip the UI off the SaaS product. Expose actions, data, and workflow as API surface. Let agents orchestrate across them, yours or someone else's. That's a different defensibility game than the SaaS era.

Most ISVs aren't designed for that game. Their data models, their permission systems, their billing all assume a human at the screen. The ones that win the next decade will be the ones that rebuild from the API surface up. The rest will become commodity execution layers underneath someone else's orchestration.

VI · Tasks Accelerate. Jobs Don't.

Here is the part the enterprise AI conversation usually misses.

Code generation accelerated tasks. The job of software engineering looks essentially the same in 2026 as it did in 2023. Figuring out what to build. How it should fit together. What the user actually needs. How the system fails. Who owns the operational consequences when it does.

What changed is the implementation cost. A two-day implementation now takes two hours. A junior engineer producing what a senior would have produced before. The leverage moved. The thinking didn't.

The same will be true of enterprise agentic work. Drafting an email is a task. Designing the customer service operation is a job. Resolving a ticket is a task. Deciding the support model, the escalation policy, the brand voice, the integration architecture, the governance layer. That is the job.

Tasks compound on a fast curve when verification is cheap. Jobs compound on a slow curve because the thinking is the constraint, not the implementation.

This is why "seventy-five percent of enterprise work will be AI in three years" gets the shape of the curve wrong. The right shape: a lot of tasks accelerate quickly inside well-bounded workflows. The jobs around those tasks don't get faster. Design, governance, integration, accountability. In many cases they get harder, because the surface area of what to think about expanded.

VII · What This Means If You're Buying

For enterprise leaders sitting in front of agent demos this year, three implications.

The agentic curve is real. The compounding will happen. But the path is governance, integration, and trust before scale, not the other way around. The enterprises that get this right will move slower than the headlines and faster than their peers.

The ones that get it wrong will have shipped ten agents and produced no business outcome.

That's the real curve worth watching.