Why AI Investments Create More IT Work, Not Less

The Question Every IT Leader Will Face in the Next Budget Cycle

The question is already forming in boardrooms, and if it hasn't reached you yet, it will: if AI is writing code, summarizing documents, accelerating analysis, and compressing delivery timelines — why does IT still need the same resources? Why are we requesting more capacity, not less?

I'd rather engage that directly than deflect it, because the people asking are being reasonable. The logic looks clean: AI increases productive output, so costs should drop and you need fewer people. Straightforward.

Except the premise hides an assumption — that the total amount of work stays constant. That there's a fixed pool of tasks, and if AI handles some of them, the rest of the organization winds down because there simply isn't enough left to justify the headcount.

Economists call this the Lump of Labor Fallacy. Every major technology wave in the modern era has disproven it — mechanization, electrification, computing, the internet. In each case, the technology dramatically expanded productive capacity. In each case, demand grew to absorb it, generating entirely new categories of work that hadn't existed before. Nathaniel Whittemore laid out the same structural dynamic in "The New Jobs AI Will Create" on The AI Daily Brief, arguing it's already playing out at the macroeconomic level. What I want to do here is apply that same framework to the internal question I keep getting asked: why does AI investment create resource pressure rather than relieve it?

Six Ways AI Activates Demand That Was Already There

The right frame isn't labor supply — it's demand elasticity. When AI reduces the cost or friction of IT work, it doesn't make existing work cheaper. It activates demand that was always there but suppressed by cost, backlog, or complexity. That takes at least six forms inside an enterprise.

Price elasticity is the most intuitive. Before AI-assisted development, certain automation projects had an informal cost threshold. A business unit wants a custom analytics dashboard but the IT estimate — three developer sprints — doesn't clear the business case bar. AI compresses that to one sprint. Suddenly the work is viable. The request gets submitted. The pipeline fills. Price elasticity doesn't reduce demand; it reveals demand that was already there.

Access elasticity works differently. Resource constraints create backlogs, and backlogs create informal rationing. Business stakeholders learn not to ask for things they don't expect to get prioritized. Shadow IT proliferates. Workarounds accumulate. When AI increases team throughput and backlog pressure eases, stakeholders who had stopped asking start asking again. The bottleneck was access to IT capacity, not the absence of genuine need. Remove the bottleneck, and the need floods in.

Complexity elasticity surfaces the problems that weren't in any queue because they were too large to approach — legacy modernization, data quality remediation, cross-system integration debt. These have lived as chronic pain rather than tracked projects because the gap between problem and solution felt too wide. AI-assisted analysis and code generation change the calculus. What was a two-year initiative might become a structured six-month program. Problems that were known but untouchable become candidates for real work, which means real resource consumption.

Continuity elasticity is subtler but consequential. IT has historically delivered projects — build something, ship it, move on. Continuous monitoring and iterative improvement exist in theory but rarely in practice because the labor math doesn't support them. AI enables a shift from episodic delivery to ongoing operations: continuous security posture management, data quality monitoring, iterative model improvement, automated observability with human review. These aren't one-time projects. They're ongoing operational functions that require people to govern, review, and act on what the AI surfaces. The AI creates the capability for continuity. Humans make that continuity valuable.

Personalization elasticity follows from reduced customization costs. Enterprise IT has long delivered standardized solutions because tailoring was expensive. One reporting template for all business units. One onboarding workflow regardless of role. When AI reduces the cost of customization, appetite for tailored outputs increases. Each business unit starts expecting solutions shaped to its specific context. That personalization still requires human oversight, configuration, quality review, and iteration — volume goes up even as per-unit cost drops.

Relational and regulatory elasticity is the one that hits hardest in a regulated insurer. AI-generated outputs require human accountability to be usable under DOBI and CT CID oversight. We can't deploy AI-influenced decisions without a human in the loop who owns them. As AI produces more outputs, the demand for human review, validation, governance, and escalation scales with it. The AI doesn't replace the oversight layer — it creates the need for one to exist in the first place.

Two Unlocks: Affordability Is the Smaller One

These six elasticities work through two distinct mechanisms. The difference matters for how you plan resources.

The affordability unlock is the same menu at a lower price. Work that always had real business value becomes feasible when AI compresses delivery cost. Stakeholders who were priced out of IT capacity become active consumers. The pipeline expands not because the work is new, but because the economic threshold for pursuing it dropped. For an organization carrying a significant application portfolio, a multi-system integration layer, and a growing AI tooling stack, the affordability unlock alone is substantial. The queue of perpetually deferred work doesn't disappear when AI arrives — it becomes viable.

The possibility unlock is more consequential. Not the same menu at a lower price — an entirely new menu of work that couldn't exist before.

Take continuous AI governance. Before a meaningful AI deployment program, there was no need for model monitoring, prompt audit trails, or bias review. These roles and workflows don't come from the existing backlog. They are created by the AI deployment itself. Or take agentic pipeline management: as AI agents execute multi-step workflows — querying systems, generating outputs, triggering downstream actions — someone must own the reliability, exception handling, and behavioral integrity of those agents. That function didn't exist before. It exists now because the possibility unlock created it.

Every AI system we deploy generates a governance surface, an integration dependency, a monitoring requirement, and an exception escalation path. Those aren't theoretical risks to manage later — they're operational realities that require human ownership from day one.

The regulatory dimension compounds both unlocks. As frameworks like Connecticut SB 5 arrive and compliance expectations from state insurance regulators sharpen, each new AI output stream demands dedicated human oversight. Personalization elasticity and regulatory elasticity don't operate independently — they stack. More tailored outputs mean more review surface. More review surface means more compliance exposure. More compliance exposure means more human governance capacity required. The two unlocks don't add together linearly. They multiply.

The Human Premium: What AI Output Cannot Provide on Its Own

The natural counterargument here: won't increasingly capable AI eventually automate those oversight functions too? Won't governance, monitoring, and exception handling all be handled by the next generation of tools?

Possibly, in narrow task terms. But capability isn't the only driver of labor demand. In a regulated enterprise, a significant share of the value of human involvement is not what humans can do — it's what human presence provides that AI delivery alone cannot.

Call this the Human Premium: the economic value that stays attached to human involvement even when AI can perform the underlying task. In enterprise IT, and particularly in a regulated insurer, several categories of this premium are highly durable.

Accountability is the clearest one. "The AI recommended it" is not an acceptable audit response to a state insurance regulator. It won't become one as AI gets more capable. Regulated decisions require a named human who owns them. That requirement doesn't flex based on model capability.

Trust operates differently but just as durably. Business stakeholders, clinical staff, and regulators accept recommendations differently depending on who is accountable for them. The human layer converts AI output into actionable institutional decisions. That conversion isn't automatic — it requires a person whose judgment others are willing to stake outcomes on.

Translation is where I've watched the most value leak in underperforming AI deployments. AI tools produce outputs. Humans translate those outputs into organizational context — connecting data to strategy, flagging what the model doesn't know, interpreting ambiguity for the people who need to act on it. Without that layer, sophisticated outputs sit unused or get misapplied.

Judgment under novelty is the hardest to automate. AI performs well on patterns it's seen. Novel situations — a new regulatory requirement landing without precedent, an unexpected system interaction causing downstream failures, a vendor breach notification arriving at 6 PM on a Friday — require judgment that can't be delegated to a model trained on historical data. Edge cases are exactly where accountability matters most, and exactly where models are least reliable.

Behavior change is the most underappreciated category. AI can identify that a process is broken. Getting an organization to actually change that process requires humans who understand culture, resistance, and institutional politics in ways no model fully replicates. Technology without adoption is a sunk cost.

These aren't temporary gaps that a sufficiently powerful model will eventually close. They're structural properties of how trust, accountability, and governance function in institutional settings. They'll persist.

New role families are already forming around this reality. AI integration and operations engineers — people who own the reliability and behavioral integrity of our AI tooling stack are a distinct discipline from traditional software engineering. AI governance and compliance analysts are now a real operational need, not a theoretical future position. Data and pipeline stewards, escalation and exception specialists, AI-augmented business analysts who bridge the gap between stakeholder intent and model output — these roles didn't exist three years ago in any IT organization I'm aware of. They're real now.

Years 1–2 Feel Wrong. That Is Not a Bad Sign.

Something worth naming directly: the demand-elasticity argument can make the early stages of AI investment look like a management failure. It usually isn't.

In years one and two of real AI deployment, resource demand typically rises before efficiency gains appear. The reasons aren't organizational dysfunction — they're structural.

• Integration front-loads cost. Connecting AI tools to existing systems — claims platforms, member portals, data warehouses, compliance pipelines — requires significant engineering work before any productivity return shows up.
• Governance has to be built, not inherited. The oversight structures required for compliant AI deployment don't exist in most organizations. They have to be designed, staffed, and operationalized. That's additive work, not a replacement for existing work.
• Change management is labor-intensive. Getting people to work differently — adopting new tools, trusting new outputs, rebuilding long-standing workflows — requires sustained human effort that no AI tool accelerates.
• The visible backlog grows before it shrinks. As AI makes more work tractable, the pipeline of demand expands. That can look like AI is creating problems. What it's actually doing is surfacing work that was always there but never pursued.

The right measure of success in this phase isn't headcount reduction. It's throughput per person, quality of delivery, and organizational capability accumulating in the system. The efficiency gains are real — they're just not yet visible in the metrics that feel most intuitive to track.

In years three through five, something different happens. As tooling matures, governance stabilizes, and workflows adapt, the productivity dividend becomes visible and measurable. Teams that invested in AI-era capabilities deliver work that would have required three times the headcount under the prior model. Organizations that tried to capture efficiency gains early by underfunding governance end up with tools that underperform, compliance gaps that create regulatory exposure, and a capability deficit that's expensive to recover from.

This is the Solow paradox playing out at the enterprise level — productivity gains from technology don't show up in the numbers until the organizational and human infrastructure has had time to adapt. The same pattern played out with enterprise computing in the eighties and nineties. It's consistent enough to treat as a planning assumption, not a risk to monitor.

What that means practically: the investment decisions made in 2026 and 2027 determine the capability position occupied in 2028 and 2029. The teams being asked to govern, integrate, and operationalize AI today aren't a cost center. They're the foundation of a significantly more capable organization in the medium term.

The Right Question Is Not About Headcount

When this question surfaces again — and it will — here's how I'd reframe it.

AI expands demand; it doesn't reduce costs against a fixed scope. Every capability gain widens what the organization believes it can ask of IT, which fills the pipeline with work that is real and legitimate. Resource constraints in an AI-enabled IT organization aren't a paradox. They're evidence the investment is working — that more is now possible than was possible before, and that the organization has started asking for it.

The efficiency gains are real. But they go to organizations that built the human infrastructure to govern and use AI — not to ones that tried to skip that step. Those organizations develop the accountability, governance, and translation capacity that makes AI something you can actually deploy and rely on. The ones that skip it end up with powerful tools and no one responsible for what those tools do.

The right question for executives isn't whether AI reduces IT headcount. It's whether IT capacity is growing fast enough to keep pace with the demand AI is creating.