Digital Labour Changes the P&L Shape

How AI agents bend the cost curve

Digital labour doesn’t remove cost. It changes the shape of it.

Most AI business cases still begin with labour substitution. Faster drafting. Fewer analysts. Reduced service headcount. Margin expansion through automation. Increasingly, those same cases are now being sold as agents doing the work.

In many cases, six months later, costs move sideways.

It shows up at forecast time, when the savings line is committed but the operating bill is still moving. The variance triggers scrutiny, not the demo. Most CFOs first meet “digital labour” as an unexplained line item variance, not as an AI initiative.

Compute rises. Vendor invoices fluctuate. Exception handling grows. Supervision absorbs senior time. Cloud line items swell without a clean causal chain. The promised savings are visible inside parts of the workflow, yet the overall P&L feels stubborn.

One reason is simple: teams speed up one cog while the surrounding cogs remain slow.

The automated step produces more volume and more variation, which piles into the next queue. Downstream teams absorb the surge through review, rework, escalation, and manual stitching. So the system doesn’t get cheaper. It gets busier. That is why local wins don’t reliably show up in the P&L.

In some cases the organisation walks back the ambition, reintroduces manual steps, or rehires capacity that was assumed to be “freed.” The system technically works, but the economics do not land cleanly.

The second reason is uncomfortable. AI agents often don’t make organisations do less. They make them do more.

Work accelerates, scope expands, throughput expectations rise, and the boundary between “possible” and “required” blurs. The gains are real, but they come bundled with demand creation and intensity. That is how “efficiency” turns into throughput, and throughput turns into a larger, more variable operating bill.

This is not a model problem. It’s a misread of the cost curve and what actually changes in the system. Many teams optimise the agent and leave the workflow largely intact.

Many organisations still model digital labour as if it behaves like salaried headcount: predictable, capped, and broadly linear.

Digital labour behaves like a metered utility bill. It is variable. It scales with usage. It responds to edge cases, retry loops, throughput spikes, and growth. It can become more expensive as it becomes more useful, because usefulness expands the surface area of use.

The shift is not “fixed cost to lower fixed cost.” It is fixed cost to variable cost plus forecast variance.

In the best cases, variance tightens, which is why it can feel like fewer forecast surprises, not just lower payroll. In the worst cases, variance widens enough to make the forecast worse even when mean cost improves.

This isn’t about being cautious. It’s about moving fast without inventing variance later. Organisations move faster after pilots when they price variability and stewardship early, rather than discovering cost ambiguity later through budget surprises and margin compression.

Three structural changes get missed.

Before we talk about savings, we need to name what changes in the cost stack.

First, compute and model usage are variable costs that scale with success.

When adoption rises, cost rises. When usage patterns shift, cost shifts. When a workflow expands into adjacent functions, cost compounds. Unlike headcount, these costs do not plateau naturally. They follow throughput. In practice that means every additional case processed, message handled, report drafted, or exception chased becomes incremental consumption and incremental cost.

Finance needs attribution at workflow-and-owner level. If you cannot attribute usage to a workflow, an owner, and an outcome, you do not have cost control. You have a moving bill with no owner. It becomes a cloud line item, the explanation becomes negotiation, and governance becomes retrospective.

I have seen this show up as “AI platform” spend being argued over in month-end reviews because nobody can tie it back to a specific workflow outcome.

Second, oversight and exception handling do not disappear. Digital labour reduces certain tasks and creates others.

Someone reviews anomalies. Someone monitors drift. Someone handles escalation. Someone maintains connectors and permissions. Someone answers the first regulatory query. That labour is often senior, and it is rarely priced into the original business case.

This is where work intensifies. Juniors and generalists can produce more drafts, more proposals, more analyses, more “first passes.” But the verification burden concentrates upward, onto domain experts. Seniors spend time reviewing, correcting, coaching, and protecting the organisation from plausible wrongness. That review work rarely appears as a named cost. It shows up as slack time disappearing and quality gates becoming informal.

This is also where simplistic governance answers fail. “Just require approvals” sounds safe, but it does not scale. As adoption grows, supervision shifts from step-by-step permissioning to monitoring, interruption, and intervention.

Oversight becomes a job with mandate, tooling, and protected time, not a checkbox. And that job needs real intervention rights: pause, throttle, safe mode, and escalation.

Third, maintenance and dependency accumulate.

Models evolve. APIs change. Vendors reprice. Contracts renew. Integrations age. What began as a pilot becomes infrastructure, and infrastructure requires care. The moment you rely on the system, you inherit its lifecycle.

None of this makes digital labour uneconomic. It changes the unit of analysis.

And this is where the unit of analysis has to change.

Headcount saved is a weak proxy. Cost per outcome is the governing metric.

“Outcome” should mean a completed business result that finance can count, price, and audit. Not an agent output. Not a task step.

It is the end-to-end unit the organisation is actually accountable for. The thing that clears a control, satisfies a regulator, closes a loop with a customer, or moves cash without rework.

In insurance, an outcome is not “a draft assessment.” It is a claim processed to decision and settlement, with the evidence captured, exceptions handled, and the audit trail intact.

In lending, an outcome is not “a summary of documents.” It is an application underwritten to a decision, with the required checks completed and the rationale recorded.

You can think of an outcome as the workflow completed to a defined standard. It includes the messy parts. Escalations, retries, reversals, human intervention, and incident handling are not outside the unit. They are inside it, because they are part of what it costs to get to “done.”

This definition matters because “agent output” is cheap and often looks impressive. Outcome is where cost, risk, and accountability meet. It is also where the P&L recognises reality.

Agents make outputs cheaper. That increases the temptation to manage volume. But value is realised at outcome level, not output level.

This is why outcome pricing is harder than it sounds.

“Outcome-based pricing” is an easy sentence. The hard work is defining the outcome unambiguously, agreeing what “good” looks like under exceptions, and agreeing who carries which risks when reality refuses to cooperate.

An outcome has two prices. There is the customer value, the benefit the business gains when the outcome is achieved. And there is the production cost, what it costs to deliver that outcome reliably under operating conditions.

The negotiation is usually the boundary of the outcome. Who pays for the tail. Missing documents. AML escalation. Abusive customers. Manual overrides. Incidents. Outcome pricing only works when you specify what is included, what is excluded, and how exceptions are priced. Otherwise it becomes fixed price plus resentment.

A practical way through is to price a corridor. Define the outcome unit, set an acceptable corridor for time, quality, compliance, and reversals, and price delivery inside it. Price the tail separately. Price the upside separately.

There is also a second category of gain that rarely lands cleanly in a labour-substitution business case: coordination cost.

Some of the largest payoffs do not come from doing a task “instead of a person.” They come from reducing the translation work between people and teams, where value leaks through reconciliation, rework, delay, and meetings whose only purpose is turning one team’s outputs into another team’s inputs.

AI agents can make that translation cheap and general, improving value flow even when nobody is “replaced.”

The P&L impact is real, but it often shows up as fewer reversals, fewer handoffs, lower delay cost, and less management time spent stitching the system together. Translation stops consuming management time and starts showing up as cycle-time reduction.

What does it cost to process a claim? To underwrite a loan? To resolve a customer issue? To produce a compliant report?

And what happens to that cost when volume doubles, edge cases increase, supervision tightens, and audit requirements arrive?

If you cannot model cost per outcome under realistic operating conditions, you are not governing digital labour. You are discovering it through variance.

In private equity-backed firms, this shows up particularly fast, because forecasts are tighter, targets are committed, and cost variance gets interrogated.

In PE-backed portfolio companies, a familiar pattern emerges. Labour savings narratives are sold early. Savings assumptions flow into forward projections. Variable costs then appear in cloud, data, or AI platform categories without a clean causal chain back to the use case. Exception handling labour grows. Switching friction rises. The initiative may be technically working, but the margin story blurs.

Once translation gets cheaper, coordination becomes something people fight to own.

One more reason the cost curve gets misread is power.

When translation becomes the coordination layer, someone will try to own it. Sometimes deliberately, sometimes by default, because visibility attracts ownership. Incumbents can become the translation layer. They can compete on accountability for outcomes. Or they can consolidate a privileged unified view internally and ration access to everyone else on paid terms.

In P&L language, that shows up as a new kind of operating bill: coordination rents. If you do not know who owns the unified view, you do not know who will price it. And because it is rarely attributable to one workflow, it shows up as “shared services” cost with a politics problem.

At that point, the relevant question is no longer “does it work?” It is “do the unit economics hold?”

For them to hold, several things must be true.

First, measurement must be real: a cost per outcome model that includes compute, supervision, maintenance, and risk overhead, with exceptions budgeted as structural, not treated as temporary noise.

Second, stewardship must be owned: clarity on who monitors drift, who intervenes, who handles incidents, and what “safe mode” looks like when the system misbehaves.

Third, reuse must be intentional: a reuse thesis that reduces marginal cost over time rather than proliferating bespoke deployments, supported by shared, validated components so every team is not rebuilding the same scaffolding in parallel.

And the organisation must know what happens to cost under stress: peaks, tail cases, retries, incident response, and the human time required to restore trust.

Digital labour does not remove cost. It changes where cost appears and how fast it scales.

You are not swapping fixed headcount for a predictable payment. You are exchanging it for a fluctuating operating bill tied directly to behaviour, usage, and growth.

When executives understand this geometry, something important becomes possible.

Agents can be governed like an operating system, not a headcount plan. CFOs regain control over margin by managing variability instead of chasing one-off savings. Reuse and standardisation start to matter because their impact on unit cost becomes visible. Boards can distinguish genuine productivity gains from cost displacement dressed up as efficiency. And across a portfolio, those outcome units become reusable measurement and control language, not bespoke “AI projects” in every company.

Digital labour can improve economics.

But only if the cost curve is understood before it is absorbed into the forecast.

The forecast doesn’t forgive ambiguity. It prices it.

If you cannot explain the new agent cost curve, you cannot govern the investment.

--

Stuart Winter-Tear

Founder, Unhyped | Author of UNHYPED

Advisory.