The Great Decoupling

Title: The Great Decoupling: The Evolution of Services Firms when the Value of Code Is Infinite, but the Value of Coding Is Zero
Author: Frank D'Souza, Co-Founder and Managing Partner, Recognize
Publication: Recognize (a private equity firm investing exclusively in digital services companies), self-published as an opinion piece
Date: April 2026
Who it's really for: Leaders of technology services firms rethinking their business model, and executives who buy technology services and need new vendor-evaluation criteria
Link: https://recognize.com/wp-content/uploads/2026/04/The-Great-Decoupling-4.15.pdf

The one-line takeaway¶

Writing code is heading toward zero cost. Making that code trustworthy is becoming the scarce, valuable thing. Services firms that keep billing for the old skill will lose to firms that get paid for the new one.

The argument¶

Frank D'Souza's thesis: technology services firms are living through a second industry-defining disruption in 40 years. The first was 1990s labor arbitrage: offshoring to places like India. This one is generative AI automating the work humans used to do. He calls the shift the "Great Decoupling": the value of code and the value of coding are moving in opposite directions (D'Souza, 2026).

The core mechanism: a "paradox of value." Code, the finished, working software, is becoming more valuable. Software increasingly runs global finance, healthcare, and infrastructure, so the stakes of getting it right keep rising. Coding, the manual task of writing that code, is becoming a commodity. LLMs can turn a business requirement into working syntax at close to zero cost. AI now touches almost every stage of the software lifecycle. The output is worth more than ever. The labor that used to produce it is worth almost nothing. Forty years of "coding scarcity" have flipped into an era of "coding abundance."

A second paradox follows: the "burden of abundance." Cheap, plentiful code is not the same thing as reliable software. As code gets more abundant, the cost of certainty actually rises. Certainty means making that code secure, ethical, fit for purpose, compliant, and properly integrated, and there's simply more surface area to get wrong. Abundant machine-generated code creates three problems. Technical debt: AI can write legacy-style code faster than humans can understand it. An accountability vacuum: no clear owner exists when machine-generated code causes a systemic failure. And security risk: more code means more vulnerabilities to patch and monitor.

The bottlenecks have not disappeared. They've moved. In the scarcity era, the bottleneck was the manual labor of writing syntax. AI has compressed that dramatically, with the biggest gains in implementation-stage tasks. But two new bottlenecks have emerged at the edges of the lifecycle. Upstream, the bottleneck is now describing intent: turning messy human problems into specs a machine can execute. Call it a "first-mile" problem: AI can build the house, but it does not know what the residents actually want. Downstream, the bottleneck is the cost of certainty: making AI-generated code secure, compliant, and properly integrated. Call it a "last-mile" problem. D'Souza's conclusion: code may be abundant, but orchestrating that code into a trusted outcome is now the industry's most valuable work.

What this implies for services firms. The market will no longer pay for the act of building, because building is now cheap. It will pay more than ever for the impact of the build. D'Souza argues this does not kill the case for services firms. It reinforces it. Firms stay valuable when they turn abundant code into trusted, secure outcomes faster than a client could alone. From this he draws several necessary changes:

Commercial models must decouple revenue from effort. Firms billing by the hour will struggle as the hours needed for any given result keep falling. Firms need to monetize outcomes, IP, or usage instead, so they keep AI's efficiency gains rather than handing them to the client for free. He proposes a three-phase transition: input pricing with hidden AI-driven margin gains, then output pricing via fixed-fee "certainty bundles." The third phase is outcome-based pricing tied to a share of the value created. The goal is avoiding the "J-curve," the period where AI efficiency shrinks billable hours and revenue before new pricing catches up.
Talent structure must shift from headcount to "hybrid teams" of humans plus AI tools and agents. Success now comes from human judgment at the right points: defining intent, ensuring compliance and safety, not from bench size.
Firms need proprietary IP (their own code, data, models, and agents) as the real differentiator. Generic AI tooling is available to everyone and gives no edge on its own.
New human roles cluster around the two new bottlenecks. Upstream: value orchestrator, enterprise architect, intent curator. These roles define and translate what should be built, and why. Downstream: results orchestrator, ethics steward, accountability steward. These roles turn AI output into a coherent, safe, owned result.
This needs a real cultural break, not a tweak. Incentives built around headcount and utilization, bonuses tied to team size, success measured by bench strength, have to go. Replace them with incentives tied to outcomes ("outcome density"). He argues this must be signaled top-down, "burning the boats," not left to emerge on its own. Clients need the same push, away from timesheets and toward outcome-based evaluation.
The talent pyramid inverts into a diamond. The old model hired large junior cohorts to do manual coding. AI removes the need for that base. The new shape is a broad middle of orchestrators and stewards managing AI output, with only a small, curated base of apprentices. He flags this as a real problem: fewer junior roles means fewer training grounds for future senior talent. He argues the industry must solve this deliberately, through apprenticeship, rather than let it erode.

What this implies for buyers of services. The old evaluation tools (rate cards, headcount audits, utilization tracking) measured effort. Effort is exactly what's now cheap and abundant, so those tools no longer separate good providers from bad ones. Buyers should shift from requesting capacity ("30 Java developers for 12 months") to specifying outcomes. An example: "reduce claims processing time by 40%, with a full security and compliance guarantee." He proposes a "certainty scorecard" across four dimensions. Outcome reliability: a track record of hitting business milestones, not just shipping code. Proprietary IP: real depth beyond commodity AI. Security posture: managing the risk of machine-generated code at scale. Accountability: owning and fixing failures. He also flags a new competitor for external firms: the client's own internal team, or Global Capability Center. If an internal team can generate "good-enough" code cheaply, the bar an external firm must clear goes up.

Closing frame. D'Souza calls this one more wave the industry has weathered before: offshoring, the internet, the cloud. He argues firms that adapted early each time came out ahead. His summary: AI is a powerful tool, but still a tool. The value of a hammer fell when the nail gun was invented; houses got more valuable, not less. The real leadership question is not "how do I produce more code?" It is "how do I turn abundant code into trusted outcomes faster than anyone else, without losing integrity or security?"

The evidence behind it¶

OpenAI valued at $840 billion, Anthropic at $380 billion, "early 2026" (D'Souza, 2026), publicly checkable, but the article names no source. These figures do check out against independent reporting for February 2026. Worth knowing: they were already out of date within weeks of the article's own April 2026 publication. By May 2026, Anthropic's valuation nearly tripled to $965 billion in a new Series H round (Anthropic, 2026). That overtook OpenAI, which had closed its own round at $852 billion in March (reporting, 2026). The reversal runs opposite to the ordering the article implies. Treat this comparison as a snapshot of a fast-moving number, not a stable fact.
"Valuations of publicly listed technology services firms have declined over the past 12 months" (D'Souza, 2026) is publicly checkable in direction. But no index, ticker, or figure is given.
AI productivity-gain chart ("AI's Uneven Impact on Software Development: Estimated time saved by task"), the author's own estimate. It is labeled "Estimated," with no stated methodology (D'Souza, 2026). Six lifecycle phases, with these aggregate figures: discovery & planning (40%), analysis & specification (50%), design (55%), implementation (70%), integration & testing (55%), deployment (45%). The single most-cited figures: boilerplate coding and unit testing show reductions of up to 90% and 85%, respectively.
Illustrative examples throughout. A "45-person team for 18 months" shrinks to "a hybrid team of six people and a suite of AI agents." A team completes "in 5 hours what used to take 40." A procurement shift moves from "30 Java developers for 12 months" to "reduce claims processing time by 40%." All are explicitly hypothetical ("might have"), not named real engagements (D'Souza, 2026).

No third-party firm, analyst, or research house, no McKinsey, Gartner, IDC, or similar, is named anywhere in the piece. The only named entities besides the author and his firm are OpenAI and Anthropic, cited only as valuation subjects.

How much to trust it¶

This is a practitioner opinion piece, not an independent study, and the author has a direct financial stake in you believing it. Frank D'Souza runs Recognize, a PE firm that invests exclusively in digital services companies. The disclosures page says plainly this "reflects his own views" and "is not an offer of investment advisory services." That's honest framing, but it does not erase the conflict. His prescriptions (stop billing by the hour, build proprietary IP, move to outcome-based pricing) are exactly the playbook a PE-backed services firm would want. A firm that pivots successfully is worth more to its investors, D'Souza included.

Almost all the evidence here is the author's own analysis, not independent data. The paper cites exactly two outside data points, the OpenAI and Anthropic valuations, and neither names a source. We independently confirmed the figures were accurate for the moment cited, but they moved sharply within weeks. Everything else, including the one chart with hard-looking numbers (30% to 90% productivity gains by task ), is D'Souza's own estimate. It is not survey data or measured research; the chart is labeled "Estimated" and cites no methodology.

Where it overreaches: the productivity chart shows precise-looking percentages with zero sourcing. The most quantified claim in the paper is also its least substantiated. The comparison between AI-company valuations and services-firm valuations reads as market "signal." But it's a suggestive correlation with no control for other factors, and (per the check above) already stale in its own framing. The piece reads as neutral, practical advice throughout. The commercial stake only shows up on the final disclosures page.

One more thing worth knowing before you treat this as a fresh 2026 insight: most of the prescription is long-circulating services-industry conventional wisdom. Outcome-based pricing, IP as a moat, inverting the talent pyramid: all forecast for well over a decade without displacing hourly billing at scale. What's genuinely new here is the framing (decoupling, first-mile/last-mile), not the recommendations themselves. That does not make the advice wrong, but it's a different claim than "this investor spotted something new."

Bottom line: treat the framework (decoupling, upstream and downstream bottlenecks, diamond versus pyramid) as one experienced investor's informed opinion, worth taking seriously. It's a genre of advice this industry has heard before. Treat the productivity chart as illustrative, not measured. And read the "adapt or die" urgency knowing the author's fund benefits if you believe it.

So what — for you¶

If you run a services or consulting firm: most of what follows is a leadership- or board-level call. It is not something a delivery lead or practice head can execute alone. If that's your seat, treat this as the case to build upward, not a personal to-do list. None of it depends on the productivity chart's exact numbers being right. The pricing and structure argument holds even if AI's gains turn out smaller than D'Souza estimates. What changes with the numbers is the urgency, not the direction.

Stop selling hours, start selling certainty. If you're still billing time-and-materials, you're exposed. AI is cutting delivery effort while your revenue stays tied to the hours that effort used to take. Move toward fixed-fee "certainty bundles," or outcome-based pricing tied to a metric the client cares about. Think a percentage reduction in claims processing time, not headcount deployed.
Retire utilization and headcount as your success metrics, visibly. D'Souza's call: formally kill bench-size and utilization as KPIs. Replace them with "outcome density": value delivered per team member per quarter. This has to come from leadership, top-down, or sales and delivery teams keep optimizing for the old incentive.
Reshape the team from a pyramid to a diamond. Stop hiring big junior cohorts for manual coding: that's the layer AI eats first. Build a thick middle of people who direct and validate AI output, plus a small, carefully chosen base of apprentices. Solve junior-talent development on purpose, or you lose your senior pipeline in five to ten years. Worth naming plainly: "the layer AI eats first" is a real cohort of jobs (graduate hires, offshore delivery staff), not just a succession-planning line item. "Burning the boats" is easier to prescribe from outside the firm than to execute against the people it costs.
Name who owns it when the AI is wrong. If your firm cannot answer "who's the accountability steward here," that's a sales liability with sophisticated buyers, not just an internal question.
Invest in your own IP, not a wrapper around someone else's AI. Generic tooling is available to every competitor and to the client's own team. Your pricing power comes from what a client cannot easily replicate in-house.
Expect margin expansion before revenue growth. Plan for the dip in between. His suggested bridge: keep familiar hourly pricing short-term, but use AI internally to shrink team size while holding rates steady. Bank the margin while you build the models that eventually replace it.

If you buy technology services:

Write outcome specs, not staffing requests. Ask for "reduce claims processing time by 40%, with a full security and compliance guarantee." Do not ask for "30 Java developers for 12 months." A vendor that can only offer capacity is pricing you into the old, exposed model.
Replace the rate card with a certainty scorecard. Check vendors on outcome reliability, proprietary IP depth, security posture, and accountability: will they actually own a fix when something breaks.
Benchmark vendors against what your own team could do with the same AI. If your in-house team can generate "good-enough" code cheaply, that's a real negotiating advantage. Use it.
Ask for certainty delivered, not hours worked. Request dashboards on security, compliance, and milestone progress instead of timesheets. A vendor whose only reporting is a timesheet has not made the shift yet.

What not to overreact to: this is a thesis from an investor whose fund benefits if the "adapt or die" framing lands hard. Read the urgency with that in mind. There's no independent survey data, no named client outcomes, and no dated adoption statistics behind it. The percentages and the whole framework are illustrative and author-original, not measured.

The fine print¶

Frank D'Souza is Co-Founder and Managing Partner of Recognize Partners LP, a private equity firm investing exclusively in digital services companies. The disclosures state the paper reflects his own views, and "does not necessarily reflect the views of any entity he represents." It's furnished "solely for informational purposes" and is not investment, legal, tax, or accounting advice, nor an offer tied to any Recognize investment. The disclosures also note some information came from "published and non-published sources prepared by other parties" that "has not been independently verified." They add that any estimates or illustrations "should be regarded as illustrative only." No individual figure gets a named source in the piece itself. And the one pair of figures we could check, the OpenAI/Anthropic valuations, had already shifted enough within weeks to reverse the comparison.

Sources¶

Frank D'Souza, Recognize (Co-Founder and Managing Partner) — The Great Decoupling: The Evolution of Services Firms when the Value of Code Is Infinite, but the Value of Coding Is Zero, 2026. In early 2026, OpenAI was valued at $840 billion and Anthropic at $380 billion... boilerplate coding and unit testing have seen reductions of up to 90% and 85%, respectively. View source · verified 2026-07-10 · primary
Anthropic — Series H announcement, 2026. Anthropic has raised $65 billion in Series H funding led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital, valuing the company at $965 billion post-money. View source · verified 2026-07-10 · primary
Yahoo Finance (reporting) — OpenAI raises $122 billion at $852 billion valuation, 2026. AI giant OpenAI announced a massive funding round of $122 billion at a valuation of a whopping $852 billion on March 31, sending shock waves throughout the tech world. View source · verified 2026-07-10 · ⚠ secondary mirror