The Pendulum Swing
Why Enterprises Are Rethinking the COTS Platform — and the Discipline That Makes Leaving One Survivable
James "Jamie" Campbell, Product Owner, Alchemist AI Pro™
ACC3 International / AI Pro Holdings · July 14, 2026
A Note on Evidence, Before We Start
This paper makes a claim you will not find in an analyst report yet: companies in large numbers are exploring getting off their COTS platforms because the value of those platforms has diminished in their view.
Some of what supports that claim is historical — dated events, public filings, published research. Those items are marked [H] and cited in the references. The rest is observable but not yet citable: market behavior, pipeline conversations, patterns any practitioner can see but no institution has yet written down. Those items are marked [A] for anecdotal, and the references section says so plainly rather than dressing them up as something they are not.
That split is deliberate. The most interesting moment in any market shift is the window after the evidence becomes visible and before it becomes citable. That is the window we are in. When the analyst reports arrive — and they will — the CIOs and consulting firms who moved during this window will already be positioned. This paper is for them.
1. The Claim - Two Facets: The CIO and the Systems Integrators
Enterprises spent two decades consolidating work onto commercial off-the-shelf platforms — Pega, ServiceNow, Adobe, Salesforce, Appian, and their peers. That consolidation made sense, and this paper will give the platforms their full due. But the consolidation logic has now inverted for a specific class of workload, and the market is repricing accordingly.
The pendulum has swung before. Custom mainframe code in the 1960s and 70s gave way to packaged ERP in the 80s and 90s. Packaged software gave way to SaaS in the 2000s, and SaaS to low-code platform suites in the 2010s. Each swing happened because the economics of one side broke. [H — this arc is standard industry history]
This swing is happening because the economics of building changed faster than the economics of buying. AI-assisted development matured dramatically after October 2025, and organizations are now asking a question that was unaskable three years ago: can we endure one season of CapEx to retire years of platform OpEx? [A — the question is showing up in real pipeline conversations; no analyst has yet quantified it]
Exhibit 1 — The Pendulum. Build gave way to buy across four prior swings — packaged ERP, SaaS, low-code — each triggered when one side's economics broke. Each swing happened because one side's economics broke. This one is no different.
2. Why CIOs Bought COTS in the First Place - and Why They Were Right To
Be honest about the original bargain, because the exit only makes sense if you understand what you are exiting.
CIOs purchase COTS platforms to shed risk on technology that is not their competitive advantage. A regional insurer does not win customers with a hand-built workflow engine, so it rents one that ten thousand other companies have already debugged. Then the CIO transfers a second layer of risk by hiring a consulting firm for the configuration spike, avoiding permanent staff for temporary demand. Two rational risk transfers, stacked.
And the platforms delivered real, specific value — not generic value. Each of the majors earned its position with something genuinely hard to replicate:
- Pega - the Situational Layer Cake™. One application definition, specialized by layer for every region, product line, and channel — a global bank runs one case-management build that behaves correctly in forty jurisdictions without forty codebases. That architecture is the reason Pega owns complex, regulated case management.
- ServiceNow - one platform, one data model. The CMDB plus a single workflow engine spanning IT service, operations, HR, and security. Its power is that every workflow in the enterprise reads and writes the same system of record — the integration tax approaches zero inside the walls.
- Adobe - the content supply chain. Creative tooling, asset management, and customer-experience data joined end to end, so the asset a designer produces is the asset the personalization engine serves, with the rights metadata intact. Nobody else connects creation to delivery at that depth.
- Salesforce - the metadata-driven platform. Objects, fields, workflows, and permissions customized declaratively on multi-tenant infrastructure, with the AppExchange ecosystem around it. Its genius was making deep customization survivable through upgrades — configuration lives above the code line.
- Appian - process plus data fabric. Low-code process automation that unifies data from systems you do not migrate. For process-heavy shops with scattered estates, that combination compressed delivery times enterprises could not achieve any other way.
Here is the pattern worth staring at: in every case, the platform's real value is buried deep inside complexity. The Layer Cake matters when you have forty jurisdictions, not four screens. The CMDB matters at ten thousand configuration items, not a hundred. When the work is genuinely complex, the COTS bargain still holds — and this paper is not going to pretend otherwise.
The problem is what got built on these platforms when the work wasn't complex.
Around 2022, the platforms and their major systems-integrator partners began selling solutions for work that did not need a heavy platform at all. Intake forms. Departmental trackers. Approval chains. Reporting portals. Point solutions with no forty-jurisdiction problem, no ten-thousand-CI estate — but the platform was already there, the license was already signed, the SI already had the bench. Pressing the Easy Button was cheaper than standing up anything new, so the periphery of the application portfolio migrated onto premium infrastructure. [A — every practitioner in the ecosystem watched this happen; you will not find an analyst note calling it out at the time, because the analysts were scoring it as platform "adoption"]
For three years this looked like growth. Platform revenue compounded, SI bookings compounded, and the quadrant charts rewarded it.
Borrowing a phrase from trading: those three years were the extended-range candle of the business application industry — the final elongated push of a long trend, printed with volume and conviction, right before the reversal. The applications that drove it were never worth platform pricing on their own merits. They were worth platform pricing when the alternative was a custom build at 2022 costs. That "when" just expired.
Exhibit 2 — The Extended-Range Candle. Illustrative, not market data. Three years of peripheral workloads bought at platform prices — the point-solution land grab of the "Easy Button" era — printed the last leg of the old trend before the October 2025 repricing began.
4. October 2025: Three Things Happened
First, Gartner moved. On October 20, 2025, at its IT Symposium/Xpo, Gartner named AI-native development platforms a Top Strategic Technology Trend for 2026 — and attached a number to it: by 2030, these platforms will lead 80% of organizations to evolve large software engineering teams into smaller teams augmented by AI, with engineers working as "forward-deployed" partners to domain experts. [H1] Read that carefully. The institution whose quadrants served as the marketing engine of the buy-side for twenty years told every CIO on the planet that small teams building custom software with AI is a strategic trend, not a hobby. When the referee changes the rules, the game changes — and CIOs cite Gartner to their boards for a living.
Second, the capability arrived. AI-assisted development matured dramatically through late 2025 and into 2026 — agentic coding tools, longer-horizon autonomy, and dramatically better code generation. This is not a claim that raw AI coding is enterprise-ready; our own research library documents exactly why it is not (a 2025 study found 27.3% of AI-generated code samples carried security weaknesses [H6], and surveyed requirements engineers report that 81.2% still manually check and approve AI suggestions [H7]). The claim is narrower and stronger: the cost floor of building fell an order of magnitude, which means the build-versus-buy calculation reopened for every workload on the periphery. What determines whether the build succeeds is no longer the coding — it is everything upstream of the coding. Hold that thought; it is the hinge of this paper.
Third, the trouble. Those 2022–2025 point solutions now sit on premium OpEx forever — license, support, and the rising cost of platform-certified resources who grow more skilled and more expensive every year, then leave for better offers. Organizations have started doing the arithmetic on whether one season of CapEx can retire that OpEx for years. [A — this is the observable insight; it lives in pipeline conversations and CFO spreadsheets, not yet in any published survey]
5. The Anecdotal Evidence: Read the Charts
I told you which parts of this paper are circumstantial. Here is the circumstantial centerpiece — and you can verify every pixel of it yourself with a free charting tool.
Pull up the stock charts of Pegasystems, ServiceNow, Adobe, Salesforce, and Appian from January 2020 to today. Five different companies, five different products, one shared shape: the COVID-era run-up, the 2022 rate-shock drawdown, the recovery — and then, from late 2025 into 2026, a divergence from the broader market that no rate story explains. The market is not repricing these companies' execution. It is repricing their model. [A1 — the charts are public fact; the interpretation is ours]
Exhibit 3 — Five Charts, One Shape. Weekly closes, January 2020–July 2026, for Pegasystems, ServiceNow, Adobe, Salesforce, and Appian: since Gartner named AI-native development platforms a top 2026 trend on October 20, 2025, each has fallen between 15% and 51%. The prices are public record; the interpretation — that the market is repricing the model, not the execution — is anecdotal. See references [A1].
Now the historical anchors that turn the shape into a story:
- The 2026 software repricing. Between January and early February 2026, software stocks lost more than $1 trillion in market value while the S&P 500 stayed flat; the iShares Expanded Tech-Software ETF fell roughly 21%, and Salesforce fell 26%. Enterprise behavior moved the same direction: the average number of SaaS applications per company declined from 112 to 106, 82% of companies reported reducing supplier count, and per-seat pricing fell from 21% to 15% of SaaS pricing models in twelve months while hybrid consumption pricing surged from 27% to 41%. [H3] Sellers do not abandon seat pricing in a market that believes in seats.
- IBM, today. On the morning this paper was finished — July 14, 2026 — IBM shares fell roughly 23–25% after a second-quarter earnings warning, pacing the company's worst trading day since October 1987 and dragging software stocks down with it. Management pointed to clients redirecting spend toward hardware at the expense of software and consulting. [H4] Whatever the proximate cause, a one-day repricing of that violence means the market has stopped extending the benefit of the doubt to legacy delivery models. Growth built on yesterday's structure is now suspect on contact, because buyers understand that new things are arriving while old things go away at an alarming rate.
- Klarna, the early tell — with the honest asterisk. In September 2024, Klarna's CEO announced the company was dropping Salesforce and Workday in favor of an internally built, AI-driven stack. The follow-through was messier than the headline: Klarna ended up consolidating onto a proprietary data core supplemented by other SaaS tools, and its CEO later said he doubted most companies would replace Salesforce with AI. [H5] Skeptics read that as the exodus debunked. Read it again. A payments company looked at two of the most entrenched platforms in enterprise software and concluded that leaving was worth the pain — in 2024, before the tooling matured. The lesson of Klarna is not "it can't be done." The lesson is that leaving a platform without a requirements discipline is how you end up with a messier version of what you left. Which brings us to the actual subject of this paper.
6. What the CIO Actually Loses - and How Each Piece Gets Replaced
Strip the COTS bargain to what it really provides, because an exit plan that only counts the license savings will fail. The platform gives a CIO six things:
- Constant feature enhancements, delivered on the vendor's schedule
- Someone to call when there is a crisis
- Someone who understands the base platform
- A community of users exploring ideas
- Reduced cost of ownership — you staff only the custom configuration and share the base's cost with every other licensee
- The soft thing nobody itemizes: a steady stream of ideas for what to do with what you "already have"
Any serious exit has to replace all six. In addition, it needs to consider the incumbency and the install base of "doing nothing." Here is how the ledger reads when the replacement is AI-assisted delivery run under the Alchemy SDLC — the full lifecycle ACC3 built around Alchemist AI Pro™: an Alchemy Business Analyst for discovery and vision architecture, an Alchemy Crew for development and functional testing, and an Alchemy Away Team for guardrail-driven repair, with requirements traceable end to end.
| What COTS provided |
The Alchemy SDLC replacement |
| Vendor feature roadmap - enhancements arrive on their schedule, priced into the license |
A roadmap you own. Requirements live in a closed loop with the code repository; drift detection keeps specifications synchronized with deployed reality, and version-controlled requirement sets carry new features forward without losing the base. You ship when the business needs it, not when the release train runs. |
| Someone to call in a crisis |
The Away Team. Guardrail-driven repair that classifies each anomaly by origin and impact before writing a line (IEEE 1044), reads the spec before the code, and monitors the four golden signals across service layers (Google SRE). A vendor's support queue is replaced by a repair discipline that traces every fix to the requirement defining correct behavior. [H7] |
| Someone who understands the base |
You own the base — documented deeper than the vendor ever documented theirs. Every table, route, and business rule traces to the requirement that demanded it. And the exit itself is where this starts: code archaeology reads the existing platform configuration and extracts the requirements buried in it, so twenty years of institutional logic comes out of the platform instead of dying inside it. |
| A community of users |
The honest entry in this ledger: you lose the vendor community, and nothing replaces it one-for-one. What you gain instead is the entire open software ecosystem plus a traceable spec library of your own — but a CIO should price this loss honestly rather than wave it away. |
| Reduced TCO via the shared base |
The economics inverted — that is the whole point. The shared-base discount stopped mattering when the license line item became the thing being retired. AI-assisted teams run smaller: Gartner projects 80% of organizations move to small AI-augmented teams by 2030 [H1]; our own delivery evidence ran 3–4 people against a planned 9. |
| Ideas for what you already have |
Socratic elicitation on your own estate. The Alchemist's guided questioning — including a Frameworks stage that surfaces identity, integration, data, observability, accessibility, and compliance concerns you had not thought to raise — does for your portfolio what the vendor's account team did for theirs, minus the incentive to sell you another module. |
The Proof This Is Not Theoretical
The Alchemy SDLC has delivery evidence. The Product Campaign Manager build — an autonomous marketing-orchestration platform with a human director holding final approval — went through the full lifecycle: Alchemist produced the design specification, 395 use cases, business specifications, and 1,086 test scripts tied to acceptance criteria in 6 hours; Alchemy Crew delivered a functional cloud-deployed system in 9 hours of build time. Final tally: all 10 features and 395 use cases delivered (100% feature completeness), 95% requirements adherence, 4 weeks against a planned 50, a 3–4 person team against a planned 9, and a $250K cost against a $2.27M plan — roughly $2M saved. [P2, P3] Alchemist AI Pro itself was assessed by the Department of Defense against a full rubric of requirements and made Tradewinds Awardable — a procurement-readiness signal, stated precisely, not an award or endorsement. [P2]
Exhibit 4 — The OpEx Crossover. Illustrative — model your own portfolio. Cumulative platform OpEx (license, support, the certified-resource premium) versus a one-season CapEx rebuild followed by maintenance: the crossover marks the point every year beyond it retires platform spend. The Product Campaign Manager case study figures [P3] are the real-world reference point — $250K actual vs. $2.27M planned, 4 weeks vs. 50, 3–4 people vs. 9.
The CIO Playbook, Concretely
- Inventory the periphery. Separate the workloads that use the platform's deep capability (the Layer Cake, the CMDB) from the 2022–2025 cohort that landed on the platform because it was there. The first group stays. The second group is the candidate list.
- Run archaeology before you run away. Extract the requirements embedded in the platform configuration — the business rules, the exception paths, the tribal decisions nobody wrote down. This is institutional-knowledge preservation, and skipping it is how modernization programs silently drop capability. [P4 — see Why Software Modernization Programs Fail and the EKE series in the ACC3 library]
- Rebuild against requirements, not against the old screens. Statement-for-statement conversion — "paving cow paths" — carries every piece of technical debt forward and ignores what the new stack does well. Capture requirements platform-independent, then map to the target's strengths. [P4]
- Sequence the CapEx per application, and measure like a portfolio. Each rebuild retires a specific slice of license OpEx on a specific date. Screen every effort through the discipline of experiment → capability in development → deployed solution → sustained outcome, and be ruthless about the last gate. Gartner predicted 30% of generative-AI projects would be abandoned after proof of concept [H2] — the exits that fail will fail there, at the pilot-to-production gate, which is precisely the gate requirements discipline exists to hold.
- Leave the core alone. The pendulum swings; it does not detach. Deep-complexity workloads still earn their license. A CIO who rebuilds the forty-jurisdiction case-management core in year one has confused a pendulum with a catapult.
7. For the Consulting Firms: The Pyramid Is the Exposure
Now the other reader of this paper, because the same swing that threatens the platforms threatens the firms that implement them — and offers them a better business than the one they are losing.
The consulting business model is a pyramid: a partner sells, a handful of seniors direct, and a wide base of junior resources executes billable tasks. The pyramid's economics require the base — that is where the margin lives — and the base is exactly what AI assistance absorbs. Clients have noticed. Organizations are balking at paying for a bench of people to do junior-level tasks an AI does in minutes, and the shift in client demand is toward outcome-based contracts. [A4 — observable in deal negotiations across the industry; the IBM repricing [H4] is the market saying the same thing at volume]
Here is the part almost nobody says out loud: outcome-based contracts were never a pricing innovation problem. They were a risk problem — specifically, the risks of humanity. A firm pricing a fixed outcome had to absorb attrition mid-project, vacations at the worst moment, the variance between its best architect and its median one, key people poached at the deadline, and estimate error compounding across all of it. Priced honestly, the risk premium made outcome contracts uncompetitive against time-and-materials. So the industry stayed on the pyramid, and everyone pretended it was a preference.
AI-assisted delivery collapses those risks. The assisted team does not take vacations from its context. There is no attrition of the knowledge held in the requirement set — continuity lives in the artifacts, through the whole lifecycle. Delivery variance compresses when every task traces to an EARS-formatted requirement with pre-derived test scripts. And that changes what is finally contractible: you cannot price an outcome you cannot define, and you cannot define an outcome without hyper-defined requirements. This is the point where the Alchemy SDLC stops being a delivery method and becomes a business model:
- Sell acceptance, not hours. EARS-formatted acceptance criteria plus test scripts derived from use cases before development starts means the contract's definition of done exists on day one — measurable, testable, adjudicable. Disputes get settled by the trace matrix, not by the steering committee.
- Restructure around judgment. The pyramid's base is gone; the top is what clients still pay for. Small senior teams directing assisted delivery — 3–4 people doing the planned work of 9 [P3] — with margins that come from outcomes delivered, not hours logged.
- Build the new offering: the COTS-exit practice. Portfolio triage, configuration archaeology, requirements extraction, staged rebuilds that retire license OpEx on a schedule. Every anxious CIO reading section 5 of this paper is a client for it. The firm that productizes the exit will take share from the firm still selling the implementation.
- Run it on the Alchemy SDLC and keep your brand on it. The lifecycle — Business Analyst, Crew, Away Team, with the independent Audit stage in which a second AI model, distinct from the first, audits the output of the first [H7, P4] — is the delivery apparatus that makes an outcome promise survivable. The firm brings the client relationship and the domain judgment; the SDLC brings the repeatability that keeps the fixed price fixed.
The firms that make this turn get a better business than the one the market is currently repricing: higher margin, defensible deliverables, and contracts clients actually prefer. The firms that do not will discover what IBM's shareholders discovered this morning — that the market reprices legacy models all at once, not gradually. [H4]
Exhibit 5 — Pyramid to Outcome. The pyramid sells hours; margin lives in the wide junior base AI now absorbs. The outcome model sells acceptance; margin lives in judgment, with a small senior team directing an assisted delivery core against requirements, a trace matrix, specs, and test scripts. Outcome contracts were never a pricing problem — they were a risk problem, and the artifacts are what make the outcome contractible.
8. Then Versus Now: The Same Six Steps, a Different Discipline
The compare-and-contrast that makes this concrete is the Alchemist AI Pro Journey — the without/with walk across all three roles of the lifecycle. [P1] Where we were then is the left column of that journey. Where we are now is the right. Three snapshots:
The Business Analyst, then: whole documents arrive unstructured; SMEs burn cycles re-explaining context; unstructured interviews let tacit assumptions slip through, and ambiguity surfaces late — found by developers, mid-sprint. Now: code repositories and reference material are read directly; Socratic dialogue drives the ambiguity out before handoff; requirements land in EARS format, which was designed to eliminate the named defect classes — ambiguity and vagueness first among them. [H7]
The Developer, then (the "vibe coding" era): prompts produce plausible-looking code with no scope baseline; each AI session is siloed with no shared data model, so today's schema contradicts last sprint's; "done" means it compiled once. Now: EARS-validated stories drive the sizing, a live ontology of actors, objects, processes, and rules is declared once and referenced everywhere, and every commit tags to its spec, test case, and use case — traceability from requirement to merge. [P1]
The Final Mile, then: fixes target the first error in the log; UX validation stops at the happy path; every patch adds silent assumptions. Now: the Away Team classifies the anomaly before writing the fix, validates the full journey including the exception paths, and treats a repair that deviates from spec as a spec update first — not a code update. [P1, H7]
Notice what changed and what did not. The six steps are the same steps enterprise software has always required. What changed is that the artifacts are now produced with a speed that matches AI-assisted development instead of dragging three sprints behind it. Requirements moving slow while coding moves fast was the failure mode; the answer was never to skip the requirements. It was controlled acceleration — speed with steering, brakes, and telemetry. [P4]
9. Why the Artifacts Matter No Matter Who — or What — Writes the Code
Everything above reduces to one claim that deserves its own section, because it is the reason the Alchemy SDLC works for the CIO's rebuild and the consultancy's outcome contract alike:
Certain artifacts produce better software regardless of whether humans or assisted AI do the building — because the defect source they eliminate is ambiguity, and ambiguity does not care who types.
Fifty years of software engineering established that requirements defects are the most expensive defects, growing 30–100× costlier when discovered late. AI raises those stakes rather than lowering them. A large language model is an amplifier of its inputs: hand it an ambiguous requirement and it does not ask the clarifying question a senior developer would — it commits to one interpretation, confidently, at scale. The research record is blunt about the current state: 27.3% of AI-generated code samples carried security weaknesses [H6]; 61.9% of requirements engineers report interpretability challenges with LLMs; 81.2% still manually check and approve AI suggestions, and 71.9% frequently correct or override AI-generated requirements — organizations doing the homework the AI was supposed to eliminate. One study found that more than half the time, an LLM could not reproduce its own results even at temperature zero. [H7]
None of that argues against AI-assisted delivery. It argues for exactly four properties in whatever surrounds it — and these are the properties the Alchemy SDLC's artifacts exist to supply:
- Sustainable — the requirement set survives the people. Attrition, the silver tsunami of retiring expertise, the consultant rolling off: the knowledge lives in versioned artifacts, and closed-loop synchronization with the repository keeps those artifacts matched to deployed reality instead of drifting into fiction.
- Repeatable — EARS-formatted requirements, a canonical ontology, and use-case-derived test scripts mean the second build behaves like the first, and the tenth. Repeatability is what lets a consultancy sign an outcome and a CIO trust a schedule.
- Auditable — every decision traces to its source; an independent Audit stage runs a second AI model against the first's output before a human ever signs; the trace matrix stands as delivery evidence a contract, a regulator, or a board can examine. No black boxes — the failure that made the last generation of conversion tooling untrustworthy.
- Human-held intent — AI does the tasks that can be measured, tested, traced, and repeated; humans hold the judgment and the sign-off. That division of labor is not a compliance gesture. It is what keeps the amplifier pointed at the right target.
We have written plenty about Alchemist AI Pro and its uses. This is the reason underneath all of it: the artifact discipline is what converts AI-assisted speed from a demo into an asset a business can depend on — on time, the first time.
Exhibit 6 — The Alchemy SDLC, One View. Eight guided stages, three roles, one closed loop: the Alchemy Business Analyst carries Capture through Frameworks, the Alchemy Crew runs Elaborate through Export (with an independent second AI model auditing the first's output at the Audit stage), and the Alchemy Away Team performs guardrail-driven repair on the live, deployed system — feeding code archaeology and drift detection back into the next iteration.
10. Where the Pendulum Stops
It does not stop at "rip out every platform." Deep complexity still earns the license, and the vendor community is a real loss that honest CIOs will price. The pendulum stops where it always stops: at the point where each workload sits on the economics that actually justify it. The 2022–2025 periphery goes home to purpose-built, AI-assisted, requirements-governed applications. The core stays where the Layer Cake and the CMDB still do work nothing else can.
What determines who navigates this swing well is not access to AI tools — everyone has those. It is whether the organization can turn what it wants into hyper-defined, traceable requirements fast enough to keep up with what the tools can now build. That has been the deciding discipline for fifty years. The only thing that changed is that the market is finally repricing everyone who skipped it.
Alchemist AI Pro is free to use. The delivery evidence, the case study, and the full whitepaper library are public. Read the proof, then reach out: jamie@acc3int.com.
References
A. Historical — Dated, Verifiable, Cited
[H1] Gartner press release, October 20, 2025 — Gartner Identifies the Top Strategic Technology Trends for 2026 (AI-native development platforms; 80%-of-organizations-by-2030 projection). gartner.com/en/newsroom/press-releases/2025-10-20-gartner-identifies-the-top-strategic-technology-trends-for-2026
[H2] Gartner, 2024 — prediction that 30% of generative-AI projects would be abandoned after proof of concept by end of 2025.
[H3] Cohan, P., Forbes, February 6, 2026 — software stocks' loss of over $1 trillion in market value in early 2026; iShares Expanded Tech-Software ETF −21% against a flat S&P 500; Salesforce −26%; SaaS apps per company 112 → 106; 82% of companies reducing suppliers; per-seat pricing 21% → 15% while hybrid pricing rose 27% → 41%.
[H4] CNBC and market coverage, July 14, 2026 — IBM shares down ~23–25% after a Q2 revenue warning ($17.2B vs. $17.86B expected); worst single day since October 1987; software +5%, consulting flat, infrastructure −7%; management citing client spend redirected to hardware.
[H5] Klarna coverage, September 2024 – March 2025 — announcement of dropping Salesforce and Workday for an internally built AI-driven stack (Salesforce Ben; Inc.); subsequent reporting that the replacement blended a proprietary data core with other SaaS tools (CX Today), and the CEO's later caution that most companies would not replace Salesforce with AI (TechCrunch, March 2025).
[H6] Fu et al., 2025 — analysis of 733 AI-generated GitHub code samples; 27.3% carried security weaknesses.
[H7] Methodology and research base as compiled in the ACC3 library: Mavin et al., IEEE RE'09 (EARS); Ferrari et al., RE 2015 and SEI Christel & Kang (Socratic elicitation); IEEE 1044-2009 (anomaly classification); Beyer et al., Site Reliability Engineering, Google, 2016 (four golden signals); Ferraiolo & Kuhn, NIST 1992 (RBAC); Mills, Dyer & Linger, 1987 (Cleanroom); Lehman & Belady, 1985 (software evolution); ISO 9241-11:2018 (usability); Murali Rani et al., 2025 (LLM interpretability 61.9%, reproducibility 52.4%, controllability 47.6%; 81.2% manual verification; 71.9% correction/override); Ouyang et al., ACM TOSEM 2025 (LLM non-reproducibility at temperature zero); 30–100× late-defect cost literature (NIST/Tassey 2002; CISQ 2022).
B. Anecdotal — Observable, Deliberately Uncited
These claims are presented as practitioner observation. They are offered for the reader to verify against their own market view, not as settled research.
[A1] The shared shape of the 2020–2026 stock charts of Pegasystems, ServiceNow, Adobe, Salesforce, and Appian, and the late-2025 divergence from the broader market. The prices are public record; the interpretation — that the market is repricing the model, not the execution — is ours.
[A2] The 2022–2025 point-solution wave: peripheral workloads implemented on premium COTS platforms because the platform was already licensed. Watched by every practitioner in the ecosystem; recorded by no analyst at the time, because it scored as "adoption."
[A3] The CapEx-to-retire-OpEx question now circulating in CIO and CFO conversations, including our own pipeline. No published survey yet quantifies it. This paper's central claim lives here, and we mark it accordingly.
[A4] The strain on pyramid-model consulting economics and client movement toward outcome-based contracts — observable in deal negotiations across the industry, with the July 14, 2026 IBM repricing [H4] as the loudest market echo to date.
C. ACC3 Primary Sources
[P1] The Alchemist AI Pro™ Journey — the without/with comparison across Business Analyst, Developer, and Final Mile roles. acc3int.com/resources/alchemy-journey-slides
[P2] Alchemist Capabilities and Proof — the public proof package, including the DoD Tradewinds Awardable assessment (a procurement-readiness signal, precisely framed). acc3int.com/resources/alchemist-capabilities-and-proof
[P3] Product Campaign Manager case study — 10 features, 395 use cases, 1,086 test scripts, 475 sequenced tasks; 100% feature completeness; 95% requirements adherence; 4 weeks vs. 50 planned; 3–4 people vs. 9 planned; $250K cost vs. $2.27M planned (~$2M saved).
[P4] The ACC3 whitepaper library — acc3int.com/library — in particular: Why Software Modernization Programs Fail (Fernandez); From Vibe Coding to Governed Delivery (Preyna); AI Coding Is Moving Fast. Requirements Are Still Moving Slow. (Brooks); The Known and Costly Problem of Rework (Edwards); From AI Potential to AI Outcomes (Preyna); Enterprise Modernization Beyond Digital Engineering, Part 1 (ACC3 team).
© 2026 ACC3 International / AI Pro Holdings. Alchemist AI Pro™ and the Situational Layer Cake™ referenced above are trademarks of their respective owners.
