Executive Synthesis
The cloud ERP migration wave (2015–2024) and the current introduction of AI into COTS and SaaS packages (2023–2026) share three deep structural parallels: both promise operational transformation but deliver it only when the operating model is redesigned around the technology; both generate a Year-3 cost trap where subscription and consumption economics diverge sharply from the original business case; and both create vendor dependency that compounds over time and becomes progressively harder to escape.
The three places the analogy breaks down are equally important. First, the failure mode is different: cloud ERP fails slowly through process rigidity and cost escalation; AI-in-COTS fails fast through hallucination, non-determinism and governance gaps that can produce audit-material errors in days, not quarters. Second, the lock-in mechanism is categorically new: cloud ERP created workflow lock-in (muscle memory, process dependency); AI-in-COTS creates cognitive lock-in (the vendor’s semantic index becomes your institutional memory, and that memory cannot be exported). Third, the competitive threat is asymmetric: no one built a credible ERP from scratch during the cloud migration wave; the post-LLM disruptor building custom finance workflows on foundation-model APIs is a live threat to the ERP ecosystem today.
The single strongest implication for boards: do not sign multi-year AI-embedded-ERP commitments until you have ring-fenced your organisational data graph in infrastructure you control. The cognitive lock-in compounding inside every vendor’s AI layer is the sharpest strategic risk in enterprise technology today, and it is one boards are not yet pricing into renewal decisions.
Parallel-by-Parallel Analysis
1. The Productivity Promise vs. the Operating-Model Prerequisite
Cloud ERP Precedent
Cloud ERP vendors promised real-time reporting, close compression and embedded analytics. These benefits materialised—but only where the operating model was redesigned around the technology. At Hotelbeds, close compression was material; at Awaze, NetSuite consolidated finance across three countries. In both cases the benefit depended on harmonising master data and reengineering processes before go-live, not on the software alone.
AI-in-COTS Analog
AI vendors promise productivity gains of 100+ hours per user per year (Forrester TEI, 2025). Microsoft reports 70% of Fortune 500 companies have deployed Copilot. But only 3.3% of Microsoft 365 users have the paid add-on, and 74% of companies cannot yet show measurable ROI. Stanford’s AI Index 2026 confirms fewer than 10% have fully scaled AI in any single business function.
Evidence Base
Forrester TEI (2025): 108 hours/year productivity gain, 116% ROI projected for large enterprises. Stanford AI Index 2026: 88% adoption, <10% full-scale deployment. Gartner (Aug 2025): 40% of agentic AI projects will be cancelled by end-2027 due to escalating costs or unclear value. Panorama Consulting (2024): 75% of ERP users need reskilling, only 35% receive adequate training.
Strength of parallel: Strong. Both waves deliver the technology faster than the organisation can absorb it. The operating-model redesign is the binding constraint in both cases.
Where the analogy breaks: Cloud ERP required months of process redesign before go-live. AI-in-COTS can be switched on with a licence toggle, which means the gap between deployment and readiness is wider and more dangerous—users start generating AI-assisted outputs before governance, training or quality controls are in place.
Board action (30 days): Commission a readiness audit across the top three AI-embedded applications. Score each on data quality, process governance and user training. Gate further AI licence expansion on readiness scores, not vendor timelines.
2. The Year-3 Cost Trap
Cloud ERP Precedent
SaaS subscription pricing escalates aggressively from Year 3. SAP RISE contracts contain FUE licensing that masks consumption inflation, mandatory BTP extension costs (€1.2–1.8M for mid-sized estates), and effective lifetime commitment clauses. Senior CIOs across Spain confirm the pattern: RTVE rejected RISE on cost and lock-in; Repsol reported cost surprises and DRP failures; when three negotiation clauses are absent (CPI cap, BTP credits, FUE ceiling), the Year-3 cost trap can erase 30–60% of projected savings.
AI-in-COTS Analog
Consumption-based AI pricing transfers compute-cost volatility to customers. A 2025 Zylo study found 78% of IT leaders report unexpected charges from AI pricing models. IDC confirms 46% cite pricing unpredictability as the primary obstacle to gen AI deployment. AI-native spending nearly doubled YoY to $1.2M average per organisation (2025), and 85% of organisations misestimate AI costs by more than 10%.
Evidence Base
Zylo/CIO.com (2025): 78% unexpected AI charges, 85% cost misestimation >10%. IDC (2025): 46% cite pricing unpredictability as primary AI obstacle. CIO Dive (2025): AI cost overruns adding up with major CIO implications. Tendam/RISE analysis (2025): Year-3 cost trap erases 30–60% of projected RISE savings.
Strength of parallel: Strong. The economic pattern is structurally identical: low initial deployment cost, aggressive Year-3 escalation, and a switching cost that makes renegotiation one-sided.
Where the analogy breaks: Cloud ERP cost escalation is predictable (annual price increases, module additions). AI cost escalation is stochastic—consumption scales with adoption, and adoption is not centrally controlled. A single department discovering an AI use case can spike the bill.
Board action (30 days): Model Year-3 and Year-5 AI total cost of ownership across all embedded AI products. Negotiate three clauses into every AI vendor contract: consumption ceiling with overage cap, CPI-linked price escalation limit (2%), and model-substitution rights (the right to switch underlying models without renegotiating the contract).
3. Vendor Lock-In: From Workflow to Cognitive
Cloud ERP Precedent
Enterprise vendor dependency passed through three generations: technical lock-in (1990s, proprietary formats), contractual lock-in (2000s, volume licensing), and workflow lock-in (2010s, process muscle memory). Gartner’s 2024 data confirms only 39% of ECC customers have migrated to S/4HANA despite a 2027 deadline—a direct measure of how deep workflow lock-in runs.
AI-in-COTS Analog
The fourth generation is cognitive lock-in: the vendor’s semantic index, agent configurations and accumulated context become the institutional memory of the finance function. SAP’s Joule, Workday’s embedded AI, Oracle’s Fusion AI agents and Salesforce’s Agentforce are all designed to accumulate context inside the vendor stack. Every interaction widens the gap between the incumbent’s understanding of your function and any alternative provider’s understanding. The switching cost is no longer data migration or contract penalties—it is intelligence loss, and intelligence loss is non-exportable.
Evidence Base
Copilot Disruption briefing (Feb 2026): four-generations lock-in framework. Gartner (Q4 2024): 39% S/4HANA adoption rate despite 2027 cliff. Kai Waehner analysis (Apr 2026): enterprises risk embedding agentic architecture into vendor runtime, governance and observability stacks. DSAG Investment Survey (2026): only 3% of SAP customers run SAP Business AI in production.
Strength of parallel: Partial—the pattern is analogous but the mechanism is categorically new. Workflow lock-in was escapable with effort (retraining, process redesign). Cognitive lock-in compounds autonomously: the AI learns more about your organisation every day without anyone deciding to deepen the dependency.
Where the analogy breaks: Cloud ERP lock-in was visible and negotiable. Cognitive lock-in is invisible and non-negotiable—there is no contract clause that exports a vendor’s accumulated understanding of your finance function.
Board action (30 days): Mandate a multi-model policy: no single LLM provider may exceed 60% of enterprise AI workloads. Ring-fence the organisational data graph in a cross-platform layer you control (e.g., Databricks, Snowflake, or a self-hosted vector store). Treat agent configurations and AI-accumulated context as exportable assets in every vendor contract, even when the vendor offers no native export path.
4. The Standardisation–Differentiation Tension
Cloud ERP Precedent
Cloud ERPs ship standardised, industry-templated processes that fit 70–80% of what a finance team does. The remaining 20–30% is where friction lives. At Hotelbeds, microservices were built around SAP for dynamic pricing because the platform could not move at the required cadence. Under RISE, “clean core” restrictions now limit this option.
AI-in-COTS Analog
AI-in-COTS delivers vendor-trained models optimised for generic use cases. Fine-tuning on proprietary data is possible but expensive; running domain-specific models alongside vendor-embedded AI creates integration complexity. The tension is the same: standardise where you are average, differentiate where you compete—but the mechanisms differ.
Strength of parallel: Strong. The trade-off is structurally identical.
Where the analogy breaks: ERP customisation is deterministic (you configure a workflow, it behaves as configured). AI customisation is probabilistic (you fine-tune a model, it behaves approximately as trained). The confidence interval around “customised” behaviour is wider with AI.
Board action (30 days): Map all AI-embedded workflows by criticality. For high-stakes processes (financial close, audit, regulatory reporting), mandate human-in-the-loop and deterministic rules engines alongside AI assistance. Reserve AI autonomy for low-consequence, high-volume tasks.
5. Change Management Debt
Cloud ERP Precedent
Training, role redesign and cultural adjustment are slower and more painful than technology delivery. Panorama (2024): 75% of employees need reskilling; only 35% receive adequate training. At Hotelbeds, the finance team post-migration was smaller in the back office but stronger in pricing strategy—requiring fundamentally different skills.
AI-in-COTS Analog
AI changes what people do faster than organisations can retrain them. Jason Lemkin’s claim that “10x engineers are now 100x engineers with AI” has a finance-function equivalent: the FP&A analyst with strong data and AI literacy is several times more productive. The skill gap is the binding constraint on value capture in both waves.
Strength of parallel: Strong. Both waves create a talent asymmetry where early adopters pull away and laggards lose the ability to catch up.
Where the analogy breaks: ERP change management had a defined endpoint (go-live plus stabilisation). AI capabilities evolve continuously—model updates, new agent capabilities, shifting interfaces. Change management becomes permanent, not project-bounded.
Board action (30 days): Allocate a minimum of 20% of AI programme budget to training and role redesign. Create an “AI fluency” competency framework for finance and mandate quarterly skill assessments.
6. Data Quality as the Binding Constraint
Cloud ERP Precedent
Cloud ERP is a data-governance project wearing a technology coat. At Awaze, three regional Wyndham instances with different revenue-recognition methods had to converge before any reporting benefit was realised. Poor master data on day one undermined every benefit the business case promised.
AI-in-COTS Analog
Stanford AI Index 2026: the primary barrier is not the AI model but the data infrastructure supporting it. Organisations lacking governed data pipelines, reliable integration layers and quality frameworks cannot move AI from pilot to production. AI amplifies data quality problems: a misconfigured allocation rule in ERP propagates silently for a quarter; a hallucinating AI model on dirty data produces plausible-sounding errors that are harder to detect.
Strength of parallel: Strong. Data quality is the prerequisite in both waves, and both waves are systematically underinvesting in it.
Where the analogy breaks: ERP data quality problems are deterministic and discoverable (run a reconciliation, find the mismatch). AI data quality problems are probabilistic and may only surface as subtle output degradation that passes human review.
Board action (30 days): Before expanding AI deployment, run a data-quality audit on every data source feeding AI models. Establish a data stewardship function (if absent) with explicit accountability for AI input quality.
7. The “Two-Speed” Problem
Cloud ERP Precedent
Operational accounting becomes near-real-time while group consolidation and statutory reporting lag. Non-finance managers see a real-time operational number; corporate finance reports a reconciled, audited number that is different. Without explicit governance of which number is used for which decision, internal arguments about “the truth” follow.
AI-in-COTS Analog
AI produces instant analysis and recommendations; human governance processes (review, approval, audit) operate on a different cadence. The gap between AI speed and governance speed is the new “two-speed” problem. An AI agent can generate a financial forecast in seconds; validating it against audit standards takes days.
Strength of parallel: Partial. The structural pattern (fast system, slow governance) is the same. The magnitude is larger with AI because the speed differential is orders of magnitude greater.
Where the analogy breaks: ERP’s two-speed problem was about data latency. AI’s two-speed problem is about judgement latency—the system moves faster than human oversight can verify.
Board action (30 days): Define a “golden source” policy for AI-assisted outputs: which outputs require human sign-off before use in external reporting, and which can flow through with automated validation. Publish the policy to all budget holders.
8. Sovereignty and Jurisdictional Risk
Cloud ERP Precedent
Data residency, sovereign cloud and US CLOUD Act exposure moved from theoretical footnotes to live board-level discussions for European businesses in 2025–2026. RISE’s DRP model does not cover client custom integrations.
AI-in-COTS Analog
AI introduces additional sovereignty dimensions: model training data provenance, cross-border inference calls, and the question of who owns the fine-tuned weights. The EU AI Act (effective August 2025) imposes obligations on high-risk AI systems that most COTS vendors have not yet fully operationalised for their embedded AI features.
Strength of parallel: Strong. Both waves push enterprise data and logic into US-headquartered vendor infrastructure, creating identical jurisdictional exposure.
Where the analogy breaks: ERP sovereignty was about data at rest. AI sovereignty includes data in motion (inference calls), data in training (model weights), and emergent behaviour (agent decisions). The surface area is larger.
Board action (30 days): Require every AI vendor to provide a data-processing map showing where inference occurs, where training data resides, and which model weights are shared across tenants. Flag any cross-border inference flows for legal review under EU AI Act requirements.
9. The Pricing Model Transition
Cloud ERP Precedent
ERP moved from perpetual licence to subscription. The transition created budget predictability for vendors and cost escalation for customers. SAP’s FUE model and Salesforce’s per-seat pricing are structurally designed to grow with usage.
AI-in-COTS Analog
AI is triggering a second pricing transition: from subscription to consumption. Salesforce Agentforce offers three models (Flex Credits at $0.10/action, per-user at $125+/month, and outcome-aligned contracts). IDC forecasts by 2028, 70% of software vendors will refactor pricing around consumption or outcome metrics. The transition transfers compute-cost volatility to customers.
Strength of parallel: Strong. Both are vendor-initiated pricing transitions that transfer risk to customers under the guise of “flexibility.”
Where the analogy breaks: Subscription pricing was predictable if aggressive. Consumption pricing is genuinely unpredictable: 85% of organisations misestimate AI costs by more than 10%, and nearly a quarter are off by 50%+.
Board action (30 days): For every AI product on consumption pricing, implement a spend cap with automated alerts at 80% utilisation. Negotiate a contractual overage ceiling (e.g., 120% of estimated annual consumption) with clawback rights if the vendor’s model degrades in accuracy.
10. The Structural Threat from Post-LLM Disruptors
Cloud ERP Precedent
No credible challenger built a new ERP from scratch during the cloud migration wave. The incumbents (SAP, Oracle, Microsoft, Workday) retained their installed bases. New entrants competed at the edges (niche planning tools, AP automation) but never threatened the core system of record.
AI-in-COTS Analog
The competitive landscape is structurally different. Rory O’Driscoll’s three-player framework identifies: the pre-AI incumbent (SAP, classical Oracle), the AI-native teenager (built AI in from 2018), and the post-LLM disruptor (custom workflows on foundation-model APIs). Amjad Masad argues systems of record survive; point-solution and vertical SaaS tools face replacement by custom apps built atop APIs. The surrounding ecosystem of niche tools is in the kill zone.
Strength of parallel: Weak—and this is where the analogy is most dangerous. The ERP precedent suggests incumbents are safe. The AI reality is that the workflow and intelligence layers above the system of record are now contestable by entrants that did not exist two years ago.
Where the analogy breaks: During the cloud ERP wave, building a competing system required massive capital and years of development. Post-LLM, a well-resourced team can build a domain-specific workflow replacement in months using foundation-model APIs. The barrier to entry collapsed.
Board action (30 days): Audit the “surrounding stack” (point solutions, niche planning tools, AP automation, expense workflows) against the Masad thesis. For each tool, ask: is this a system of record with a defensible data moat, or a workflow tool a well-prompted agent could replace? Begin piloting replacements for the weakest links.
Counter-Narrative: Where the ERP Analogy Actively Misleads
1. Speed of Harm
The ERP analogy implies that problems surface slowly and can be fixed during stabilisation. AI problems surface immediately and can cause material damage before anyone notices. A misconfigured ERP allocation rule propagates silently across a group for a quarter. An AI hallucination in a financial close package can produce a plausible-sounding but fictitious number that passes human review because it “looks right.” The 2024 finding that LLMs hallucinate in up to 41% of finance-related queries is not a rounding error—it is a structural risk that has no ERP precedent. The correct mental model is not “ERP implementation risk” but “operational risk with non-deterministic systems,” closer to model risk management in banking than to IT project management.
2. The Compounding Nature of Cognitive Lock-In
The ERP analogy implies that lock-in stabilises once migration is complete. Cognitive lock-in does not stabilise—it compounds. Every meeting summarised by the vendor’s AI, every document indexed, every agent action executed widens the incumbent’s knowledge advantage. The switching cost grows daily without any deliberate decision to deepen the dependency. The correct mental model is not “contract lock-in” but “knowledge compound interest”: the longer you wait, the more expensive it becomes to leave, and the rate accelerates rather than decelerating.
3. The Barrier-to-Entry Collapse
The ERP analogy implies that incumbent vendors have structural moats. For systems of record, they do. For workflow and intelligence layers, the moat has been breached. The post-LLM disruptor can build a domain-specific workflow in months, not years, using foundation-model APIs. Tuhin Srivastava of Baseten notes most enterprises are still in the “first-generation CoPilot phase” with genuine agentic adoption 12–18 months out. ERP vendors betting on AI agents working flawlessly inside their suite by mid-2026 are betting against the consensus of the people building the stack. The correct mental model is not “ERP vendor competition” but “platform-versus-API competition,” where the question is whether the incumbent’s bundled suite can outrun the open ecosystem’s composability. History—from IBM mainframes to Salesforce’s rise to the current LLM wave—does not favour the bundled incumbent once the API layer matures.
Source Bibliography
Stanford HAI. AI Index Report 2026. April 2026. https://hai.stanford.edu/ai-index/2026-ai-index-report
Gartner. “Quick Answer: What Is the Level of Adoption of SAP S/4HANA in Q424?” Q4 2024. Via Rimini Street.
Gartner. “Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026.” August 2025.
Gartner. “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027.” June 2025.
Forrester. “The Projected Total Economic Impact of Microsoft 365 Copilot.” 2025. tei.forrester.com
Panorama Consulting Group. 2024 ERP Report. 2024.
Zylo. “AI Pricing: What’s the True AI Cost for Businesses in 2026?” 2025. zylo.com
IDC. Survey on AI pricing unpredictability as deployment obstacle. 2025. Via CIO.com.
CIO.com. “AI cost overruns are adding up—with major implications for CIOs.” 2025.
DSAG. Investment Survey 2026: 3% of SAP customers run SAP Business AI in production.
Salesforce. “Salesforce Introduces New Flexible Agentforce Pricing.” May 2025. salesforce.com
Kai Waehner. “Enterprise Agentic AI Landscape 2026: Trust, Flexibility, and Vendor Lock-in.” April 2026.
Rory O’Driscoll, Scale Venture Partners. 20VC podcast, March 2026.
Amjad Masad, Replit CEO. 20VC podcast, April 2026.
Tuhin Srivastava, Baseten. No Priors podcast, April 2026.
Jason Lemkin, SaaStr. 20VC podcast, March 2026.
Brian Chesky, Airbnb CEO. Invest Like the Best / Prof G Pod, May 2026.
De Nicolás, Á. “RISE with SAP – Tendam Analysis.” Revamp Advisors, 2025.
De Nicolás, Á. “Copilot Disruption: Strategic Briefing.” Revamp Advisors, February 2026.
BizTech Magazine. “LLM Hallucinations: Implications for Financial Institutions.” August 2025.
JurisTech. “Best LLM Tools for Financial Analysis 2026: Hallucination Benchmark Report.” 2026.