Executive Summary
By Paul Allard, MBA, CEO at Persevere Consulting — November 27, 2025
The promise of AI is simple: automate routine work, augment decision-making, and free humans for higher-value tasks. But as recent reporting in HBR shows, AI can also generate "work-slop" — noisy, low-value outputs that create more rework than they eliminate.
Complementing that critique, Dr. Claude Coulombe's article "GPT-5 — a damp squib? LLMs reaching a plateau" argues that foundation models may be hitting a performance ceiling, and that incremental model upgrades alone will not solve practical enterprise problems.
Together, these views point to the same conclusion: productivity gains depend less on headline model versions and more on data quality, governance, sovereignty controls, and operational integration.
Why AI Sometimes Reduces Productivity
AI systems can deliver large gains — but often produce:
- Hallucinations and incorrect outputs that require human correction.
- Inconsistent results from identical prompts due to model drift or varying fine-tuning.
- Tool sprawl and duplicated efforts as teams assemble ad-hoc solutions.
- Compliance and sovereignty headaches when data is used without proper controls.
Dr. Coulombe's analysis emphasizes another root cause: chasing marginal model improvements without addressing systemic issues leads to disappointment. If models are plateauing in capability, the leverage shifts decisively to data, integration, and governance.
Data Governance and Sovereignty: The Productivity Levers
Three governance failures drive work-slop:
1. Poor Provenance and Lineage
Without lineage, teams can't verify what produced an output, whether cleansing occurred, or which biases exist — leading to repeated validation and rework.
2. Inconsistent Access and Stewardship
Ad-hoc copying into local notebooks and multiple competing models arise when stewardship roles and metadata standards are absent.
3. Cross-Border and Jurisdictional Ambiguity
Data sovereignty constraints complicate training and fine-tuning; ignoring them can block deployments and trigger lengthy remediations.
If improved model architecture is not delivering proportional value, clean, well-governed data and robust deployment controls become the primary path to productivity.
Why You Need Sovereign AI
Gen AI is not optional to fix the productivity gap. Canadian organizations cannot rely 100% on foreign proprietary Gen-AI solutions to process intellectual property, confidential information, or sensitive personal information:
- Physics beat Law: Contractual safeguards ≠ technical safeguards
- Residency ≠ sovereignty: local hosting does not prevent foreign access
- Dependence on foreign technology is a business continuity risk: weakened bargaining power, technological lock-in, geopolitical and supply chain risk
- A chain is only as strong as its weakest link: the need for a 100% open-source and auditable Gen-AI stack
How to Prepare for a Sustainable AI Deployment
1. Strategic Blueprint & Risk Assessment
Identify your organization's unique competitive competencies. For each, map out the associated processes and clearly delineate where AI-driven innovations can tolerate initial mistakes versus those critical operations where absolute reliability is paramount.
2. Invest in Data Excellence
Data is the decisive lever. Prioritize investments in assessing and enhancing the readiness of both structured and unstructured data. Establishing robust data lineage, quality, and governance protocols is essential to combat the GIGO effect.
3. Nurture Internal Leadership
Identify and empower your internal AI champions. Provide them with the motivation, resources, and training necessary to lead the charge. These individuals will be instrumental in mobilizing subject matter experts and driving AI adoption.
4. Build a Sovereign AI Foundation
Implement a private and sovereign AI infrastructure. This is not merely a technical choice but a strategic imperative to protect your intellectual property, mitigate geopolitical and supply chain risks, and ensure the long-term integrity and trust in your AI-driven productivity gains.
A Practical Four-Lens Framework
To assess AI's net productivity impact, evaluate across four dimensions:
| Lens | Key Actions |
|------|-------------|
| People | Defined stewardship roles, training in prompt design, human-in-the-loop validation |
| Process | Validation gates for outputs, incident response for hallucinations/exfiltration, user feedback loops |
| Data Governance | Policies for provenance, access controls, regional enforcement, vendor SLAs, Sovereign AI infrastructure |
| Technology | Central catalog with lineage, model registry, CI/CD with drift monitoring, regionally constrained pipelines |
Mini Case Vignettes
Failure: Marketing fine-tuned a vendor model with a scraped dataset lacking consent metadata. Outputs breached consent terms, requiring legal remediation and campaign rollbacks.
Success: A healthcare consortium used a Private Sovereign AI infrastructure to pool de-identified patient data with strict provenance and use policies. Validated inputs shortened model validation cycles and accelerated clinical insights.
KPIs to Measure Progress
- Reduction in time spent fixing AI outputs
- % of models with verified provenance
- Number of sovereignty incidents per quarter
- Time from model discovery to validated deployment
Conclusion
Can AI save productivity? Yes — but not by relying solely on new model releases.
Treat data governance as a productivity investment. Build your Sovereign AI infrastructure to protect and grow your IP. Focus less on chasing model hype and more on cleaning the data plumbing, enforcing provenance, and baking in human oversight — then AI becomes a durable productivity engine rather than a source of rework.
Disclaimer: AI contributed to the creation of this article, but it was guided, reviewed and fact-checked by Persevere Consulting's human experts. The content is for general information purposes only and is not legal or management advice.