Deployed AI Doesn’t Learn. We Fixed That For Payroll Data.

Executive summary

Emmanuel Dupoux, Yann LeCun, and Jitendra Malik, writing from META FAIR, NYU, and UC Berkeley, published a paper with a blunt claim: AI models learn nothing once deployed. According to the paper, the learning happens in the MLOps pipeline, through data scientists and retraining cycles, while the deployed model stays fixed.

The authors propose a theoretical architecture to fix the problem, with learning from observation, learning from action, and autonomous switching between the two modes. By the authors’ own admission, the architecture is decades from fruition.

datascalehr has already built a version of the architecture the authors describe and runs that version in production for global payroll data.

The central insight

The paper identifies “domain mismatch” as a core failure mode. AI trained on internet-scale data fails in specific environments because real-world data contains unseen cases and keeps changing, and a larger training set does not fix the failure. In global payroll, domain mismatch is the whole problem. datascalehr designed KMod™ to solve domain mismatch by learning continuously from the decisions payroll specialists confirm every day.

Source: Dupoux, LeCun, and Malik (2026), “Why AI systems don’t learn and what to do about it,” META FAIR / NYU / UC Berkeley. arXiv:2603.15381v1.

Why the paper supports datascalehr’s architecture

The paper describes four properties an AI system needs to learn from its environment after deployment. Each property maps to how KMod™ works in production today:

1. Domain mismatch requires domain-specific learning

Anyone who has connected a Workday instance to a German payroll engine knows the domain mismatch problem. Every country has different compensation components, field names, and formats. A model trained on generic HR data fails on its first Brazilian payroll file. Even after a team fixes the model for Brazil this quarter, the Brazilian rules change next quarter.

In datascalehr, KMod™ learns from the decisions payroll specialists confirm every day: which source field maps to which target field, how a format converts, and where a country rule applies. Each confirmed correction is captured at once and is available across every tenant. When a specialist in Munich corrects a mapping at 9:00 AM, a colleague in Vienna gets the improved suggestion at 9:01, with no retraining and no batch job.

2. Learning from action and observation

In datascalehr, KMod™ learns from observation when KMod™ reads the structure of each incoming file, and learns from action when a payroll specialist confirms or rejects each suggested mapping.

3. Alignment hacking and active data selection

In datascalehr, every learning event requires human validation, and the architecture is built around that requirement. Only validated mappings enter the knowledge model. A user must confirm or correct every suggestion before the suggestion becomes a rule. Because payroll errors cost money and break compliance, datascalehr makes the payroll specialist part of the learning cycle.

4. Evolutionary curriculum: progressive complexity

The paper proposes an “evolutionary curriculum” in which the environment gradually grows more complex, so the system moves from simple environments to harder ones as competence increases.

In datascalehr, new clients start with their own data patterns (local knowledge). As the system processes more integrations, KMod™ draws on anonymized patterns from the whole client base (federated knowledge). The learning cascade moves from simple local patterns to complex cross-jurisdiction patterns through the normal progression of enterprise deployments.

What the research proposes and what KMod™ delivers

The researchers describe a future architecture for AI that learns autonomously. datascalehr built a domain-specific version of the architecture, constrained by human validation and running in production.

The research agenda (decades away)

• General-purpose autonomous learning across any domain
• Learns like children: open-ended and domain-agnostic
• Theoretical, proposed as a multi-decade research program
• Alignment risk remains open, and researchers must design the curriculum

KMod™ in production today

• Domain-specific continuous learning for payroll data
• Learns like a payroll expert: constrained, validated, and compounding
• In production: 180 countries, 1,400+ connectors, 4 of top 5 global PSPs
• Alignment handled: a human validates every learning event
• Curriculum emerges from natural enterprise deployments

How KMod™ learns in production

Each integration is a learning event, and each confirmed correction adds to what KMod™ knows. The learning loop below shows how datascalehr’s AI improves with every deployment:

1. Ingest any file in any format

datascalehr reads each file by meaning as the file arrives, instead of forcing the file into a fixed format. The platform learns file structure, headers, data types, and hierarchical relationships automatically, without predefined connectors. Workday PECI files, SAP HCM extracts, CSV exports, and PDF payslips all go through the same process.

2. KMod™ predicts the mapping

The four-layer learning cascade runs in order: local knowledge first (your data patterns), then proprietary algorithms, then federated knowledge (anonymized patterns across all clients), then an LLM as a constrained fallback. datascalehr’s deterministic algorithms handle 91% of cases, covering mapping, validation, and transformation, which keeps the approach predictive, not generative. The LLM proposes a mapping, and the predictive stack decides.

3. A human validates or corrects

A payroll specialist confirms or corrects the suggested mapping. The specialist’s confirmation or correction is the learning signal KMod™ trains on, and each decision enters the knowledge model immediately.

4. Knowledge compounds immediately

The validated mapping is available immediately to every user, tenant, and country, with no retraining cycle and no batch job. A correction in Munich at 9:00 AM improves suggestions in Vienna at 9:01 AM, so each deployment starts with more knowledge than the deployment before.

Production beats theory

Fully autonomous, domain-agnostic AI learning is worth pursuing. Companies that have made AI learn within one domain, with specific constraints and human supervision, are delivering value now.

For global payroll directors:

Stop waiting up to 12 months for each new country integration. KMod™ has already learned mapping patterns across 180 countries and more than 11,500 schemas, and setup takes under an hour per country.

For payroll service providers:

Implementation margins erode with every new client. Because datascalehr’s AI keeps learning from expert-confirmed decisions, client #50 takes a fraction of the effort of client #1.

For CIOs and integration architects:

Traditional iPaaS tools and unified APIs are not built for the semantic complexity of payroll data. The context layer adds the payroll domain knowledge that a general integration stack lacks.

For HR technology vendors:

Your customers need payroll connectivity across 50+ countries. datascalehr’s API and MCP connector let your platform reach any payroll engine without building country-by-country integrations.