TL;DR

• When a frontier model is Tessera-embedded, accuracy rises by +28.9 points on average across five frontier models on a benchmark of 42 ERP-transformation tasks.

• On contamination-proof tasks (designed so the answer cannot be recalled from pretraining) the lift is +64 points. On tasks with publicly available ERP knowledge the lift is about +18 points.

• All models improved, with the biggest jump going to Claude Haiku (+38). The highest Tessera-embedded accuracy went to Claude Opus 4.8 (90.3%).

• Tessera lifts the floor of enterprise capability by supplying an enterprise world model; Claude, working inside it, reaches the highest ceiling.

Enterprise intelligence is not like natural science.

Frontier LLMs have broken boundaries in mathematics, biology, and physics. But enterprise - with its human complexity and tribal knowledge - is a different sort of challenge.

Enterprise value creation requires modeling the fragmented human knowledge layer. Gathering scattered context across hundreds of systems, reconciling conflicting records, understanding the evolution of undocumented processes, and accounting for decades of institutional history when making complex decisions.

Tessera has invested tremendous resources in closing this gap - pairing frontier models with the enterprise world model - context graph - they need to reason well: domain-specialized datasets, process mining, enterprise entity graphs, and curated knowledge bases.

Combining the Tessera context graphs with the right base model for reasoning and orchestration, we have seen outcomes that a general-purpose model cannot generate on its own with enterprise-grade reliability or accuracy.

Model Intelligence is Horsepower; Context Graphs are the Guide

The best people - and correspondingly, the best systems - combine both intelligence and wisdom.

Intelligence is the ability to analyze, reason, generate, and solve problems. Frontier models do this exceptionally well. Given the right information, they can synthesize documents, write software, decompose complex workflows, and coordinate sophisticated sequences of actions. But their intelligence is jagged rather than uniform, and horsepower alone doesn't get you anywhere if you don't know the terrain.

Wisdom is different; it understands the world model. Not just what exists but why — recognizing that two systems which look disconnected actually exchange data through a nightly batch job written a decade ago, and remembering that the last time someone touched it, finance reconciliation broke. Organizations build this context graph slowly. Every implementation, migration, audit, customization, and production incident leaves behind another small piece of context, and it compounds: the more history you carry, the better you read the next situation.

Figure 1: Enterprise Context Graph

Tessera's critical component is this context graph. A living map that links schemas, code, SOPs, and systems - one that deepens with every interaction. The agentic harnesses on top provide a working and comprehensible model for AI to leverage in practice.

Intelligence is horsepower, but wisdom and context are a map—recording memory of every road already traveled, with the judgment to drive it. Tessera harnesses the intelligence available and provides the wisdom.

Our System Architecture

We begin with a simple question: what actually exists?

In a large enterprise, the answer is spread across source code, configuration tables, database schemas, documentation, and the operational knowledge of people who have lived through prior transformations. Our platform architecture brings together — and converts into a durable asset — these various components.

Tessera unifies these fragments into one working model. It connects to systems like SAP, Oracle, Salesforce, and Workday, converts their raw artifacts into structured form, and links them into an entity graph — objects within a system and across systems, including migration mappings, shared processes like Order-to-Cash, and the custom programs that touch each entity. On top of that model, specialized agents plan and act, drawing on both the current session and an organizational memory that sharpens with every engagement, plus a curated knowledge base of practitioner expertise that general models don't have.

The full architecture including ingestion, the enterprise entity graph, the domain-expert knowledge base, and the multi-agent reasoning layer is detailed in our technical whitepaper.

As the Tessera platform came together, we needed a way to measure it. Traditionally, frontier models have been benchmarked on tasks such as coding, web search, and computer use, but not on the practical enterprise workflows that our systems handle at scale every day during transformation projects. Thus, we built our own benchmark.

Figure 2: Knowledge Graph

TILES: The Enterprise Benchmark

TILES, our enterprise benchmark, contains forty-two tasks drawn from real transformation work across seven suites: SAP ABAP Coding, SAP Configuration Analysis, Finance Schema Modeling and Cleansing, Manufacturing Schema Modeling and Cleansing, Cross-System Schema Reconciliation, Tabular Data Prediction, and Test Requirement and Test-Data Generation.

Each of the forty-two tasks were also categorized by class.

Contamination-proof tasks are tasks where the answer exists only inside the customer's environment not seen in public training data. A raw model cannot know these answers in advance. It can only reason from given context.

Public tasks depend on general ERP knowledge. A strong model may already know the answer from public documentation, or at least have seen enough similar material during pretraining to make an informed attempt.

Control tasks are tasks where Tessera's knowledge base should not help. They are included as a guardrail. If embedding improves them too much, something is probably wrong: the system may be adding verbosity, leaking information, or shifting the grading rather than supplying useful knowledge.

We evaluated five frontier models in two configurations. In the first, the model was run directly on the task. In the second, the same model was embedded with Tessera's agent layer. Every task is graded against an expert-authored rubric where we decompose a correct answer into discrete, verifiable claims and score the fraction satisfied, using a three-judge LLM ensemble (majority vote; judges disjoint from the models under test).

Across all seven suites and all five models, embedding the model in Tessera improved task accuracy by an average of 28.9 percentage points:

• On contamination-proof tasks, the lift was 64.3 points.

• On public tasks, the lift was 18.0 points. That is still meaningful, but it is smaller for the reason we would expect. Frontier models already know a fair amount about standard SAP behavior and published ERP concepts. Tessera supplies relevant domain context and reduces ambiguity, but there is less missing knowledge to supply.

• On control tasks, the lift was 0.5 points and not statistically significant. That is the result we wanted to see. Tessera does not improve tasks that do not need Tessera.

Full multi-agent architecture and per-suite analysis in our technical whitepaper.

Figure 3: Lift by task class.

We also analyzed the lift across suites. Tabular Data Prediction improved by more than sixty points. Test Requirement and Test-Data Generation improved by more than forty-five. SAP Configuration Analysis and Cross-System Schema Reconciliation both improved by more than thirty.

Figure 4: Knowledge-base lift by task suite.

A concrete case. We asked a model to review a custom ABAP routine that runs after an S/4HANA conversion. On its own, even a strong model misses that an INSERT into the MSEG table now silently does nothing — in S/4HANA those tables became compatibility views over MATDOC, so goods movements vanish without an error. It's not a reasoning failure; the model has no way to know this specific system's behavior. Embedded in the knowledge base, the same model flags the silent data loss, cites the governing SAP Note, and recommends the fix. That is the difference between horsepower and knowing where to go.

Tessera is model-agnostic by design. The enterprise graph, knowledge base, and agent framework can embed different frontier models, with the benchmark illustrating improvement in every model.

But model-agnostic does not mean model-indifferent.

Tessera supplies the harnesses and enterprise knowledge, but leverages a frontier model for orchestration. Our experiments indicated that all else being equal - Claude was the most effective user of Tessera's context graphs.

Why Building With the Right Model Matters

Embedding the model in Tessera delivers a large, significant lift for every engine we tested. But model choice still shapes the outcome in two ways:

• The smallest model gained the most. Claude Haiku 4.5, the lightest and least expensive engine, jumped +38 points with embedding.

• The strongest model sets the ceiling. Claude Opus 4.8 reaches the highest embedded accuracy of any engine tested (90.3%), with Claude Sonnet 5 just behind, because a stronger reasoner exploits the same knowledge more fully.

Figure 5: Raw vs. embedded accuracy by model.

Tessera is able to give the model a working picture of the enterprise ecosystem, including factors like how the systems depend on each other and what changed over time.

The frontier LLM used in Tessera leverages that context to execute work. It reasons across the evidence, maintains the thread through long agentic workflows, follows constraints, uses tools, and produces recommendations that a human expert can inspect.

Our experiments showed Tessera and Claude together to be most complementary, allowing the agentic system to understand and reshape the sprawling systems a business depends on.

The broader lesson is architectural: enterprise AI performance comes from aligning frontier model intelligence with the wisdom and context embedded in the platform around it.

The Future of Transformation as Wisdom and Context Compound

Tessera is not stateless. Each interaction is a deposit into an organizational memory that compounds over time.

The platform learns which tables matter to this specific enterprise, which processes depend on them, which risks appear in real migrations, and which recommendations human experts accept or correct. That knowledge becomes part of the enterprise model.

If Tessera discovers that a custom ERP table is not just used for reporting, but also feeds downstream finance reconciliation - the fact is useful immediately. But it is also useful later: when slicing migration data, generating tests, assessing risk, or tracing code impact. The next transformation does not have to rediscover the tribal knowledge from scratch.

The advantage of memory compounds; it is easy to adopt newer, stronger models. But you cannot recreate a history accumulated over years.