AgentsAutonomy & tool use
Celonis Chief Scientist Devavrat Shah Argues Enterprise AI Agents Need an Operational Context Layer
Celonis chief scientist Devavrat Shah argues enterprise AI agents stumble without operational context, and pitches a layered design: process data from systems of record, role-scoped agents within guardrails, and human approval, delivered through the Celonis Context Model launched in May 2026.
From the MIT Technology Review · AI article of October 22, 2026 and cited sources: Shah's central argument is that agents without operational context miss business constraints or spend compute working out what they need to know. [1]
SOURCES[1] Making the case for context-aware AI technologyreview.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Shah's central argument is that agents without operational context miss business constraints or spend compute working out what they need to know. [1]
- Original source
- MIT Technology Review · AI article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 22, 2026 article published
- Covered
- Oct 22 edition
- Evidence
- 7 sources across 7 sites
- 1Making the case for context-aware AItechnologyreview.com
- 2ActuIA — Celonis rachète Ikigai Labsactuia.com
- 3MIT LIDS — Researcher posters recognized at NIPSlids.mit.edu
- 4alphaXiv — PM-LLM-Benchmark (Berti, Kourani, van der Aalst)alphaxiv.org
- 5Constellation Research — Celonis acquires Ikigai Labs…constellationr.com
- 6diginomica — Celonis chief scientist on wrong AI for numbersdiginomica.com
Reach: not measured yet
Shah's central argument is that agents without operational context miss business constraints or spend compute working out what they need to know. His proposed design has three tiers. At the base are systems of record and their event logs. In the middle, an agent-first context layer runs role-limited agents inside guardrails. On top, people review and approve outputs. [1]
Celonis launched the Context Model in May 2026 and announced its acquisition of Shah's company, Ikigai Labs, at the same time. Terms were not disclosed. The Context Model is described as a real-time digital twin of operations. It connects to AWS, Databricks, Microsoft Fabric, Snowflake and ERP systems without copying data, and it plugs into Copilot, Bedrock, watsonx and Claude. In the deal, MIT took Celonis shares in exchange for exclusive patent rights. [5] [7] [2]
For numerical work, Shah says enterprises misuse language models and promotes Large Graphical Models, which model tabular and time-series data. He says they match or beat Google's time-series models at about one-hundredth of the compute or less. This is his own claim; he credits it to NeurIPS research, but the specific paper was not identified. [6]
Shah's deployment advice is to pick one decision, map its process, and apply agents to its hardest step. Teams should measure speed and value before scaling, and clear a standard information-security review of vendor access to core systems first. [1]
the architecture and failure modes are Shah's arguments, with no published benchmark or customer outcome data.
Read the full assessment
Implication: leaders weighing agent deployments can use the narrow-start, role-scoped playbook but should require benchmarks before trusting cost or speed claims.
Executive brief
The MIT Technology Review piece is sponsored content. It was produced by Insights, the publication's custom-content arm, in partnership with Celonis, and the guest is Celonis's own chief scientist, Devavrat Shah. It contains no performance data (MIT Technology Review). Shah argues that AI agents fail in enterprises when they lack operational context: they miss constraints and burn compute working out what they need to know. That argument promotes the Celonis Context Model, launched in May 2026 together with Celonis's purchase of Shah's company, Ikigai Labs. In that deal MIT received Celonis shares in exchange for exclusive patent rights (ActuIA). No independent evaluation of the product was found.
What changed and event timeline
Matrix-estimation forecasting recognized
Shah and co-authors won best poster at the NIPS Time Series Workshop for "Time Series Forecasting via Matrix Estimation". This is early groundwork for the tabular and time-series modeling Ikigai later sold commercially ().
Process-mining LLM benchmark published
RWTH Aachen researchers released PM-LLM-Benchmark: 78 prompts in seven categories. Process-model generation proved much harder than domain-knowledge questions, especially for small models ().
Context Model launched, Ikigai deal announced
Celonis introduced the Celonis Context Model (CCM) and agreed to acquire Ikigai Labs. Terms were not disclosed. Shah became Chief Scientist of Enterprise AI (;).
Shah argues LLMs are the wrong tool for numbers
In an interview, Shah said enterprises misuse language models for tabular forecasting and pitched Large Graphical Models (LGMs) as the alternative ().
Sponsored Business Lab episode
Shah lays out the "decision intelligence" pitch: context, role-scoped agents, and human approval. The publication labels the episode as Insights content, not editorial ().
Capabilities and access
- Described as a real-time "digital twin" of operations. It pulls process data and business context from source systems (Constellation Research).
- Zero-copy links to AWS, Databricks, Microsoft Fabric and Snowflake, plus connectors for Oracle and other ERP/CRM systems (ERP Today).
- Microsoft Copilot, Amazon Bedrock, IBM watsonx and Claude (ERP Today).
Read the full section
- CCM. Described as a real-time "digital twin" of operations. It pulls process data and business context from source systems (Constellation Research).
- Data connections. Zero-copy links to AWS, Databricks, Microsoft Fabric and Snowflake, plus connectors for Oracle and other ERP/CRM systems (ERP Today).
- Agent integrations. Microsoft Copilot, Amazon Bedrock, IBM watsonx and Claude (ERP Today).
- Ikigai tools. aiMatch (data reconciliation), aiCast (prediction) and aiPlan (scenario planning). AI Co-Planner is live; Co-Trader and Co-Marketer were described as coming soon (Constellation Research).
- No version numbers or pricing were found.
Technical analysis for researchers and developers
- Three layers, as Shah describes them. At the bottom, systems of record: workflow tools and their event logs.
- These model relationships among numbers in tabular and time-series data in a "data-cube" fashion, unlike token-sequence models (diginomica).
- No public architecture paper, benchmark or evaluation protocol for the CCM was found.
Read the full section
- Three layers, as Shah describes them. At the bottom, systems of record: workflow tools and their event logs. In the middle, a context model that abstracts those systems into pieces agents can work on, runs "role agents" inside guardrails, and is built "agent-first" so outside agents can connect. On top, humans review and approve outputs (MIT Technology Review).
- LGMs. These model relationships among numbers in tabular and time-series data in a "data-cube" fashion, unlike token-sequence models (diginomica).
- Reproducibility. No public architecture paper, benchmark or evaluation protocol for the CCM was found.
Claims and evidence
- Agents without context make costly mistakes or waste compute. This is an argument from Shah, not a measured result (MIT Technology Review).
- LGMs match or beat Google's time-series models at "100 times or more cheaper compute."
- Some supply-chain cycles cut from months to minutes. This is Ikigai's own claim (ActuIA).
Read the full section
- Agents without context make costly mistakes or waste compute. This is an argument from Shah, not a measured result (MIT Technology Review).
- LGMs match or beat Google's time-series models at "100 times or more cheaper compute." Shah attributes this to NeurIPS research, but the paper itself was not identified (diginomica).
- Some supply-chain cycles cut from months to minutes. This is Ikigai's own claim (ActuIA).
- Customer support. Cardinal Health, an early CCM deployer, is quoted endorsing the idea of context, but no outcomes are reported (ERP Today).
- None of these claims has independent corroboration.
Context and prior work
- Celonis built its business on process mining: reconstructing how work actually flows from system event logs.
- PM-LLM-Benchmark gives an independent view of how well LLMs handle process-mining tasks.
- One cited estimate puts the process-mining market at $1.4B in 2024 (ActuIA).
Read the full section
- Process mining. Celonis built its business on process mining: reconstructing how work actually flows from system event logs. The "context layer" repackages that event-log foundation for AI agents (MIT Technology Review).
- Academic precedent. PM-LLM-Benchmark gives an independent view of how well LLMs handle process-mining tasks. Top commercial models scored about 8.3–8.4 out of 10, and the study found that very small models were not adequate (alphaXiv).
- Market size. One cited estimate puts the process-mining market at $1.4B in 2024 (ActuIA).
Limitations, safety and contested findings
- The episode is vendor-sponsored. MIT holds Celonis equity and Shah is an MIT professor, which entangles academic and commercial interests (ActuIA).
- Deal terms and closing timing were not disclosed (ActuIA).
- No skeptical coverage found. Trade coverage of the launch included no independent skepticism (ERP Today).
Read the full section
- Conflicts of interest. The episode is vendor-sponsored. MIT holds Celonis equity and Shah is an MIT professor, which entangles academic and commercial interests (ActuIA).
- Missing disclosures. Deal terms and closing timing were not disclosed (ActuIA).
- No skeptical coverage found. Trade coverage of the launch included no independent skepticism (ERP Today).
- Security review is a gate. Shah names classic information-security review of vendor access to systems of record as a key step before deployment (MIT Technology Review).
Business and practitioner implications
- Shah's playbook: pick one decision, map its process, apply agents to its hardest step, then measure speed and value before scaling (MIT Technology Review).
- Define clearly what each agent can and cannot do, and budget for training staff to interpret AI outputs.
- Shah warns that pointing LLMs at every problem could create an "even worse zoo" of AI infrastructure than today's data infrastructure (diginomica).
Read the full section
- Start narrow. Shah's playbook: pick one decision, map its process, apply agents to its hardest step, then measure speed and value before scaling (MIT Technology Review).
- Scope agents by role. Define clearly what each agent can and cannot do, and budget for training staff to interpret AI outputs.
- Match model to data. Shah warns that pointing LLMs at every problem could create an "even worse zoo" of AI infrastructure than today's data infrastructure (diginomica).
- Ask for proof. Request independent benchmarks before relying on cost or speed claims.
The source trail.
Sources (7)
Making the case for context-aware AI
Article text retrieved; extracted text may omit tables or interactive elements.
technologyreview.com