The momentum of Claude Cowork marks another step in AI’s move from chat-based assistance to hands-on business tasks, raising a question for ambitious enterprises: what happens when general-purpose AI is applied to the real-life, high-stakes workflows businesses run every day?
We evaluated Claude Cowork against real finance use cases requiring massive spreadsheets, PDFs, repeatable data preparation, multi-step processes, maintenance, and governance. That includes reconciliations, consolidations, accruals, sales tax, transfer pricing, and broader month-end close and tax provisioning workflows.
Our conclusion was that, although the AI is remarkably capable, capability alone is not operationalization. The challenge is turning that intelligence into a workflow an enterprise can run, trust, and govern.
See also: AI Without Governance Is Just Faster Risk
A better way to evaluate enterprise AI
In our view, enterprise AI should be evaluated across three distinct layers: thinking, planning, and execution.
Thinking is the model’s ability to understand intent, reason through ambiguity, generate logic, and propose an answer. This is where modern AI already shines. It can interpret natural language, make sense of messy inputs, and often produce strong first-pass outputs.
Planning is what sits between intelligence and action. It is the ability to translate that thinking into a repeatable, human-readable process that a business can understand and trust. A real plan should define workflow steps, account for validation and exceptions, incorporate human review where needed, and remain maintainable over time.
Execution is the ability to run that plan reliably in the real world across large files, business systems, and recurring workflows, with the controls enterprises need. That includes scale, auditability, versioning, approvals, security, and governance.
Our research suggests that AI is already strong at thinking, less reliable at planning, and still limited when it comes to execution in enterprise-grade workflows.
Thinking is already strong
We first looked at the reasoning layer. Can the model understand business intent, interpret unstructured inputs, and generate a plausible approach to the task?
On that dimension, Claude Cowork is genuinely impressive.
In data-heavy workflows, we tested whether the model could take natural-language instructions and translate them into meaningful business logic across spreadsheet and analytics tasks. What we found was encouraging. Claude was often able to handle structuring, filtering, joining, reshaping, and other common transformations from plain-English instructions. Testers repeatedly described the experience as intuitive and accessible.
For document-heavy workflows, we assessed the ability to reason through unstructured inputs such as long PDFs and invoice-style documents. Here too, the results were directionally strong. Claude could often identify structure, separate distinct records, and produce a useful first-pass extraction from files.
We also tested whether the model could improve its reasoning when given feedback or additional instructions. That turned out to be another area of strength. In many cases, it could revise its output, explain what it got wrong, and improve on a second pass.
In many enterprise workflows, the bottleneck is not knowing what needs to happen. It is knowing how to express that logic in the syntax of legacy tools. AI lowers that barrier and compresses the distance between business intent and technical action.
Planning is where the cracks begin to show
The next question was whether the model could turn a plausible answer into a repeatable, human-readable workflow that a business team could inspect, validate, and maintain.
When we tested whether Claude would produce a business-readable workflow definition, the “plan” was not a plan in any business sense. It was code, fragments of logic, or a mix of narrative and generated scripts. In some cases, Claude created step-by-step documents or custom skills, but these were not the same as a workflow definition a team could truly own and manage.
We also found that even when the output looked promising, the underlying logic was often difficult to follow. In spreadsheet-oriented workflows, Claude regularly dropped into custom JavaScript, Python, bash, or XML manipulation. That makes the result fragile and hard to maintain, especially for business users who need to understand how the workflow actually works.
The core issue was not that AI had no plan. It was that the plan often lived inside the model or inside generated code, rather than as a human-readable process the business could inspect and trust.
Enterprise workflows require turning probabilistic reasoning into deterministic business processes. A model may generate a plausible answer, but businesses cannot run critical workflows on plausibility alone. The logic has to be made explicit. Assumptions have to be surfaced. Validation has to be built in. The process has to be repeatable enough to produce dependable outcomes over time.
Execution remains the real bottleneck
The final question was whether the system could run those workflows reliably under real-world enterprise conditions.
In one test, we evaluated whether the system could handle a large, business-critical spreadsheet and perform a relatively simple update without damaging the file. The result was a meaningful breakdown in execution reliability. The output file was corrupted and could not be reliably opened. The process took longer than the manual alternative, and significant human effort was still required to validate and recover the output.
In another test, we looked at whether the system could extract useful information from long PDFs with enough accuracy to reduce review effort. Claude produced a directionally useful first pass, but accuracy became uneven when tables, line items, and visually structured content mattered. More checking and rework were still required before the output could be trusted.
We also tested whether workflows could be reused and rerun over time. Claude could document what it had done or attempt to create reusable skills, but this was not the same as having a durable execution layer. Rerunning the process often required reconstructing context or using documentation as a workaround.
Finally, we tested whether the system could operate inside the enterprise environment by connecting to upstream and downstream business systems. Claude was most effective when working with local files or directly accessed files. Deeper connectivity to systems of record such as ERPs, close platforms, tax engines, and downstream operational systems was much more limited in practice.
Across these tests, the pattern was consistent. The issue was not whether the AI understood the task. In many cases, it clearly did. The issue was whether it could execute reliably under enterprise conditions.
What enterprises should take from this
The lesson here is not to be so wary of AI that we dismiss it, but that there needs to be an awareness of where it still needs support.
General-purpose AI is already a strong reasoning layer. It can help people think faster, prototype faster, and get to a plausible first pass with much less friction.
But enterprise value is created further downstream. It comes from workflows that can be run the same way every month, inspected without decoding generated code, integrated with systems of record, and defended in a compliance review.
That requires an operating framework around AI that provides repeatability, human-readable workflow definitions, low review burden, durable automation, connectivity, and governance. In other words, the real challenge is not intelligence. It is execution.
