This article reviews how we used Anthropic’s Claude to automate vendor spend reconciliation across two deployment models: Anthropic’s desktop tool for internal work and a custom application on Amazon Web Services (AWS) for client data. It covers what each approach can and cannot do, and how we moved from prototype to production.

The Capability Is Already Here

Most accounting teams perform vendor spend reconciliation as part of the monthly close. The work is well understood: match invoices against contracts, flag variances, update registers and produce a summary. It is also relentlessly manual. Staff accountants spend the final week of every month pulling data from the accounting system, cross-referencing contracts and assembling workpapers in the shared drives.

Artificial intelligence (AI) can now handle this type of structured, rules-based work. Anthropic’s research shows that a large share of finance tasks falls within current AI capability. When researchers measured the share of tasks that current AI models are technically capable of handling, finance roles scored high, yet actual adoption remains low.

The Gap Is Not About Technology

The gap exists because different types of accounting data require different levels of protection, and most firms treat this as a binary choice: use AI with no safeguards, or do not use it at all. That framing is wrong. The right approach is to match the deployment model to the data sensitivity.

Tier 1: Start With Cowork for Internal Workflows

We began by prototyping the reconciliation workflow using Claude’s desktop tool, Cowork, with sample data. Cowork runs on Anthropic’s infrastructure under their data processing agreement (contractual protection, not architectural). For internal process development and non-client data, this is where to start. The setup took 15 minutes. Cowork provides:

• Markdown instruction files that act as onboarding: a context file describing the workflow, a task file listing the reconciliation checklist and a memory file that retains decisions between sessions.
• Skills that produce specific outputs: formatted reconciliation reports, exception summaries and client-ready narratives.
• Connectors that integrate directly with QuickBooks, OneDrive and SharePoint.

  Using sample vendor data (273 invoices, 14 contracts), Cowork extracted terms from a new contract PDF, compared spend against contract values, flagged three variances and built a reusable dashboard. This prototype proved the workflow before we moved any client data.

Tier 2: Move to AWS Bedrock for Client Data

For client financials, we built a custom application inside an AWS Virtual Private Cloud that calls the same Claude model through AWS Bedrock. This required engineering investment, but the architectural guarantees are fundamentally different:

• Anthropic delivers model weights to AWS, which copies them into an isolated account Anthropic cannot access.
• Prompts never leave the AWS backbone. Traffic flows through VPC PrivateLink with no public internet exposure.
• Encryption keys and audit logs remain under our control.

Our application replicates the Cowork workflow patterns (instruction files in S3, persistent memory in DynamoDB, a knowledge base through Bedrock Knowledge Bases) but every component runs inside our VPC. The connectors to QuickBooks and SharePoint are custom integrations we built against those platforms’ Application Programming Interface (API)s. The markdown files, skills and memory logic that took 15 minutes in Cowork required real development work to replicate on Bedrock, but the workflow the accounting team interacts with looks and feels the same.

The gap does not close with a single tool. It closes when firms match the right deployment to the right data. Thus, it’s best to do the following: Start with Cowork and sample data to prove the workflow. Move to a private cloud when client data is involved. The capability is here. The architecture to use it safely is here. What remains is for domain experts in accounting to write the instructions, because no one else understands the process well enough to get them right.