The Assistant and the Agent
"Excel is not a spreadsheet application. It is the operating system of the financial profession. Everything else — the presentations, the memos, the reports — is built from what lives in a spreadsheet first."
In Chapter 16, you learned the Knowledge Extraction Method — the structured process for surfacing tacit professional knowledge and encoding it in a SKILL.md that produces reliable agent behaviour. Now you will apply that methodology to the domain where AI agents have the most immediate, measurable impact on professional output: finance.
Every finance professional knows the feeling. The corporate controller reconciles accounts in Excel. The investment banker builds deal models in Excel. The equity research analyst updates earnings forecasts in Excel. The FP&A manager presents variance analysis to the board from a PowerPoint deck built from an Excel model. If you work in finance, accounting, treasury, or investment management, Excel is the medium through which your professional judgement is expressed. This chapter begins there — not because Excel is the only tool that matters, but because it is the tool where the distinction between an AI assistant and an AI agent becomes concrete and consequential.
Anthropic offers two architecturally different ways to bring Claude's intelligence into your Excel workflow. Confusing them will cause real frustration: you will look for features in the wrong place and miss capabilities that are available to you. Understanding the distinction will also give you a framework that applies far beyond finance — the difference between intelligence embedded in a single tool and intelligence orchestrating across tools is one of the most important architectural decisions in any domain agent deployment.
The Core Distinction
Claude in Excel is an AI assistant embedded directly inside Microsoft Excel. It lives in a sidebar within your workbook. It reads your spreadsheet — all sheets, all formulas, all data — and applies reasoning to answer questions, trace dependencies, test scenarios, debug errors, and build new model structures from plain-language descriptions. Through its pre-built Agent Skills, it generates professional financial deliverables like DCF valuations and comparable company analyses. When configured with data connectors, it can incorporate market data directly into your workbooks. It is a standalone Microsoft add-in. It does not require Cowork. It works within one application.
Cowork with Excel involves Claude acting as an AI agent that orchestrates across multiple applications. Excel is one node in a multi-step autonomous workflow. The agent can perform analysis in Excel, then carry that context forward and autonomously build a presentation in PowerPoint — without you switching applications or copying anything manually. This capability requires the Cowork platform and operates through finance plugins that bundle domain knowledge, data connectors, and workflow logic.
The shorthand: Claude in Excel is a deep companion for financial modelling work. Cowork with Excel is a workflow orchestrator that treats Excel as one stop in a larger automated process.
What Makes Them Different
| Claude in Excel | Cowork with Excel | |
|---|---|---|
| What it is | Standalone Microsoft Excel add-in (sidebar) | Part of the Cowork agent platform |
| Architecture | AI assistant embedded in one application | AI agent orchestrating across applications |
| Scope | One workbook at a time | Multiple applications (Excel, PowerPoint, Word) |
| Pre-built financial skills | 11 Agent Skills including comps, DCF, 3-statement models, due diligence packs, earnings analyses, initiating coverage, and presentation tools | Domain plugins with bundled skills and slash commands |
| Data connectors | Any connector configured in your Claude settings | Same connectors, applied across multi-app workflows |
| Cross-app capability | Excel only | Excel to PowerPoint orchestration (research preview) |
| Who it is for | Financial analysts doing modelling work directly in Excel | Enterprise teams running agent workflows across tools |
| Plan requirement | Pro, Max, Team, Enterprise (beta) | Team and Enterprise |
| Requires Cowork? | No | Yes |
One detail in this table deserves emphasis. Claude in Excel and Cowork do not maintain separate connector ecosystems. Any data connector you configure in your Claude settings — S&P Global, FactSet, PitchBook, Morningstar, Daloopa, LSEG, or any other — works in both environments automatically. The difference is not which connectors are available. The difference is scope: Claude in Excel uses those connectors to serve analysis within one workbook; Cowork uses the same connectors to serve workflows that span multiple applications.
This is a useful mental model beyond finance. In any domain, the question is not "which AI tool has the most features?" but "does my workflow live inside one application or across several?" The answer determines whether you need an embedded assistant or an orchestrating agent.
The Two Layers of Claude in Excel
Claude in Excel operates at two distinct layers. Understanding the difference helps you use it effectively.
Layer 1: General workbook intelligence. Claude reads the open workbook — all sheets, all formulas, all data — and applies general reasoning to answer questions about it, trace formula dependencies, test scenario changes, debug errors, and build new model structures from plain-language descriptions. This layer requires no configuration and works on any workbook. It includes native Excel operations: pivot table editing, chart adjustments, conditional formatting, sort and filter, data validation, and print area configuration. Every action Claude takes is traceable — when Claude references revenue being driven by cell B12, clicking that reference in the sidebar navigates Excel directly to it.
Layer 2: Pre-built Agent Skills. Claude in Excel ships with specialist financial workflows purpose-designed for investment banking, equity research, and financial analysis. These Agent Skills go beyond general assistance: they produce professionally structured financial deliverables that follow industry conventions. The current set of Agent Skills includes:
- 3-Statement Financial Modelling — integrated income statement, balance sheet, and cash flow statement
- Comparable Company Analysis — valuation multiples, operating metrics across public and private peers
- Competitive Landscape Analysis — market positioning and competitive dynamics
- Discounted Cash Flow Model — FCF projections, WACC calculations, scenario toggles, sensitivity tables
- Due Diligence Data Pack — processes data room documents into structured Excel with financials, customer lists, contract terms
- Earnings Analysis — quarterly transcripts extraction: metrics, guidance changes, management commentary
- Initiating Coverage Report — industry analysis, company deep-dives, valuation frameworks
- Strip Profile / Business Overview — condensed company overviews for pitch books and buyer lists
- Pitch Deck Creation — structured presentation content from financial analysis
- PPT Template Creator — presentation templates following industry formatting conventions
- Presentation Quality Checker — reviews and validates presentation deliverables
When connecting to external data sources such as FactSet, Capital IQ, or Daloopa, these skills can incorporate live data into your workbooks. The specific data connectors available depend on your organisation's connector configuration.
The financial services Agent Skills are currently available to Claude for Financial Services users. Check the Claude for Financial Services page for current access requirements and waitlist information.
Chapter Roadmap
This chapter is structured in three parts, each serving a different professional context.
Part One: Claude in Excel — The Embedded Assistant. Covers general workbook intelligence (understanding models you did not build, scenario testing, error debugging) and the pre-built Agent Skills for financial analysis. This is where most finance professionals will spend their initial time. If you work directly in Excel building models and analysing data, Part One gives you immediately applicable skills.
Part Two: Cowork Finance Plugins — The Orchestrating Agent. Covers the Cowork platform's finance capabilities — the corporate finance plugin, the financial services plugin suite, data connectors, and cross-app orchestration from Excel to PowerPoint. This part matters when your deliverables span multiple applications and your workflow involves carrying analysis from a spreadsheet into a presentation, memo, or report.
Part Three: Enterprise Extensions. Covers how to extend the pre-built capabilities with custom SKILL.md files for your firm's specific needs — applying the Knowledge Extraction Method from Chapter 16 to finance domain expertise. This part connects everything: the platform capabilities from Parts One and Two, the extraction methodology from Chapter 16, and the agent architecture from Chapter 15.
Getting Started
Step 1 — Install the add-in. Go to the Claude in Excel page and click Install Now — this takes you to the Microsoft AppSource listing. Click Get it now and sign in with your Microsoft account to add the add-in.
Step 2 — Find the add-in in Excel. After installation, the add-in location depends on your platform:
- Windows (Excel 365): Home tab → Add-ins in the ribbon
- Mac (Excel 16.46+): Tools menu → Add-ins
- Excel on the web: Insert tab → Add-ins → My Add-ins
- iPad (Excel 2.51+): Insert tab → Add-ins
Click the Insert tab → My Add-ins (labelled My Apps in some Office versions). If the add-in still does not appear, click Refresh in the My Add-ins menu. Organisation administrators who have disabled the Office Store can deploy via the manifest XML file through the Microsoft 365 Admin Center instead.
Step 3 — Open the Claude panel. Click the Claude icon in the ribbon, or use the keyboard shortcut:
- Mac: Control + Option + C
- Windows: Control + Alt + C
A sidebar opens on the right side of your Excel window. Sign in with your Claude account. This is where you type prompts and where Claude's responses appear.
Step 4 — Open a practice workbook. Download the chapter practice workbook to follow along with the exercises in this chapter:
Download Practice Workbook (ch17_veeva_practice_workbook.xlsx)Open the workbook in Excel (desktop or web), then open the Claude panel using the shortcut above.
Step 5 — Say hello. Type your first prompt to verify the connection and explore the workbook:
Walk me through the structure of this workbook: summarize the key inputs,
calculations, and outputs, any important relationships between sheets,
and the overall purpose of the analysis.
If Claude responds with a structured description of the workbook, the add-in is working. You are ready for the rest of this chapter.
Plan requirement. Claude in Excel is available in beta for Pro, Max, Team, and Enterprise subscribers.
Supported platforms. Excel on the web, Excel on Windows (Microsoft 365, build 16.0.13127.20296+), Excel on Mac (version 16.46+), and Excel on iPad (version 2.51+).
Supported formats. .xlsx and .xlsm. File size limits apply by plan.
Claude in Excel reads the contents of your workbook — including cell values, formulas, and comments. This creates a prompt injection risk when opening spreadsheets from sources you do not fully trust.
Malicious instructions hidden in cells, formulas, or comments can attempt to trick Claude into extracting sensitive information, modifying critical financial records, or performing destructive actions. This is particularly relevant in finance, where spreadsheets are routinely shared between organisations during deal processes, audits, and due diligence.
Protections in place: Claude will show you a confirmation pop-up before executing functions that fetch external data — including WEBSERVICE, STOCKHISTORY, IMPORTDATA, and others that reach outside the workbook. Do not dismiss these prompts without reading them.
Best practice: When opening a spreadsheet from an external source, review its contents before engaging Claude. If the workbook was not created by someone you trust, treat any unusual formulas or hidden sheets with the same caution you would apply to an email attachment from an unknown sender.
Try With AI
Use these prompts in Claude in Excel or your preferred AI assistant to explore this lesson's concepts.
Prompt 1: Architecture Assessment
I am a [YOUR ROLE] in [YOUR INDUSTRY — e.g., corporate finance,
investment banking, equity research, FP&A]. Describe my typical
workflow for producing a [KEY DELIVERABLE — e.g., quarterly board
deck, deal book, earnings model update].
For each step in the workflow, classify it as:
1. Single-application work (stays inside Excel) — suited to an
embedded AI assistant
2. Cross-application work (moves between Excel, PowerPoint, Word,
email) — suited to an orchestrating AI agent
Based on the classification, recommend whether I should start with
Claude in Excel, Cowork, or both, and explain why.
What you are learning: The assistant-vs-agent distinction is not abstract when applied to your own workflow. By mapping your actual deliverable production process against the two architectures, you develop an intuition for which problems each solves — and where the boundary between them falls in your specific context.
Prompt 2: Agent Skills Discovery
I work in [YOUR FINANCE ROLE]. Claude in Excel has pre-built
Agent Skills including: 3-Statement Financial Modelling, Comparable
Company Analysis, Competitive Landscape Analysis, Discounted Cash
Flow Model, Due Diligence Data Pack, Earnings Analysis, Initiating
Coverage Report, Strip Profile / Business Overview, Pitch Deck
Creation, PPT Template Creator, and Presentation Quality Checker.
For each Agent Skill:
1. Explain what it produces in one sentence
2. Rate its relevance to my role (High / Medium / Low / Not Applicable)
3. Describe one specific situation where I would invoke it
Then suggest which two Agent Skills I should learn first and why.
What you are learning: The Agent Skills are not generic features — they are purpose-built for specific financial workflows. By mapping them to your role, you identify which capabilities will produce immediate value and which you can defer, preventing the common mistake of trying to learn everything at once instead of mastering the tools that match your actual work.
Prompt 3: Connector Scope Analysis
Explain the data connector architecture that Claude in Excel and
Cowork share. Specifically:
1. What are data connectors and what do they provide?
2. Why do the same connectors work in both Claude in Excel and Cowork?
3. What changes between the two environments — if the connectors are
the same, what is actually different?
4. Give me a concrete example: if I have S&P Global configured as a
data connector, describe what I can do with it in Claude in Excel
versus what I can do with it in Cowork.
Use language appropriate for a finance professional who understands
data terminals but not software architecture.
What you are learning: The shared connector architecture is one of the most commonly misunderstood aspects of Claude's finance offerings. By working through how the same data source behaves differently depending on whether it serves an embedded assistant or an orchestrating agent, you build the mental model that prevents confusion throughout the rest of this chapter.
Flashcards Study Aid
Continue to Lesson 2: Understanding Workbooks You Didn't Build →