By 2026 the three most common ways a finance team stands up a reusable AI workflow are a Custom GPT inside ChatGPT Business, a Project inside Claude Team or Enterprise, and a Copilot Agent inside Microsoft 365. They look similar in a demo. They are different tools once you actually use them for month-end.
This is a direct comparison from real finance work. Pricing and features are as of 2026.
The one-line version
- Claude Project. A shared context window. Drop 20 files in and the model reasons across all of them for every conversation. Best for long-form thinking.
- Custom GPT. A packaged prompt plus a small file library plus optional actions (API calls). Best for a single repeatable job.
- Copilot Agent. A Microsoft-hosted agent that reads your Graph data (Outlook, OneDrive, Teams, SharePoint, Excel) with your existing tenant permissions. Best when the data cannot leave the tenant.
Comparison matrix
| Dimension | Claude Project | Custom GPT | Copilot Agent |
|---|---|---|---|
| Excel handling | Reads XLSX and CSV, reasons over rows well. No formula edits. | Reads XLSX and CSV. Code Interpreter can run pandas on it. | Lives inside Excel. Can write formulas, refresh pivots, edit cells. |
| PDF handling | Native PDF ingestion, scanned pages okay for text. | Native, similar quality. | Yes, but weakest of the three on scanned docs. |
| Context window | 200K tokens (Sonnet), 1M on some tiers. Project files stay pinned across chats. | 128K on GPT-4 class, larger on GPT-5. Custom GPT knowledge files loaded per conversation. | Limited per call. Relies on Graph search rather than a full context window. |
| Memory across sessions | Project holds files and system prompt. No per-user cross-chat memory by default. | Optional memory. Custom GPT instructions apply every session. | Ties to your M365 identity. Uses your recent files and calendar. |
| Security posture | Team and Enterprise tenants do not train on your data. SOC 2 Type II. DPA available. | Business and Enterprise tenants do not train on your data. SOC 2 Type II. DPA available. | Data stays inside your M365 tenant. Respects existing SharePoint permissions. |
| Sharing model | Project shared with your workspace or a subset. Every viewer sees the same file set. | Publish inside your workspace or make public. Anyone with access runs the same tool. | Deploy to a team or the whole tenant. Runs inside Teams and Outlook. |
| Cost per seat (2026) | $25-30 Team; Enterprise custom. | $25-30 Business; Enterprise custom. | $30 per user on top of M365 E3/E5. |
| API access | Yes, direct Anthropic API for building your own tools. | Yes, OpenAI API and Custom GPT Actions. | Azure OpenAI plus Graph. More setup. |
Decision table by finance use case
| Use case | Winner | Why |
|---|---|---|
| Month-end variance narrative | Claude Project | Pin the plan, actuals, and prior narrative. Ask for the writeup. It stays in style. |
| Payroll reconciliation | Copilot Agent | Payroll registers cannot leave the tenant. Copilot reads them in place. |
| QoE prep | Claude Project | Long documents, multiple periods, needs cross-reference reasoning. |
| Invoice coding automation | Custom GPT with Actions | Single job, well-defined inputs, needs an API call to your ledger. |
| Weekly flash to team | Custom GPT | Repeat template, same inputs weekly, easy to share. |
| Contract review before signing | Claude Project | Reasoning quality on legal prose is still the strongest. |
| Answering “where is the deck from Q2 board?” | Copilot Agent | Searches SharePoint and Teams by permission, returns the actual file. |
The trap most finance teams fall into
The trap is picking based on what IT already licensed instead of what the job needs. If you have M365 E3, Copilot feels free. So you use it for everything, including tasks it is bad at (long reasoning, prose narrative, multi-file synthesis). Six weeks later your team decides “AI does not work for finance.”
The right pattern is picking the tool per job. Most PE-backed finance orgs I have seen end up with all three. Claude for the CFO’s own reasoning work. A Custom GPT or two for the analyst-level repeat tasks. Copilot for anything touching data that has to stay in the M365 tenant.
What “file handling” actually means in practice
The demos all look similar. Upload a spreadsheet, ask a question, get an answer. The differences show up on real files.
Excel with 40 tabs and dependent formulas. Claude and ChatGPT will read the values but not the dependencies. Copilot can traverse the formula graph because it is Excel. If your model has a lot of internal cross-tab logic, Copilot understands it best.
A 150-page credit agreement. Claude wins on long PDFs. It holds the whole document in context. ChatGPT is comparable. Copilot Agent tends to summarize and lose specific clauses.
A folder of 25 vendor invoices. Custom GPT with a strict extraction prompt is fastest. Claude Project works if you want to also compare against contract terms. Copilot works if the invoices live in SharePoint already.
Security in one paragraph each
Claude Team and Enterprise. Data is not used to train models. SOC 2 Type II, HIPAA available on higher tiers, DPA on request. See the trust center for the current list of certifications.
ChatGPT Business and Enterprise. Same guarantee on training. SOC 2 Type II. Enterprise adds SSO, audit logs, and admin controls.
Copilot for M365. Runs on your existing tenant. Data does not leave your Microsoft 365 boundary. Respects the file and folder permissions you have already set. This is the strongest posture for regulated data if your tenant is set up correctly. The word “if” is doing work there. Copilot honors SharePoint permissions faithfully, which means if your permissions are wrong it will happily show a payroll file to someone who was accidentally added to the folder.
Deployment: single user vs team
For a single CFO or a small finance team of one to three, Claude Projects wins on time to value. Set up in an hour, start using immediately.
For a team of five to fifteen where you want everyone using the same variance-writing prompt, Custom GPTs are easier to distribute. Publish once, everyone runs the same tool.
For an org where finance is not standalone (part of a shared services function, or reporting into a business unit) and files live in SharePoint, Copilot Agents win because you do not have to teach anyone anything new. It shows up in Excel and Outlook.
Push back on this.
Every operator’s situation is a little different. If you run this differently, disagree with the methodology, or think we got something wrong, tell us. We publish the best counter-approaches on our Reader Contributions page, credited or anonymous, your call. Email hello@thepragmaticcfo.com.
FAQ
Can I use all three? Yes and I would recommend it above a certain scale. About $60-90 per seat per month total. For a 5-person finance team the ROI is one avoided hire.
Which one has the best pricing forecasting? None of them will replace a real forecast model. Claude reasons best about scenarios. Copilot updates the model. Custom GPT lets you build a repeatable narrative wrapper.
What about Gemini in Google Workspace? Similar shape to Copilot for M365 tenants. Good if your company runs on Google. Weaker on Excel-native workflows because Sheets is not the primary finance workbench.
Is Custom GPT going away in 2026? No. OpenAI has continued to support and expand the Custom GPT surface. The naming may shift but the pattern is stable.
Does Claude have anything like Copilot Agents (native app integration)? Claude has an MCP (Model Context Protocol) that lets you connect it to internal tools. It is more developer-heavy than Copilot Agents but more portable.
Related reading
- Claude vs ChatGPT vs Copilot vs Gemini for CFO Work
- The CFO’s Guide to Uploading Financials to LLMs Without Privacy Risk
- The Prompt Library Every CFO Should Steal
- The AI-Native Finance Team: 2026 Job Descriptions
Sources
- Anthropic, Claude Projects documentation, docs.anthropic.com
- OpenAI, Custom GPTs and Actions documentation, platform.openai.com
- Microsoft, Copilot for Microsoft 365 documentation, learn.microsoft.com
Written by The Pragmatic CFO. 15+ years running P&Ls and building AI-native finance workflows across portfolio companies.
The Pragmatic CFO 