Assistant platforms for finance teams — Copilot, Claude Projects, ChatGPT, Gems

An assistant platform is a hub where you build an AI assistant with its own knowledge base, instructions and connectors for a specific team or task — the four big ones are Microsoft 365 Copilot, Claude Projects, ChatGPT Projects/Custom GPTs, and Gemini Gems. For finance teams they differ fundamentally on context size, Office/365 integration, governance controls and pricing — the choice determines whether you get a working assistant or an empty promise.

A finance team wanting to set up its own AI assistant — not for the whole organization, but for the close cycle, accounts receivable, or reporting work — faces a choice between four serious platforms: Microsoft 365 Copilot (Agents), ChatGPT (Custom GPTs and Projects), Gemini Gems, and Claude Projects. At a high level they do the same thing: capture instructions, attach documents, switch on tools, share with colleagues. The differences sit in points that are often decisive in a finance context: context window, integration with the accounting system and MS365, sharing model, audit logging, and data governance.

The shortest summary for finance: pick the platform your team already works in. A Microsoft 365 shop builds on Copilot, a Google Workspace team on Gems, a team already using Claude on Projects. Pure ChatGPT/Custom GPTs are rarely the first pick for finance.

The four platforms through a finance lens

Microsoft 365 Copilot

Agents inside Microsoft 365 Copilot, built through Copilot Studio or directly in Teams/SharePoint. The core differentiator for finance: native access to Microsoft Graph — Outlook, calendar, SharePoint, OneDrive, Teams — and the option to support Excel and Word work directly.

Strong for finance: sits exactly where the work already happens. The controller living in Excel models, the finance manager with a full Outlook, the whole team sharing documents on SharePoint — Copilot runs on top of that without anyone learning a new tool. Inherits tenant-level security and compliance (data residency, DLP, audit logs, SSO) — a big plus for finance.

Weak for finance: model capabilities typically lag what you get directly from OpenAI or Anthropic. For heavy reasoning (IFRS analyses, transfer pricing) Copilot is mediocre. The builder experience (Copilot Studio) is more powerful than plug-and-play. One important watchpoint: since April 2026, Flex Routing is on by default — queries get routed to US servers as soon as EU capacity is saturated. For finance data, an administrator has to switch this off explicitly.

Claude Projects

A workspace with custom instructions, knowledge files (up to 200K tokens on the standard tier, 1M on enterprise), and shared chats. No built-in tools for image generation, but strong document analysis and reasoning quality.

Strong for finance: the best full-context reading of the four — knowledge files are loaded into the window in full, no RAG surprises. Zero data retention on the Enterprise tier. Excellent for annual-report analyses, contract portfolios, IFRS disclosures. For finance work that requires nuance and where every clause matters, this is by far the strongest platform.

Weak for finance: no native link with Outlook or Office. No public sharing (no "Claude Store"). Less suited for broadly shared inboxes or Excel work than Copilot. Limited built-in tools — for tool integrations, Anthropic leans on MCP and the API.

ChatGPT — Custom GPTs and Projects

Two variants. Custom GPTs: a ChatGPT persona with its own instructions, knowledge (up to 20 files), capabilities (web search, code interpreter, DALL-E), and Actions for API integrations. Publicly shareable. ChatGPT Projects: a workspace with instructions, files, and a series of chats in which team members can pick up each other's conversations.

Strong for finance: strong code execution (Python) makes it powerful for data analysis on large Excel extracts. Deep Research for benchmarks and market information. ChatGPT Projects are useful as a workspace per client or per large dossier (think: an M&A process with multiple colleagues).

Weak for finance: no native integration with Outlook or an accounting package. Custom GPTs are often overkill for finance (public sharing is unwelcome in a finance context). Knowledge at larger volumes gets RAG'd, not always transparently — less suited for close reading of an annual report.

Gemini Gems

Google's variant: custom personas inside Gemini, with instructions, knowledge (up to ten files or 10+ million tokens on paid tiers), and — with Workspace integration — direct access to Drive, Gmail, Docs, and Calendar.

Strong for finance: very large context window (whole dossiers in one go), strong Workspace integration, strong multimodal capability (images, video, audio — useful for dashboard screenshots or scanned documents).

Weak for finance: only relevant if your company runs Google Workspace. The sharing model inside Workspace is less developed than ChatGPT's Custom GPTs. Outside Workspace, the product still feels young.

Comparison on dimensions that matter for finance

Context and knowledge

• Copilot Agents: leans on SharePoint and Graph — fetched dynamically, whatever the user is allowed to see by permissions. Good for "consult our close manual."
• Claude Projects: up to 200K (standard) or 1M (enterprise) full context — the most predictable for careful reading.
• ChatGPT Projects: ~200K worth of documents, medium-precision retrieval.
• Custom GPTs: 20 files, RAG at larger volumes. Fragments can get lost — less suited for close reading of a contract.
• Gemini Gems: the largest window (up to 1M tokens), full context viable for heavy dossiers.

Rule: above 50 pages of reference material and any depth of content (annual report, transfer-pricing report, long client contracts), Claude or Gemini wins. For scattered short documents the gap is small.

Integration with finance systems

• Copilot: native MS365, but no out-of-the-box Exact/Twinfield/AFAS integration.
• Claude: no native MS365, but strong MCP support for custom integrations.
• ChatGPT: API and Custom GPT Actions for custom integration.
• Gemini: native Google Workspace, no out-of-the-box accounting link.

For a direct link to your accounting system you need, in all four cases, either MCP, a specific connector, or a platform with the link already built in. This is where products like Saldus come in — delivering the accounting integration out of the box.

Sharing model — relevant for finance teams

• Copilot Agents: tenant-scoped under MS365 governance, strong for internal rollout, not publicly shareable.
• ChatGPT Projects: shareable inside the workspace with view/edit/manage rights.
• Custom GPTs: private, inside the workspace, or public. Flexible, but "public" is rarely desirable in a finance context.
• Gemini Gems: shareable inside Workspace, public sharing limited.
• Claude Projects: shareable inside Teams/Enterprise environments, no public link.

For finance teams sharing internally, the sharing model is rarely the bottleneck. For customer-facing assistants (think: a chatbot for AR queries), Custom GPTs are technically suitable, but the privacy implications have to be fully worked through.

Governance and audit logging

• Consumer tier (free Claude.ai, ChatGPT Plus, free Gemini): no finance data, ever.
• Team/Business tier: the minimum for customer data.
• Enterprise tier (Claude Enterprise, ChatGPT Enterprise, Copilot for Business, Gemini Enterprise): contractual guarantees, SSO, audit logs, data residency.

For finance, Business/Team tier is the minimum the moment customer data enters the assistant. Audit logging is available on every Enterprise tier, but it has to be switched on and reviewed periodically — otherwise it's security theatre.

Decision order for finance teams

A workable order:

1. Where does the team already work? MS365 → Copilot is the first pick. Google Workspace → Gems. Mix → look at the heaviest finance use case.
2. What is the main task? Consulting the close manual, mail/calendar work → Copilot. Reading long contracts, IFRS analyses → Claude Projects. Data analysis on large Excel extracts → ChatGPT Projects or Saldus.
3. Who will use the assistant? Finance only, internal → Copilot or Projects. Customer-facing → Custom GPTs (with privacy justification), or better: a custom frontend on the API.
4. Which accounting integration do you need? None → any of the four works. Yes → build the MCP integration yourself, or pick a platform that delivers it out of the box.
5. What is the governance requirement? Strict EU-compliant, no data exchange outside the tenant → Copilot, an Enterprise tier of the other three, or an embedded deployment.

In practice many finance teams run two tracks: Copilot for everything tied to Outlook and Excel, plus a second platform (often Claude Team or a finance-specific AI platform) for work that sits outside the Microsoft world or needs more model quality. That isn't sloppy — that's pragmatic.