Editor’s note: This is Part 2 of a two-part series on choosing the most appropriate AI tool for finance and accounting use. It looks at considerations for the enterprisewide decision, e.g., where AI should sit, what it should be allowed to access, and how far it should be trusted to act before a human steps in. Part 1 provides a simple tour of the major AI tools and includes their strengths and where they fall short.
In Part 1, I argued that there is no single AI winner. OpenAI’s ChatGPT, Anthropic’s Claude, Google Gemini, Microsoft Copilot, and the specialist research tools all have their place. However, that conclusion is only half the story. Choosing a tool for yourself is one thing; choosing one for a finance team, practice, or organisation is quite another.
For individuals, the key question is usually, “Which tool gives me the best answer?” For organisations, the better question is, “Where should AI sit in the way we work?”
The major tools are no longer sitting politely in a separate browser tab waiting for someone to ask a question. They are being embedded into documents, meetings, workflow systems, enterprise data sources, and more. The assistant is moving from answering work to participating in work. That changes the risk profile rather dramatically.
For accountants, that means the decision on which tool to use should be based on three practical questions:
- Where it sits: If the major tools are no longer positioned in a separate browser tab, the practical question is where do they live now — inside your documents, inbox, meetings, spreadsheets, and chat tools, or still somewhere separate that somebody has to remember to open?
- What it can reach: Which systems, files, and data sources can it genuinely access, and whose permission(s) determines what it can access?
- How far it runs: Does it draft and wait, or does it take a goal, complete several steps, and return with the work?
Those questions matter because they determine governance. A tool that drafts a paragraph and waits is one thing. A tool that can access files, email a client, update a spreadsheet, or trigger a workflow is something else entirely. If the tool introduces errors, the first may embarrass you but the second compounds that at scale.
Here, I discuss the capabilities and governance issues of the major AI tools.
ChatGPT: A flexible workbench with growing connections
ChatGPT is increasingly becoming a general-purpose workbench rather than just a chat interface. For a finance team, the attraction is flexibility: It can help draft policies, explain spreadsheet logic, prepare training material, summarise research, and work with connected files and applications where enabled. In the real world, that might mean taking a rough month-end commentary, a list of budget variances, and three anxious emails from operations and then producing a revised draft of the commentary that is at least coherent enough for a manager to refine.
This matters because ChatGPT is increasingly being extended through tools, integrations, and connections that make it more useful in typical finance work. Examples might include:
- Plugins and connected apps: Package reusable AI agent skills (i.e.,reusable modules of instructions, tools, and workflows) with approved apps that can search workplace sources, conduct research with cited sources, and take permitted actions across connected services.
- Financial data integrations: Bring trusted market and company data from FactSet, Dow Jones Factiva, LSEG, Daloopa, S&P Global, and other providers into research and modelling workflows.
- Agent-style multistep tasks: In eligible plans, hand over an outcome, let the assistant work across approved applications and files, and review the resulting documents, spreadsheets, presentations, or reports.
- Custom GPTs (generative pre-trained transformers) shared across the team: Create an IFRS, US GAAP, tax, or policy assistant once so colleagues can use the same instructions, knowledge, and tools consistently.
- Model Context Protocol for internal systems: Connect approved internal data and tools through governed applications, with administrators controlling deployment, permissions, and permitted actions.
- Spreadsheet and presentation workflows: Where supported, work directly in spreadsheet and presentation environments, including approved Microsoft or Google Workspace flows.
Governance point: Whether you use ChatGPT or another AI tool, define what data may be uploaded, which connectors are approved, who may create shared agents, and what review is required before outputs are circulated externally. Without those rules, you do not have an AI strategy; you have a very enthusiastic experiment.
Claude: Strong for long documents and polished deliverables
Claude’s organisational appeal lies in its handling of lengthy material and its ability to produce polished artefacts. In finance terms, think policy reviews, board papers, management commentary, control narratives, contract summaries, and other documents where the first draft often takes longer than the analysis itself. It is particularly handy when the raw material is not one neat file but a pile of minutes, draft papers, prior-year reports, and comments from people who all believe their paragraph is the important one.
Capabilities include:
- Approved collaboration integrations: Where enabled, bring Claude into approved collaboration channels, use shared context and connected tools, then return the work for review.
- Agent-style multistep work: In eligible plans, hand over a broader objective across approved files and applications, then review the result later.
- Model Context Protocol connectors: Give Claude governed access to approved business systems and let compatible tools operate inside the workflow rather than relying on stale exports.
- Microsoft 365 integrations: In supported environments, read from or write to Microsoft 365 files and carry relevant context between applications without repeated explanations.
- Skills for recurring work: Package instructions, scripts, and reference material so Claude follows the same method for repeatable work across the team.
- Word review workflow: Where supported, revise selected text, preserve surrounding formatting, and use tracked changes so colleagues can accept or reject edits in Word.
Governance point: Be clear about which files Claude may access, which outputs it may edit directly, and whether tracked changes or approval workflows are required. If the tool is helping rewrite client-facing or board-facing documents, you need version control and accountability.
Google Gemini: Strongest where Google Workspace is home
Gemini makes most sense where the organisation already works in Google Workspace. In that setting, it can sit inside Gmail, Docs, Sheets, Slides, and Drive, helping with email drafting, document summaries, spreadsheet explanations, and source-grounded research. For a finance team that keeps forecasts in Sheets, commentary in Docs, and supporting evidence in Drive, the convenience is obvious. The caveat is equally obvious: Convenience is not the same as control.
In practical terms, the main capabilities to consider are:
- Workspace context: Where enabled, brings together context from Gmail, Chat, Drive, and Google Workspace applications so Gemini can better understand relevant projects and work patterns.
- Enterprise connectors: In supported Gemini Enterprise or Google Workspace environments, search or take permitted actions across approved third-party systems.
- Gems (customised Gemini AI assistant) and workflow automation: Create Gems for recurring work and, where available, use them inside an approved automated process that can review context, generate content, and route the next step.
- Gemini Notebook for source-grounded work: Load contracts, prior-year files, or regulatory guidance and ask questions grounded in those sources, with enterprise protections available through eligible Google Workspace plans.
- Continuous-work features: In eligible environments, manage selected tasks and schedules across connected applications, with confirmation required where supervision is appropriate.
- Custom and third-party connections: Connect supported services and tools, including custom applications, through approved connection standards such as the Model Context Protocol.
Governance point: Distinguish clearly between consumer and business features. A personal AI subscription should not become the shadow system for sensitive work. If Gemini is to be used in an organisation, it needs administrator controls, approved data sources, and clear rules about what may be connected.
Microsoft Copilot: Strongest where Microsoft 365 is the work operating system
Copilot’s organisational advantage is obvious: Many finance teams already work in Microsoft 365. Word, Excel, PowerPoint, Outlook, Teams, OneDrive, and SharePoint are not optional extras; they are where the work happens. Copilot’s value is that it can sit inside those applications and use permission-aware organisational context. In a practical finance workflow, that could mean moving from a Teams discussion to an Excel variance analysis, then into a Word commentary, and finally a PowerPoint board slide without treating each step as a separate island. In other words, it should not provide a user something they do not already have permission to access. That is a rather important sentence.
This advantage is most apparent in the practical tools and controls that surround Copilot. These include:
- Work IQ as organisational context: Helps Copilot and agents understand the organisation’s working context by drawing upon relevant emails, meetings, files, chats, people (because Work IQ can use organisational directories and relationship contexts), and connected business systems that the user is already permitted to access.
- Delegated work features: In eligible environments, plan and execute longer-running tasks across Microsoft 365, with checkpoints before actions such as sending emails or making significant updates.
- Copilot Studio for custom agents: Build agents that retrieve live information, carry out actions, and run specialised workflows across Microsoft 365 and connected business systems.
- Power Platform continuity: Reuse Power Automate flows, Power Apps, Dataverse, and existing connectors as parts of agent-led processes rather than rebuilding them elsewhere.
- Agent governance tools: In supported Microsoft environments, observe, secure, and govern agents across the organisation, including some agents built outside Microsoft’s own tools.
- Enterprise data protection by default: In eligible Microsoft 365 Copilot experiences, keeps organisational data within Microsoft 365 controls, honours existing permissions, and does not use prompts or responses to train foundation models.
Governance point: Copilot inherits the strengths and weaknesses of your Microsoft 365 environment. If permissions, sensitivity labels, and file hygiene are poor, Copilot may expose the mess rather than create it. This is not an argument against Copilot; it is an argument for cleaning the cupboard before installing a spotlight.
The real battleground: Access, action, and accountability
Across all four platforms, the marketing names will keep changing. That is the easy prediction. The more important shift is that AI tools are moving in three directions at once:
- Model Context Protocol is becoming a common connection standard: Anthropic introduced it, and other major AI providers have announced support or integration plans. Internal connections are therefore becoming more portable and less dependent on one vendor.
- Your productivity suite no longer fully determines your AI assistant: AI assistants are increasingly crossing suite boundaries: Claude has Microsoft 365 integrations; ChatGPT supports business connectors where enabled; and Microsoft 365 Copilot Researcher supports model choice in eligible environments. The suite and the assistant are becoming separate choices.
- Reach and governance create the real differences: Copilot is strongest when permission-aware Microsoft 365 context matters, whereas Claude may work deeply with the files and applications placed in front of it. Neither approach is inherently better; the right one depends on the control your organisation needs.
How an accounting or finance team should choose
If you are making this decision for a team, do not start with model rankings. Start with the workflow. Pick one painful, repetitive, well-understood process and test AI there. Month-end close, board reporting, management commentary, reconciliations, client onboarding, cash flow forecasting, and variance analysis are all obvious candidates. Do not begin with “transform the finance function”. That phrase has caused enough PowerPoint damage already.
Ask these questions to help make a choice between AI models:
- Where does the work actually happen? Choose a tool that meets the team inside the applications and channels they already use rather than adding another destination.
- What does it need to access and whose permissions should it inherit? Broader context makes an assistant more useful, but it also increases the importance of access controls and data governance.
- How far should it run without supervision? Drafting and waiting carries a different risk from executing a workflow, so the approval process must match the level of autonomy.
- Who can monitor what the agents are doing? Decide early how agents are approved, supervised, and reviewed before colleagues begin building their own.
- What happens if you change direction? The Model Context Protocol and wider cross-platform support are reducing switching costs, so avoid treating today’s choice as permanent.
Once you have selected a workflow, document the controls. What data may the tool access? Who reviews the output? What actions require approval? How are prompts, outputs, and decisions retained? Who is responsible when the tool is wrong? If that last question feels uncomfortable, good — it should. Accountability cannot be outsourced to software.
A useful pilot might be deliberately unglamorous: Take last month’s management accounts, ask the tool to identify the five largest budget variances, draft commentary for each one, and list the evidence it relied upon. Then compare its work with the finance team’s final pack. Where did it save time? Where did it invent causality? Where did it miss the obvious because the obvious was buried in an email thread? That is a better test than asking it to write a strategy paper.
The following summary matrix can help with decision-making.
Word to the wise
The sensible approach is therefore simple: Start small, choose a real workflow, connect the tool carefully, review everything, and expand only once you understand what it does well and where it fails. AI adoption is not a race to install the most impressive demo. It is a process of making work better without making risk invisible.
Liam Bastick, FCMA, CGMA, FCA, is director of SumProduct, a global consultancy specialising in Excel training. He is also an Excel MVP (as appointed by Microsoft) and coauthor of Python in Excel: Unlocking Powerful Data Analysis and Automation Solutions. To comment on this article or to suggest an idea for another article, contact Oliver Rowe at Oliver.Rowe@aicpa-cima.com.
LEARNING RESOURCE
AI in Action: Your First Steps to Smart Adoption
In this session, you’ll discover easy-to-use tools, see real-world examples, and leave ready to take your first confident steps towards using AI to save time, boost creativity, and make smarter decisions.
WEBCAST
MEMBER RESOURCE
“AI Tools for Finance Professionals — Which One and When? Part 1”, FM magazine, 29 September 2026
Leave a Reply