Chat Picker

Choosing Financial AI Tools: Risk Models and Compliance Checks

A practical afternoon test compares different vendor-documented file limits, source access, plan controls, and human-review requirements before an AI tool enters a risk or compliance workflow.

Sources checked 2 Oct 2026

Choosing among ChatGPT, Claude, and Gemini for financial risk or compliance work should start with a controlled test of file handling, source traceability, plan limits, data controls, and human review. The vendors’ pages we read describe product features and restrictions, but they do not say that any assistant is fit for a particular risk model, AML/KYC process, or regulated decision. Chat Picker has not tested the assistants for this use.

Know these limits before you start

Use synthetic records; this is an evaluation plan, not financial, legal, or compliance advice. OpenAI’s accuracy guidance says ChatGPT can be incorrect or misleading and sound confident, so important information needs verification. Anthropic’s guidance says not to rely on Claude as the only source of truth and to inspect cited sources in high-stakes work. Google’s source-panel help says Gemini Apps may show sources and related content when available. Treat source visibility as a test item, not a correctness guarantee.

Policy boundaries matter too. OpenAI’s usage policies bar automated high-stakes decisions in sensitive areas without human review. Anthropic’s Usage Policy calls for relevant human expertise in elevated-risk uses and qualified review of covered advice, recommendations, and subjective decisions before dissemination or finalization. Gemini’s guidelines say it should not generate factually inaccurate output that could cause significant harm to a person’s finances, but also describe LLM outputs as probabilistic.

File ceilings differ. As read on October 1, 2026, OpenAI’s File Uploads FAQ gives a 512 MB limit per file, a 2 million-token cap for text and document files, and an approximately 50 MB limit for spreadsheets. Claude’s upload guidance lists 500 MB per chat file and 30 MB per project file, with possible additional token limits. Gemini’s file guidance lists up to 10 files per prompt, up to 100 MB per non-video file, and up to 2 GB per video, subject to availability.

Training settings also need a direct check. As read on October 1, ChatGPT’s pricing page lists an opt-out from training on chats for Free, Go, Plus, and Pro. Claude’s pricing page lists an opt-out for Free, Pro, and Max and says Team is not trained on by default. Google’s US plans page points to data-handling information but states no training setting, so mark that point unverified.

The test plan

Use the same synthetic cases, a clean chat or project, and a simple log for every assistant.

  1. Accuracy and explainability. Prepare a synthetic small-business credit file and a fictional policy memo. Prompt: “Review both attachments. List every risk factor, cite its exact field or page, separate facts from inferences, identify missing data, and do not approve or decline the applicant.” Run it twice in fresh chats. Look for unsupported claims, broken citations, hidden assumptions, and changed conclusions. Log each claim, its source, omissions, corrections, and reviewer minutes.

  2. AML and KYC checks. Prepare a synthetic customer record, a small transaction sample, and a written screening policy. Prompt: “Identify possible anti-money-laundering or know-your-customer review signals. Quote the source field, name the relevant attached rule, and label the signal unconfirmed or confirmed by the evidence. Do not infer traits absent from the record.” Look for unsupported matches and confusion between an alert and a finding. Log the signal, evidence, rule reference, and reviewed disposition.

  3. Data handling. Create clean files containing unique synthetic markers, then record account, plan, privacy, connector, and tool settings. Prompt: “Use only the attached synthetic files. Name each file used, quote its marker, and do not use web search or connected apps.” Look for unexpected sources, permission requests, and settings that differ from your intended configuration. Log the settings, files used, limits encountered, and available deletion information.

  4. Model governance and version control. Reuse one frozen test file. Prompt: “List only the model or version labels and tools visibly exposed in this interface. If either is not visible, write ‘not exposed’; do not infer a version.” Treat the interface—not the assistant’s self-report—as the record. Log the visible label, plan, tool state, date, time, workspace role, and any information the interface does not expose.

  5. Vendor support and viability. Send each vendor the same written question: “For confidential financial analysis, do your current consumer and team terms provide contractual retention periods, incident notices, audit rights, model-change notices, and support response targets? Identify the plan and contract clause for each answer.” Log the respondent, date, answer, linked contract language, and unresolved points. Do not treat a general marketing answer as a contract term.

  6. Total cost of ownership. Build a workload sheet from the first two tests, then rerun their prompts with unchanged inputs. Prompt: “Process both attached synthetic cases under their attached rules. Begin with a source checklist, show each calculation, and end with unresolved items. Do not browse or use connected apps.” Log verified subscription cost, usage-counter changes, reviewer time, permissions required, and any limit reached. Calculate total cost only from current vendor figures and your own logged work.

How to read your results

Set pass-or-fail requirements before reviewing the answers: required source traceability, permitted data settings, approved tools, human sign-off, acceptable review effort, and a spending ceiling. A failed non-negotiable control rules out that configuration regardless of how polished the response appears.

Open every material citation, recompute the calculations, and record corrections. Keep an observed error separate from a point the vendor pages leave undocumented. Your results describe these prepared cases, one plan, and one point in time; they do not establish general accuracy, reliability, regulatory status, or a ranking. If none passes, narrow the task, add controls, or retain a manual process.

Where the plans differ

Start with these rows in Chat Picker’s Claude vs ChatGPT vs Gemini comparison matrix: Free plan, Low-cost tier, Main paid plan, Heavy-use plans, Team plan, Context window in the app, How usage limits are described, Training on your chats, and Ads. Each matrix figure includes its vendor source and read date; an unconfirmed cell is marked “Not verified” rather than filled with an estimate.

As read on October 1, 2026, ChatGPT Plus is listed at $20 per month when billed monthly. Claude Pro is $20 monthly or $17 per month with annual billing and $200 up front. Google AI Pro is $19.99 per month. These are entry prices, not total operating costs; include your logged review time, extra usage, seat needs, and administration.

Context labels are not directly comparable. Read on October 1, 2026, Claude’s pricing page says context can reach 1 million tokens on every plan and varies by model; Google’s plans page gives no app token figure; and ChatGPT’s pricing page separates instant and reasoning context figures by plan. Confirm the current matrix cells before subscribing.

Sources

For current prices and limits, see the dated comparison pages.

Compare assistants