AI for financial advisors is best understood as a staffing decision. Your main AI tool is not one employee who does everything . It is more like a small team:  a quick operations specialist for routine work, a dependable manager for most assignments, a senior strategist for unusually difficult problems, and a quirky art director for creative exploration. The goal is not to put the most powerful teammate on every job. It is to match the teammate—and the amount of thinking time—to the work. That's how financial advisors ought to think of AI. 

It's wise to use an efficient model for clear, repeatable work. Give a capable general-purpose model most substantive assignments. However, you want to also bring in the most capable model and greater reasoning effort when the stakes justify it by verifying every material fact against its source. A more capable model can improve the odds of a good result, but no model or setting turns an unverified answer into an accurate one.

AI for financial advisors: meet the team

Think of the AI models as the members of staff you and think of the reasoning effort as the amount of time and computational work you authorize that person to spend thinking before answering. Those are two separate choices. Calling it a “strength setting” is a useful conversational shortcut, but it is not OpenAI's product term and it can be misleading: High or XHigh does not make the model physically stronger, automatically faster, or guaranteed to be right.

OpenAI calls the "strength" setting reasoning effort. Higher effort lets a reasoning model generate more internal reasoning before producing the answer. On a genuinely difficult assignment, that can improve thoroughness and the quality of the result. The trade-offs are usually more time and more usage. Lower effort generally favors speed and efficiency. The best choice is therefore not “highest.”  It is the lowest level that reliably completes the particular job to the required standard. You're not using a sledgehammer to hang a picture frame, or calling on a brain surgeon to book your appointments or take notes of your next meeting. 

The team metaphor makes the choices easier to remember:

• Luna at Low: the nimble operations teammate who handles clearly specified filing, formatting, extraction, and other routine work quickly.
• Sol at Medium: the experienced manager who is a sensible starting point for drafting, analysis, tool use, and most multi-step assignments.
• Astra at High or XHigh: the senior strategist brought in for the hardest ambiguities, consequential comparisons, and final challenges—not to rename a folder or polish a short list.
• Terra for a creative pass: the quirky art director with the sketchbook and three unexpected ways to frame an idea. OpenAI does not designate Terra, or any other model, as “the creative model.” Creativity depends on the task, prompt, examples, taste, and revision. The character is a staffing metaphor, not a product guarantee.

This does not mean the speedy teammate is careless or the senior strategist is infallible. Different models have different capabilities and operating profiles, and reasoning effort changes how much computation is devoted to the response. Accuracy still comes from job fit, clear instructions, authoritative sources, testing, and human review. For AI for financial advisors, that final review is indispensable whenever the answer affects a client, public claim, compliance obligation, investment decision, or professional judgment.

I learned the cost of using the senior team for everything

I learned the hard way about all this when I ran out of my Codex allowance for the first time. I had been leaning on a powerful model at a high reasoning setting even when a simpler assignment did not need it. It was like asking the firm's top engineer to staple meeting packets: impressive talent, expensive deployment, and no corresponding improvement in the staples. 

The useful question was not merely, “How do I spend less?” It was, “How do I keep accuracy first, get answers quickly, and avoid running out halfway through an important project?” That is a familiar advisory-business question. Advisors must allocate attention to the decisions that matter most. We should allocate AI effort the same way.

Allowance, credits, tokens, and resets are not the same thing

OpenAI's customer-facing guidance uses usage limits, included usage, allowance, credits, and resets for different parts of the system. Tokens are the underlying units a model processes—pieces of input, output, and sometimes reasoning—not the best everyday name for what a subscriber purchases in the Codex interface. In this article, “allowance” means the usage included with the plan, while “credits” means a separate balance that may pay for supported usage after included limits are exhausted.

A reset refreshes an eligible usage window. A normal reset happens on the schedule shown in the product. A banked reset is a one-time reset saved to an eligible account until it is used or expires. OpenAI says that applying a full banked reset refreshes the eligible five-hour and weekly Codex windows and changes the weekly reset date. A paid instant reset is different from buying credits: it restores the weekly allowance immediately and pulls the normal weekly period forward rather than adding a separate bucket of usage.

For AI for financial advisors, the practical recovery question is simple: what should you do first? Open Settings → Usage and read the actual meter. 

If the normal reset is near and the work can wait, then wait. 

If the work is urgent and the account shows an available banked reset, review its expiration and use the reset before buying extra credits. 

If there is no usable reset—or the refreshed allowance will not cover the additional work—then consider adding credits if that option appears on the account. 

Do not casually say “buy more tokens”; use the choices OpenAI actually shows: wait for the usage limit to reset, apply an available reset, buy an instant reset where eligible, or add credits.

The account's available options depend on the plan and can change. The authoritative place to check is the usage page, supported by OpenAI's current guidance on using Codex with a ChatGPT plan and banked Codex resets.

AI for financial advisors starts with the risk of being wrong

Before assigning an AI teammate, ask what happens if the answer is wrong. That consequences should guide model selection and reasoning effort in AI for financial advisors. Reformatting an internal outline and interpreting a regulatory requirement are not the same job. A fast first draft may be fine for the former. The latter calls for current primary sources, careful reasoning, and a qualified human review before anyone acts on it.

For my everyday Codex work, I use GPT-5.6 Sol at Medium reasoning as the starting point. I use a lighter model for repetitive, well-specified work, and I consider Astra at High reasoning for the hardest final reviews. These are operating choices, not a promise that a particular setting will always be cheaper, faster, or correct. The scope of the prompt, files supplied, tools used, and revisions also affect usage.

The table is a routing guide, not an automatic decision engine. If a “simple” formatting job includes private client records, privacy and authorization may make it high stakes. If a difficult-sounding task is merely brainstorming headings, it may not deserve the strongest setting.

AI for financial advisors: write a useful assignment

Even the right teammate needs a good assignment. Many wasted turns with AI for financial advisors begin with an underspecified instruction: “Write the email,” “fix the spreadsheet,” or “research this rule.” The model must guess the audience, sources, output, and stop point. A better request names all four.

For example: “Using only the attached approved portfolio update description, draft a 150-word invitation forclients. State the date and place exactly as supplied. Flag any missing links. Do not send or publish anything.” This is faster to review because its boundaries are visible. The AI  has fewer opportunities to invent a detail or take an unwanted action.

For a more demanding assignment, add an evidence requirement: “Give me a short answer first. Then list each material assertion, the supporting source, and what still requires my approval. Give me the monte carlo analysis and other MPT statistics, including the rsquared.” That turns an attractive paragraph into an auditable work product. 

Make workflow handoffs clear before automating 

Think of the workflow as the team’s operating chart: each handoff needs an owner, a system, and an approval point. To send a marketing email, an advisor does not want a single “send email” step. There is an offer, an audience, an enrollment or purchase path, a confirmation, a reminder and follow-up. Spell out that chain before asking AI to automate it. Record which system owns each step and which human approves it. 

For instance, say you're planning a webinar for clients and prospects. Your CRM might be the place to manage marketing and client emails, while Zoom handles webinars. The workflow design must specify what Zoom will do with your CRM .  . The same discipline applies to a blog: a draft in DropInBlog is not a published post, and FAQs in a planning file are not necessarily native FAQ entries on the page.

At each handoff, give your AI a precise stop rule: “Prepare the draft and show me the recipient segment; do not send,” or “Enter the article as an unpublished draft; wait for my review before publishing.” The person who owns the communication can then approve both the message and its route. 

Protect client information and professional judgment

Any use of AI for financial advisors must protect client information. Avoid pasting client account numbers, private financial records, or sensitive personal information into a tool merely because the task would be convenient. Use authorized systems, approved data-handling rules, and de-identified data where possible. FA for advisors is great for de-identifying data and is something advisors should always consider in reports.  

AI can help organize evidence, surface inconsistencies, and draft plain-language explanations. It cannot assume a fiduciary's duty, however. It can't decide whether a recommendation is suitable, or certify that a disclosure meets every current requirement. Where a claim depends on law, regulation, product terms, or a client's circumstances, check a current primary source and involve the appropriate professional. A confident sentence without a citation is still an unverified sentence.

Keep a small operating reserve allowance, then adjust it

Treat the AI allowance like a staffing budget. When budgeting AI for financial advisors, start the week by looking at the remaining allowance and the next reset date in the product, not by guessing how many credits or tokens a single message costs. Usage may be shared across local and cloud activity, and the account display may show percentages rather than a predictable message count. Keep a reserve: do not spend the entire allowance on exploratory drafts just in case a project or urgent client task needs a careful final review later.

Try this simple rhythm:

1. Route: decide whether the assignment is routine, substantive, creative, or consequential.
2. Scope: supply the approved sources, desired length, audience, output, and stop rule.
3. Draft: use the middle setting for the normal pass.
4. Challenge: run a claim-by-claim or field-by-field review.
5. Escalate: use deeper reasoning only for unresolved issues that justify it.
6. Approve: a responsible human checks the result before sending, publishing, or acting.

At the end of a week, note what consumed the most allowance. Was it a genuinely difficult analysis, giant attached files, repeated searches, or five rounds of avoidable revisions? Fix the underlying workflow. Simply downgrading every task may save usage while creating a more expensive review problem.

The goal of AI for financial advisors is not to use the smallest model at all costs. It is to use the right amount of intelligence at the right moment. Financial advisors can move faster without giving up the habits that make their work trustworthy: clear scope, reliable sources, explicit approval, and a record of what was checked.

Craig Turner's live webinar, Use AI to Clean Spreadsheet Data and Build Dashboards, is Tuesday, September 22, 2026, from 4–5 p.m. Eastern. It is the eighth distinct live class in Craig's continuing AI for financial advisors series.

If you want to practice these methods in an advisor-focused setting, see Advisors4Advisors' Codex self-study course. Check the current course page for its content and applicable credit details before enrolling.

Editorial note: Model availability, account limits, and product controls can change. Verify current settings and usage in your own Codex account.