The Finance AI Operating System

The productivity evidence is substantial and consistent. In a controlled experiment on mid-level professional writing tasks, Noy and Zhang (2023) found that assistance from a language model reduced completion time and raised the graded quality of the output, while narrowing the gap between stronger and weaker writers. In a larger field experiment, Dell'Acqua and colleagues (2023) had 758 Boston Consulting Group consultants complete tasks with and without AI; on work that fell inside the model's competence, the AI-assisted group produced more, finished faster, and earned higher quality ratings. For the drafting and structured analysis that occupy much of an analyst's week, the case for productivity gains is well supported.

The same studies are equally precise about the limits. Dell'Acqua and colleagues describe a jagged technological frontier: capability is uneven, and the boundary rarely announces itself in the output. On a task deliberately designed to fall just outside the model's competence, the AI-assisted consultants were wrong more often than those working without the tool, and the incorrect answers read with the same fluency as the correct ones. In these systems, fluent output is a property of how the model writes rather than a signal that its content is correct.

Three further findings bear directly on how the output should be reviewed. Professionals misjudge their own results with these tools. A 2025 randomized study of experienced open-source developers (Becker and colleagues, METR) found that participants worked more slowly with AI assistance while believing it had sped them up, so a sense that the work went faster is not evidence that it did. Fabrication persists even in software built for the job: a Stanford RegLab study (Magesh and colleagues, 2025) evaluated legal-research products marketed as reducing hallucinations and still found incorrect or unsupported answers on a meaningful share of queries. More context does not reliably help either. Liu and colleagues (2024), in "Lost in the Middle," showed that models attend most to the beginning and end of a long input and can overlook material placed in between. Alongside these accuracy limits sits a documented security surface, the OWASP Top 10 for Large Language Model Applications, which covers risks such as prompt injection and the disclosure of sensitive information.

The picture is a capable but uneven collaborator whose output has to be verified rather than trusted. That is not a case against the tool. It is the reason finance and accounting are unusually good candidates for AI assistance. The work is structured and example-rich, so a prior period can show the model what a good result looks like. Much of it is document-heavy drafting and extraction, the tasks these tools handle best. Its deterministic pieces can be walled off into formulas and code. And the profession already runs on a preparer-and-reviewer culture, which the tool joins as a preparer rather than a replacement for the reviewer. Those four properties are what the rest of this module turns into a repeatable method.

Each task in this course runs through the same five moves. Every later module applies these same five moves to a new workflow, so the parts that follow take each one in depth.

1. Map it to a worked example. A task has the shape of a journey with three anchors: where you are (the materials you start with), where you are going (a concrete example of the finished output), and the route between them. The destination anchor does the most work. A friend's strong resume shows what "good" looks like; your draft and the job posting are the starting point; the model handles the route. A finance process has the identical shape: last period's inputs and finished output are the worked example, this period's inputs are the new starting point, and the model replicates the pattern. When you genuinely have no destination, the model becomes a brainstorming partner that widens the option set, but the choice of endpoint stays with an expert, because choosing the conclusion is on the not-for list.

2. Split the deterministic math into code. Where a step has one correct answer for a given input, a variance, an amortization line, a tax-table lookup, have the model write the formula or spreadsheet and let that artifact run the numbers, rather than asking the model to do the arithmetic itself. The model is excellent at writing the formula; it is a weak engine for running it.

3. Fuel it with a minimal folder. Give the tool exactly what the task needs, delivered as a curated folder: the inputs you normally receive, one example of the output, and little else. "Lost in the Middle" is the reason a wider pile hurts, because it buries the few lines that actually matter.

4. Guard it, then 5. review it. Confirm the data is cleared for an approved instance and the folder is scoped, keep the work inside the normal sign-off, and review before anything ships. The model is a fast preparer, not an approver.

Scaffold it so it repeats. Once a task runs cleanly two or three times, capture it: a reusable prompt or a packaged skill that encodes the framing, the worked example, and the deterministic split, plus a short review checklist you run each time (verify each source, own the conclusion, trace every number, reconcile the totals, strip the AI phrasing). The scaffolding is what turns a good one-off into a process a teammate can rerun next period.

Resolving it means changing what the AI is asked to produce. When a workflow hits a deterministic component, do not have the AI perform the calculation; have the AI build the formula, spreadsheet, or code for it. The artifact then runs deterministically and is validated once, rather than re-verified every time you use it. The model is excellent at writing the formula; it should not be the thing that runs it.

This single split shapes every later module. In close and reporting, a template computes the variances and the model narrates them. In cash forecasting, the model builds the 13-week model and then writes scenarios on top of it. In valuation, the model constructs the discounted-cash-flow skeleton and you validate the math against an assumptions ledger. In every case, the deterministic part becomes code you can trust, and the model is reserved for language and judgment.

Context is the single largest determinant of output quality, and the instinct to give more of it is usually wrong. Research on long contexts shows that accuracy degrades when the relevant information is buried in the middle of a large pile; models attend most to the beginning and the end. Oversupplying context wastes tokens and lowers the quality of the output.

So the principle is minimum necessary context: give the tool exactly what the task needs and leave the rest out. A resume does not require your full life story. The cleanest way to deliver that context is a curated folder: the inputs you normally receive, one example of the output you expect, and nothing else. The curation happens before an agent ever sees the folder.

Live connectors to systems like an ERP or CRM sound ideal, but they often underdeliver today. The moment an AI is connected to a full instance of a complex system, it is drowning in context, which is the opposite of the minimum-context principle. Broad, live integrations tend to bury the model in far more than the task needs.

The dependable pattern is simpler: run the report in the source system, download it, and feed the file. A targeted export is pre-curated context; it embodies the minimum-context principle automatically, because you have already narrowed it to what the task requires. Narrowly scoped connectors already work well, and the calculus will keep improving as designs mature. For now, when in doubt, download the report.

No output ships without a human review, and the review has five points. Verify every source: even purpose-built professional AI tools have been found wrong on a meaningful share of queries, and invented citations have drawn court sanctions. Own every conclusion: the model supplies options, but the judgment is yours. Strip AI writing patterns: remove em dashes and formulaic framings; rewrite until it reads the way you would explain it aloud. Trace every number: follow each figure back to its source. Reconcile: totals foot, cross-references agree, and no conclusion overstates the evidence.

Governance sits alongside review. Company, client, and personal data belong only in approved enterprise instances, never in consumer accounts or on self-published sites; when in doubt, ask before you upload. Agents raise the stakes because they can see and modify a whole folder, so point them at working copies, keep untrusted files out, and remember that minimum context is a security control as much as a quality one. And AI-assisted work stays inside existing controls: the same preparer and reviewer sign-offs, the same audit trail. The model is a fast preparer, never an approver.

The moves are clean on the page; live work is messier, and a few failure modes recur often enough to name.

• The adversarial checker agrees for the wrong reason. A second model trained on similar data can share the first model's blind spot, so it confirms a fabricated figure or citation with the same confidence it would show for a correct one. Vary the vendor and the framing, and read a clean adversarial pass as a lowered risk rather than a cleared one.
• The deterministic split quietly leaks. A model asked only to narrate will sometimes restate a number in prose, and if that restated figure drifts from the template, two versions of the truth now sit in one memo. Tie each figure in the narrative back to the template rather than trusting the sentence it sits in.
• The curated export is stale or silently filtered. A folder can tie out internally and still be wrong if the export was last quarter's, or a saved filter dropped a region, so the numbers reconcile to the wrong source. Check the export's date range and row count against the source system before you trust the tie-out.

These are not reasons to skip the workflow; they are why the guard and review steps are not optional.

Below is the five-move method drawn as the pattern you will see in every module: inputs, an AI step, a human checkpoint, and a finished artifact. In later modules the diagram is specific to that workflow. Here it is generic, so you can map it onto a recurring task of your own. Open each step to see how the move plays out.

Every module opens with a thirty-second governance check. Run it here on a task from your own work: is the data cleared for the tool, is the folder scoped, are approvals and the audit trail in place. Doing this until it is automatic is half the point of the course.

This module's lab is reflective rather than data-driven. Download the folder, map one of your own recurring tasks as a three-anchor journey, mark which parts are deterministic, and fill in a personal red-lines card you can keep at your desk. The card is the deliverable, and it is the deliverable you keep using after the course ends.

Even a reflective deliverable gets checked. Work the validation checklist against your journey worksheet and card, then score yourself on the rubric. Your scores start filling in the workflow maturity dashboard on the course hub, which tracks how each of the eight learning objectives is developing as you move through the modules.