the new toolbox

AI Field Kit

Knowledge bases, skills, and agents — explained in finance language

AI did not replace the analytics map. It gave you a junior team member who can follow it — fast, tireless, and only as good as the instructions and review you provide. This page explains the AI concepts every finance professional now bumps into, using the work you already know: exports, reconciliations, close packages, and workpapers.

The field kit, in six pieces

Prompts & context 01 The instruction memo you hand the AI — scope, sources, and definitions

Skills (.md files) 02 Reusable procedures written in markdown that an agent can execute

Agents 03 AI that takes actions — runs code, reads files, produces workpapers

Knowledge bases 04 Your policies, mappings, and procedures organized so AI can cite them

RAG & citations 05 How AI grounds answers in your documents instead of guessing

The human reviewer 06 Control totals, professional skepticism, and sign-off stay with you

Prompts & context

A prompt is the memo you would write for a new staff accountant: what the task is, which files to use, what the rules are, and what the output should look like. Vague memo, vague work — same as always.

The PANDAUDIT route you already know is a great prompt structure:

“Here is the source (AP export, March, 4,812 rows, control total $2,914,003.17). Here is the rule (match to the payment file on vendor + amount, dates within 3 business days). Here is what counts as an exception. Give me Summary, Exceptions, Detail, and the control totals tying back.”

Skills (.md files)

A skill is a procedure written in a plain markdown file — a SKILL.md — that tells an agent when to use a technique, what inputs it expects, the steps, the code, and the validation checks. Think of it as a workpaper template plus a procedure narrative that a machine can actually follow.

This is the biggest shift for this site: every PANDAUDIT script now lives in the Skills Library as a skill document. You can read one like a recipe, run the code yourself, or drop the file into your agent’s skills folder and delegate the task.

Why markdown files? They are readable by humans, versionable in git, reviewable like any workpaper, and every major agent framework (Claude Code, and others) treats them as the native way to package know-how.

Agents

A chatbot answers questions. An agent takes actions: it reads your export, writes and runs the cleanup code, checks the totals, and drafts the exception memo — then shows you its work. If a skill is the procedure narrative, the agent is the staff member executing it.

Rules of engagement that will feel familiar from supervising juniors:

  • Give the agent one defined task with named sources, not “fix my close.”
  • Insist on control totals in every output.
  • Sanitize confidential data before it leaves your environment.
  • Review the exception list yourself. Judgment is not delegated.

Knowledge bases

An AI knowledge base is your institutional knowledge — chart of accounts logic, mapping tables, close checklists, policy memos, prior-period workpapers — organized so an AI can search it and cite it. The finance teams getting real value from AI are not the ones with the fanciest model; they are the ones whose knowledge is written down.

If you’ve heard of Obsidian, you already know the shape: a vault — one folder of plain .md files that link to each other with [[wiki-links]]. No database, no vendor lock-in, readable in any editor, versionable in git. That vault is a knowledge base, and it’s the ideal one for AI: agents can search it, follow the links, and cite the exact note an answer came from. Skills are simply the executable notes in the vault — procedures instead of prose.

Start embarrassingly small: one folder of markdown files. Your account-classification rules. Your recurring reconciliation procedures. Your fiscal-calendar definitions. Each document you write becomes something both new hires and agents can use. The knowledge-vault tutorial walks through the whole setup in five steps.

RAG & citations

Retrieval-Augmented Generation (RAG) means the AI first retrieves relevant passages from your knowledge base, then answers using those passages — with citations back to the source. It is the difference between a colleague who says “per the policy memo dated March 3, section 2…” and one who says “I’m pretty sure it’s fine.”

For finance work the rule is simple: no citation, no reliance. An AI answer about your data or your policy should point at the file it came from, the same way a workpaper points at its support.

The human reviewer

Nothing on this page removes the reviewer. It moves you up a level — from preparer to reviewer of a very fast preparer:

  1. Control totals first. Rows in, rows out, amounts tied. If they don’t tie, stop.
  2. Exceptions second. The agent’s job is to shrink the review list, not hide it.
  3. Spot-check detail. Sample the matched items, not just the flagged ones.
  4. Document the delegation. The skill file + the agent’s output is your audit trail.

Where to go next