Teach Hermes an accounting procedure once. Improve it every time.

Most AI accounting experiments have the same frustrating ending: the assistant finally understands what you wanted, the output is useful, and then the conversation ends.

All of the guidance that made the result good—the accounts to inspect, the exceptions to ignore, the evidence to include, the format the reviewer expects—is left behind in the chat.

Hermes Agent offers a much more interesting possibility. After you teach it how to handle a procedure, Hermes can write that procedure into a reusable skill. The next review does not have to begin with another blank prompt.

Use Connect with Hermes, and that skill can work with current QuickBooks data across the companies a user is authorized to access. A good accounting procedure can become something the firm runs, reviews, and improves over time.

That is where this starts to feel less like “chat with QuickBooks” and more like a new way to build accounting workflows.

The most interesting thing about Hermes happens after the first run

Hermes Agent is an open-source AI agent from Nous Research. Like other tool-using agents, it can investigate a question, call external systems, work through multiple steps, and return a finished result.

What makes Hermes especially compelling is what its creators call a closed learning loop.

You give Hermes a task. It attempts the task. You point out what it missed and show it what a better result looks like. Hermes can then save the refined procedure as a skill and load it the next time a similar request comes up.

Hermes describes skills as “procedural memory.” For an accounting firm, that could mean preserving much more than a clever prompt. It could mean preserving how the firm actually performs a review:

• start with a P&L comparison;
• investigate movements above the reviewer’s threshold;
• exclude locations that opened during the comparison period;
• trace unusual balances to the source transactions;
• show the current account beside every proposed reclassification;
• prepare changes without posting them;
• return a short controller summary and a detailed staff schedule.

Those instructions are not generic AI knowledge. They are the decisions that turn a plausible answer into a useful piece of accounting work.

From one good review to a repeatable workflow

Imagine a CAS manager asks Hermes:

Review repairs and maintenance across this restaurant group. Explain the increase from last quarter and identify transactions that may need reclassification.

Using Connect, Hermes can gather the reports, move into the relevant general-ledger accounts, and inspect the transactions behind the increase.

The first result probably will not be perfect. The manager may explain that new locations should be removed from the comparison, recurring vendor increases should be separated from unusual purchases, and equipment above the firm’s capitalization threshold should be flagged. They may ask for a different summary or require source transaction links for every recommendation.

Normally, that feedback improves one conversation. With Hermes, it can improve the procedure itself.

Hermes can write the finished approach into a quarterly-repairs-review skill. When the next quarter closes, the manager can ask for the review again without reconstructing every instruction. Hermes loads the skill, Connect supplies the current books, and the reviewer gets a familiar result built from current data.

If the reviewer adds a better test on the next run, the skill can improve again.

This is the part we think accounting firms should pay attention to. The skill is not merely automation. It is a place to accumulate the small corrections, judgment calls, and presentation choices that make a recurring procedure dependable.

What could a firm teach Hermes?

Start with procedures that are frequent enough to be tedious and structured enough to review.

A month-end variance review skill could remember which comparisons to run, how the firm applies materiality, when to inspect ledger detail, and how to distinguish recurring, timing, and unusual movements. It could prepare the same close-note format each month while still investigating the facts in the current books.

A stale receivables review skill could gather the aging report, inspect invoices and customer activity, identify unapplied payments or credits, and separate disputed balances from straightforward collection issues. It could prepare the staff follow-up list while leaving customer communication for approval.

A multi-entity management fee skill could preserve the allocation method, participating companies, due-to and due-from treatment, rounding rules, and required tie-out. One request could prepare the coordinated entries across the group for review.

A year-end vendor review skill could check vendor records and payments, surface missing information, identify duplicate records, and prepare the exception list needed for 1099 follow-up. The accountant still makes the tax determination; Hermes handles the gathering and first-pass review.

None of these procedures becomes static. That is the appeal. A senior reviewer can keep teaching the skill as new exceptions appear instead of burying those lessons in old chats or asking every staff member to rediscover them.

Connect gives the skill somewhere useful to work

A well-written skill still needs reliable access to the accounting system. This is where Connect changes what Hermes can do.

Through one connection, Hermes can work with live reports, accounts, customers, vendors, and transaction detail from the QuickBooks companies the user is allowed to access. It can compare companies, follow a balance into the underlying activity, and prepare coordinated changes without asking someone to assemble exports first.

Just as importantly, Connect keeps the safeguards around that access. Intuit credentials remain outside the Hermes conversation. Proposed changes can be dry-run first. Real writes remain subject to human approval and limits.

That pairing lets each product do what it is good at. Hermes remembers and carries out the procedure. Connect gives the procedure controlled access to the books.

Better models make the workflow better

Frontier AI models are becoming remarkably good at long investigations. They can move between reports and transaction detail, keep track of several entities, notice exceptions, and revise their approach when a reviewer explains what went wrong.

But a better model alone does not give a firm a repeatable process. Without a durable procedure, every impressive result is still a one-off.

Hermes skills give that reasoning somewhere to accumulate. The underlying model can change while the firm’s procedure remains. A reviewer can improve one instruction and make the next run better. A new staff member can begin with the firm’s established approach instead of a blank conversation.

The result is not an AI replacing the reviewer. It is a reviewer turning their guidance into a workflow the rest of the team can run.

The skills should be reviewed like the procedures they contain

Hermes can create, update, and delete its own skills. That makes the system adaptable, but it also means an accounting firm should treat important skills as controlled procedures.

A firm should know who owns each skill, review meaningful changes, keep versions, and test updates against a sandbox or a known company. A skill should say clearly what it may prepare, what it must never post or send, and when it should stop for a person’s judgment.

The goal is not to freeze the workflow. The goal is to let it improve without losing control of what changed.

What we would try first

Choose one read-only procedure that a senior reviewer knows well. A monthly variance review is a good candidate because the inputs, investigation path, and expected output are easy to inspect.

Run it with Hermes and Connect. Correct the first result as you would correct a staff workpaper: show Hermes which movements matter, what evidence is missing, and how the final review should be organized.

When the result is dependable, ask Hermes to save the procedure as a skill. Run it again for another company or period. See what carries over, what needs to remain client-specific, and what new instruction would make the workflow stronger.

Only then add prepared changes, keeping Connect approval in place.

The exciting possibility is not that Hermes can answer a QuickBooks question. Plenty of AI tools can produce an answer. It is that an accounting firm can teach Hermes a valuable procedure once, connect that procedure to current books through Connect, and keep making it better.

For more about using an AI agent with current QuickBooks data, see Connect for firms, security and approvals, or the broader QuickBooks AI agent evaluation guide.