The Trustworthy Path to Agentic Month End

Just over five years ago, David and I came together to start sketching what would become Mayday. I was new to the world of accounting tech and dove in head-first. I couldn’t have found myself in a more exciting swimming pool of problems to solve.

I discovered a world where people would spend half their working life (and often more) in spreadsheets, manually reviewing transaction data, calculating adjustments, and pasting the resulting numbers back into their accounting system.

Accounting is a high-trust, high-accuracy domain. The numbers have to be correct, and not just because the accountant is responsible for disclosing them. The business will make decisions based on the numbers presented, and as we say in the software world: garbage in, garbage out. Errors in one set of numbers cascade to future sets built on top, leading to the mother of all headaches when the issue is spotted and must be unpicked.

Whilst the accounting system contains a lot of context and precedent for these adjustments, it cannot ever be complete. There will always be business context which is unique and dynamic, and which cannot always be predicted. That’s why spreadsheets won: they gave the accountant complete control over the numbers, at the cost of time, effort, and the risk of human error.

The product guy in me saw huge opportunities to take these processes and not just make them incrementally better, but to build a whole new way of doing this work.

The Month-End Close

Finance teams spend too long closing the books at month end: on average around 7 working days, but well into the teens isn’t uncommon. The business pays the price in making decisions on stale data, and let’s face it, spending that long every month in spreadsheet hell isn’t why anyone got into accounting in the first place.

Transactions come into the accounting system throughout the month, meaning finance teams could be preparing the adjustments for next month end piecemeal throughout the month. However, they wait until the month has finished before starting on these processes, for two main reasons. First, they’re too busy closing the prior month’s books to think about getting ahead of the next month. And secondly, it's tedious, important work, so it's easier to block out a few days and get it done in one sitting.

A lot of the work at month end is automatable. Well-built models can apply Generally Accepted Accounting Principles to bills, invoices and the like with a high degree of accuracy. The accounting system contains a wealth of precedent for those adjustments which happen frequently, allowing the system to learn business context rapidly.

We believe there is a route to what we call the zero-day close: books closed on the very first day of the new month, giving finance teams their time back, and allowing the business to make better decisions with more accurate and timely data.

We call this Agentic Month End, or AMEN. Building AMEN is the journey we’ve been on over the last five years, and we can now begin to see the zero-day close on the horizon.

Our Model: Manual, Assisted, Agentic

Our philosophy for building the zero-day close has intentionally followed this pattern.

Manual

As I’ve said, accounting is a high-accuracy domain. The accountant must be in complete control. They must be able to review, approve, and override any adjustment. They must be able to create an adjustment manually within the system - they will always have more business context than the system, and there will always be edge cases.

Assisted

Once there’s a system in place which allows the accountant to complete the process entirely manually, we layer on intelligence to help them prepare the adjustments. This could be in predicting the deferral treatments for prepayments and revenue recognition, or recognising potential reversals for open accruals. In this way, the accountant remains in full control, but the system starts to take away some of the mental processing load. This is the approach behind Mayday AIR, our Artificially Intelligent Recognition technology, which combines accounting logic with business context to surface recommendations.

Agentic

As the assistance layer approaches 100% accuracy, agents can be built on top of it to review its suggestions and prepare the adjustments. The system must do a good job of presenting its decisions and flagging cases which need human review. The human must be in control of which processes are completed agentically, and which the accountant wants to continue doing full manual review of.

Building the Zero-day Close

New teammates always require training and oversight, even if they come with all the qualifications required to do the job. This is to make sure that they find all the required data from the business’ systems, and are applying the rules correctly in line with the business’ policies.

Agentic software requires the same training and oversight. You wouldn’t trust a new hire to close your books on their first day - and software shouldn’t earn that right just because it’s fast and shiny. It must earn your trust through two mechanisms:

• Control: you must retain full control, by being capable of reviewing, approving, and overriding any adjustment, as well as creating new adjustments the system doesn’t have awareness of.
• Accuracy: over time it must approach 100% accuracy, learning from your corrections as it goes. When it can’t accurately predict something, it must flag the transaction for manual review.

Neither substitutes the other: control over an incorrect system creates more review work; accuracy without control is a liability.

A zero-day close is the outcome of getting this right

A zero-day close is a by-product of automating the close properly. Completing routine work throughout the period leaves less to do at month end. Without trust and control, agents can create as much checking and correction as the spreadsheets they replace.

We’re not promising a zero-day close tomorrow. With Agentic Month End, we’re building towards one, with the accountant in control of each step.