Nearly New Cashmere · Finance
An agentic finance department that catches a wrong turn in time.
A forecast built with Claude sets the targets. Every month, the actuals and the assumptions behind the plan are checked against it, and so are the decisions that depend on it, from marketing spend to stock. So the business can change course before it buys too much or cuts the wrong budget.
1 hour
to close the month, by one person, where it took a day and several people
Within a week
of month-end, the month is reviewed, where it used to wait until the end of the following month
Every assumption
in the forecast checked against what actually happened, every month

The business
Nearly New Cashmere restores and recycles discarded cashmere and sells it online, alongside its own recycled cashmere range. It sells in the UK, the US and beyond, online, in its shop and at events.
Growing fast means making big calls early: how much to spend winning new customers against keeping existing ones, and how much stock to buy for a season that has not happened yet.
The challenge
The business already had a strong forecast, built from its customers: how many new and returning customers each market brings, how often they order and what they spend. Keeping it honest was the hard part.
Which forecast was the real one? As the plan was revised through the year, several versions built up, each sensible at the time, and it was not always clear which one the targets should come from.
The forecast and the accounts spoke different languages. The forecast was in total sales. The accountant works in sales net of VAT, and the bank sees cash after the payment platforms have taken their fees. Turning one into the other was done by hand, so there was no clear cash-in-the-bank target to hold the month against.
Closing the month took several people most of a day, and the figures never quite matched the accountant's. Meanwhile, the decisions that depend on the forecast were made without a live check against it. When an assumption is wrong, the cost shows up later: stock bought for demand that did not come, or marketing cut back when it was working.
What we built
An agentic finance department, in five connected parts. The maths always stays in the model. The agents never calculate a number themselves: they ask the model, explain the answer and propose. A person approves every change.
1. One forecast, in every language. Built with Claude from the levers the business controls, by market: new and returning customers, how often they order, what they spend. The owners choose the growth plays, and every version is saved, dated and compared with the last, so there is always one frozen plan that the targets come from. Each target is carried through three lines: total sales for the owners, sales net of VAT for the accountant, and cash in the bank after platform fees. The translation follows the accountant's own method, down to VAT applying only to sales shipped within the UK.
2. Close the month. Pure automation, no AI. Sales, payments, shop takings and the accounts land in one finance store, on the accountant's own method, and the actuals are written against the targets in about an hour.
3. Watch the assumptions. Every driver in the forecast is tracked against the assumption behind it. When one drifts, it is flagged at month-end, not discovered at year-end.
4. Check the decisions. Marketing spend and margin are checked against the plan and agreed limits, including the balance between winning new customers and keeping existing ones, so neither is overcorrected on a single month. Stock buying is the next decision to connect, so a purchase can be tested against the demand the forecast expects.
5. Explain and propose. When the business misses or beats the plan, agents break the gap into its causes and investigate releases, campaigns and returns. They work inside firm boundaries: an explanation can only rest on the assumptions in the forecast and the evidence in the data, every number must trace back to a source, and when the cause is not clear the agent says so rather than guessing. Any reforecast it proposes goes to the owners to decide. Parts 4 and 5 are being built now.
Why it can be trusted
It is the source of truth. The month's figures match the accountant's books. The only differences we found came from time zones at the very edge of a month, an order placed just before midnight in one system landing just after it in another, not from anything missing or wrong.
Far less is done by hand, so there are far fewer chances for a mistake. Nobody retypes a figure from one system into another, and nobody translates the forecast into the accountant's terms on a calculator. The figures arrive already checked, on one set of definitions.
In practice
In August, new customers in one overseas market came in at less than half the plan. The forecast's assumptions flagged it as the month closed, and the plan was reforecast within ten days, so the targets for the rest of the year reflected what was really happening.
What changed
One person closes the month in about an hour, and the owners review it within a week of the month ending, instead of at the end of the following month. They get the same one-page summary every time: whether the business hit its plan, and what made up the result.
That speed is what makes the business nimble. Decisions are checked against the plan before they are made, not after, so a wrong turn on spend or stock can be corrected while there is still time.
Where it is today
Live now: the forecast and its growth plays, the monthly close, the one-page summary and the assumptions tracker.
Which process is growth making harder for you?
Bring it to a first conversation. We will tell you honestly whether AI, automation or a better-connected system is the right answer, and what it could be worth.
Book a call