The agentic AI phase is upon us and this year looks to set to be the year when this begins to really dominate. We are very excited to see Cowork in Windows available at last. Much has been waxed lyrical about Claude Code and Cowork and we have high hopes that it will work better for us than ChatGPT's agent which we tested last year.
Contrary to conventional logic, we start out by giving Cowork no quarter and asked it to perform a multi-step task with a one-shot long prompt. That experiment failed and it hanged at a certain instruction step. We waited a day with no visible progress so we decided to cancel and start afresh from the other extreme instead : baby steps.
We first used Claude in Excel to prep the data. We were impressed with how the agent took in our instruction and it even gave us the option to execute its plan step-by-step. It also gave us the option to let it just execute everything all at once but it was nice to see it can give this level of assurance to first time users.
With the data prep, we continue with the baby step approach for Cowork. You must be wondering what we were trying to do to make the agent work. The task was to prepare from management accounts, a consolidation of a group with one subsidiary and then compile the entire consolidated financial statements using our template.
The consolidation data entry was performed with Claude in Excel. We then manually passed the necessary consolidation journals into the template. Cowork was used to mock up the figures into our Word template. We started with the individual financial statements first; the income statement, balance sheet and statement of changes in equity. We saved the cashflow statement for last (baby steps remember) and was amazed that Claude could do a simple cashflow statement using the indirect method.
However, at this stage we hit the first token limit of our Pro subscription. After the reset, it finished up the cashflow statement and we then tasked it to do up the entire financial statements with accounting policies and disclosure notes to accounts.
It remarked that this was a very large piece of work and true enough we hit the token limit again. While we were surprised at how fast the tokens are being depleted, it is also a good gauge for us to know much compute is going into the tasks it was doing.
So you can imagine by now we have a good idea on the scope and scale of what it can or cannot do with the subscription tier we're on.
At this point, a quintessential question might be "are we in danger of being replaced yet?". I would say we are definitely in the "washing machine" stage for this industry/profession. Just how good and intelligent the "washing machine" is will depend on our instructions and the fuzzy logic the machine has.