Alamin Mollick is SVP, Finance at Posh AI. He has been the first finance hire at five startups, helping companies scale from $0 to more than $25 million in ARR. Before joining Posh, he led finance and corporate development at Rewind and held finance and operations roles at Combyne.
“The more repetitive operational work? Yes, it has aided. But it’s marginally aided in FP&A.”
I’ve had mixed results using AI across finance. It’s made a real difference in recurring work across AR, AP and accounting, but it’s been harder to get the same results in FP&A.
The issue isn’t necessarily access to information. I can give an agent Slack conversations, emails, meeting notes and financial data and have it find most of what I need. The agent can struggle with what matters once you have the information, whether that’s understanding why we’re off plan or how the customer relationship should shape a collections follow-up.
Giving a collections agent the full customer picture
“When you’re dealing with $100,000-plus invoices, these are buyers that are the director or VP. I still think there’s an element of human relationship involved.”
We built an agent that runs weekly and helps us manage collections. It connects our billing system with HubSpot, Slack, Gmail and other places where information about a customer might live.
It starts with the AR data. It might tell us we have $2 million outstanding across 18 customers, show which are 30 to 60 or 60 to 90 days overdue and pull the invoices involved.
It’s set up to pull things like how long the customer has been with us, what we’ve been working on with them and recent conversations across email and Slack.
It uses that context to recommend who should follow up, whether it’s someone in finance, me directly, the CSM or whoever has the strongest relationship with the customer.
Then it drafts the email and posts it in our billing channel in Slack for that person to review, edit if needed and send.
Finding the right context takes work
Connecting the agent to Slack or HubSpot doesn’t mean it automatically knows where the useful information is.
A customer relationship can be spread across a health channel, an account-specific Slack channel, HubSpot emails and conversations with different people internally. The CSM might be working with one stakeholder, the AE with another, while our CEO has a separate relationship with someone in the C-suite.
“You have to point it to these specific things to be able to get the detail.”
We’ve had to keep refining where the agent looks. Sometimes I’m not even in the Slack channel with the relevant conversation, so I’ll have it find which channels exist first and then decide where else it needs to look.
The same applies to the invoices themselves. A customer might have four outstanding invoices tied to different projects or buying centers. We don’t want the agent sending four separate reminders just because there are four invoices.
We have to account for who owns each relationship, whether the invoices should be grouped and how we normally engage with that customer.
There are still things it won’t catch, which is why someone reviews the recommendation before anything goes out.
FP&A is harder
I’ve had a harder time getting the same value from AI in FP&A.
For budget versus actuals, it’s useful for investigating a variance. I can ask what affected a number and have it search Slack, meeting notes and other sources for the decisions or events behind the variance.
“It’ll get the key things that impacted the number. It can pull together the inputs pretty well.”
What I haven’t been able to get it to do consistently is turn those inputs into the one or two sentences I actually need.
AI token spend is a good example. Say we’re materially above what we budgeted at the beginning of the year. Five months ago, we might have decided to invest more because it was working. Revenue has grown alongside usage, and we expect token costs to get cheaper over time.
What I actually need to explain is why we moved away from the original plan and what changed in the business.
“I’m finding these juicy bites of insights to really frame where the company is heading from what the original plan’s hypothesis was versus where we’re trying to head. It still does not get these kind of bites.”
The model might find that something was discussed in a meeting, pull the relevant Slack conversations and surface when a decision was made. But I still end up talking to one of our engineering leads, CTO or whoever owns the area to understand which parts of that actually explain the variance.
“It’ll grab all the pieces, but it has trouble… it just becomes convoluted, bloated.”
Accounting has a clearer output
The results have been different in accounting, where we’ve been able to give AI work with a more defined output.
We recently launched forward deployed engineering as a new revenue stream. For revenue recognition, we can use AI to interpret the contract, build the schedule and prepare the journal entry.
“Maybe it was three, four hours, like a half day, to really figure it out, build the schedule, test it, understand it. And now you can run it through AI and just do it in like 15, 20 minutes.”
That’s also changed how much work the same finance team can support.
“I’ve had the same team size for like three, two years, three years because of what’s happening with AI.”
I don’t think there’s one headcount ratio that applies to every company. But I do think a company growing from $5 million to $20 million could run with a two- or three-person finance team where it might have needed four or five before.