Josh Bligh leads finance at Black Forest Labs, the company behind the FLUX image models, founded in 2024 by researchers behind Stable Diffusion. It raised $300 million at a $3.25 billion valuation in December 2025. He has spent more than 20 years in accounting and finance operations, including credit and collections at 3PAR, Arista Networks and Elastic. He opened an AI tool for the first time in February 2025, and has since been building agents into a finance function that runs on an outsourced accounting team and an AI-native stack.
An agent can do most of the collections function but a person should still be the one who presses send.
I ran credit and collections for years at large public companies, and the work looked the same in each company. A team on the phones and in the inbox worked through the aging list, most people paid, and a handful of hard accounts at the bottom needed skill and attention.
“I don’t allow it to automatically send the emails. That is one area of collections I don’t know that you’ll ever get around. There’s always nuance, there’s context, and I don’t want to create a customer problem with an email.”
So the collector I’m building at Black Forest Labs runs on Claude with an MCP connection into our ERP. It drafts every collection email across 6 levels, from the day an invoice comes due to 90 days past due. It sends none of them without my review.
Let the agent work the 90%
“It’s dialing for dollars. 90% of it is phone calls and emails, and people will pay. Then the 10% left over is your focus. But in the old days you had to have 6 people on the phones making those calls.”
The agent takes the volume. It watches the AR aging on a set cadence, roughly weekly, though there is always a discussion about what the cadence should be. Every invoice that comes due gets a friendly follow-up. 15 days later the ask ratchets up, and it keeps climbing until about 90 days. Anything still open then drops out of the automated process and gets worked from a dashboard by a person.
The output goes to Slack. Each day the agent posts a summary into a channel with the day’s work and the drafted emails. Someone reviews them, clicks send, and the emails go to customers from there. There is no Zapier and no middle layer, and I like it for the access. Whoever should be doing collections work gets added to the channel, and that is the setup.
The next step is replies. Because the agent is connected to the ERP, a customer who writes back saying they don’t have a copy of the invoice should get the copy or a statement without anyone going to look for it. I’ve tested it internally and while it works, it’s quite clunky. The friction right now is me getting it finished myself.
I wrote all 6 emails myself
The first version used the agent’s own drafts. They were fine. They weren’t offensive and they weren’t wrong. They just didn’t feel like something I’d want to receive as a customer. It was 30 minutes of work, it wasn’t quite what I wanted, so I wrote them myself.
That sounds like a small complaint. With anything agentic that emails people, it’s the main risk. What the agent writes is what your company looks like to the customer, and there’s a fine line between over-communicating and communicating the right amount.
“The problem with collections is you’re already the aggressor any time you send an email as a collector. And you don’t want to make the salesperson’s job much more difficult by creating a problem in the relationship when there really doesn’t need to be one.”
If someone hasn’t paid in 90 days and is giving me the runaround, I’m willing to burn the bridge. On the first email in 15 days I’m not. If I write to say an invoice was due 15 days ago, the customer replies that they never received it, and the agent comes back with something odd, I’ve made a customer problem out of nothing. That reply needs to be a lot softer.
Now the agent holds my 6 emails. It swaps in the customer name and the invoice numbers and leaves the language alone.
Use an agent to hold the purchasing process
We didn’t build every agent we run ourselves. Ramp’s procurement agent sits in front of purchasing and follows the workflow we’d have written down anyway.
In a startup, purchasing isn’t always an interesting conversation. Companies say they want a buyer and they want control, and then they say yes to everything. What goes wrong is quieter. There’s a lot of data to collect when you set up a new vendor, and you can buy without collecting it. A year later finance has a problem because the W-9 was never collected, or someone signed a contract nobody knew about.
The Ramp agent captures that detail during the purchase. I wouldn’t say it made anyone’s life better. It solved an admin problem that was starting, and I didn’t have to build out a purchasing department to do it.
The agent told me it all looked good
I was going through our fixed asset register with an agent and it came back saying everything looked good. Then I noticed a kitchen remodel had been fully expensed in 5 months. The ERP takes useful life in months, so 5 years has to go in as 5 times 12, which is 60. Someone had typed 5.
When I put it back through, the agent found it. A smarter prompt would have caught it the first time and the pattern worried me.
“You can put garbage data in and it flows all the way through. And as we become less connected to the data, because it’s just agentic or it’s SaaS driven, there’s no human eye on it.”
That’s why the interface between the CRM and the ERP interests me. In the old days that was the deal desk and order management. A salesperson filled in a form, the deal desk said yes or no, and someone keyed in the details. Then software connected Salesforce to NetSuite and the review in the middle quietly went away. An agent can put some of it back by reading the contract and checking it matches what’s in the ERP. I’m a control freak about this, and one of the things I liked about having humans in that seat was the review.
Nobody had 3 to 5 years of AI experience
My AI story starts in February 2025. Before that I worked at a company where we were forbidden from using AI in the building or during working hours, and I had no need for it at home, so I never touched it.
In October 2024 the investors told us they were only putting money into AI companies now, and that we should package ourselves for sale or close. As head of finance, however I drew the chart, it hit zero. I had about 4 months either way. So I started looking, and every job posting was AI.
“The ridiculous part is they’re saying we need people with 3 to 5 years of AI experience. There’s nobody out there, except for a very few people.”
I went to ChatGPT, Gemini and Claude and started putting my daily problems through them. Some things worked really well and some didn’t, and I figured out where it fit.
One thing I know now is that agents love formatting. In a past life I worked somewhere billing about 3,000 invoices a month at $9.99, all on card, and some would fail. The bank statement came back with 2,873 transactions of $9.99, told apart only by bank numbers and transaction IDs. The first time I reconciled it I spent 16 to 20 hours comparing 2 spreadsheets and marking each cell green. Now I could hand both statements to an agent and tell it to fix it.
I think a lot of this settles the way SaaS did. It was going to solve everything, everyone bought SaaS for everything, and then it consolidated. 5 years ago a finance person could set up a whole department with the same 5 products, and that was light years better than 2000, which was light years better than 1990.
“Specifically to finance, the results of your work don’t really change. We have to report, we have to be audit-ready, we have to be a dashboard to the company. It’s the tools we use behind the scenes that change.”