Capital

The Agent Paying Your Invoices Is Not Software

Freehand raised 75 million dollars to run supply chain spend with AI agents. It is priced against the labour budget, not the software budget.

CapitalBy Servola Tech Desk2026-07-316 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

The Agent Paying Your Invoices Is Not Software

Key takeaways

Freehand's own pitch sets $348 billion of supply chain labour against $16 billion of supply chain software, a ratio of nearly 22 to 1, and the product is aimed at the larger number. An agent bought out of the labour budget does not automatically pass the software review that would have examined it.
The company describes its product as reading contracts, negotiating with suppliers, catching leakage and processing payments. Negotiating a rate creates an obligation and processing the payment settles it, and internal control exists to keep those two in different hands.
Neither the funding announcement nor the independent reporting names an approval threshold, a human sign-off or an audit trail. Nothing says the controls are missing. They are unspecified, and an unspecified control is one you write into the contract or nobody writes at all.
The headline benefit is a savings claim: 5 to 10 percent of spend recovered in complex categories. On a European group with 400 million euros of addressable category spend that is 20 to 40 million euros, and it is measured on money saved, never on a decision correctly refused.

Nitin Jayakrishnan sold the software company, then raised for the labour

Freehand said on 29 July that it had raised $75 million in a Series B co-led by Battery Ventures and NewRoad Capital Partners, with PSP Growth, the vehicle led by former United States Commerce Secretary Penny Pritzker, and Nexus Venture Partners taking part. The San Francisco company has now raised $100 million in total, following a $25 million Series A in March 2024. Crunchbase reported the valuation as undisclosed, with the chief executive describing it as a significant step up.

The founders had been here before, on the other side of the trade. Their previous company, Pando, sold transportation management and procure-to-pay software, and it went to a strategic buyer in early 2026. Nitin Jayakrishnan, co-founder and chief executive, explained the pivot in terms of what he thought was ending: "Instead of trying to catch them up to a technology paradigm that was sunsetting, we thought we could leapfrog them into a technology paradigm that was just rising." His co-founder Abhijeet Manohar put the same idea more bluntly. "Shifting from building software for the user to building software that is the user meant we could drive far deeper transformation."

Roughly fifty customers are on the platform, among them Meta, Unilever, Johnson & Johnson, Pfizer, Dunkin' and Cardinal Health. Freehand says it processes billions in payments across sixty to seventy countries. Those are large, audited groups with mature procurement functions, which is what makes the next number worth reading carefully.

The ratio buried in the founder's own sentence

The number that explains this round is not the 75 million. It is a ratio Jayakrishnan states himself: "Enterprises spend $16 billion a year on supply chain software and another $348 billion hiring people to do what this software cannot." Set those two against each other and the labour line is 21.75 times the software line. Software is 4.4 percent of the $364 billion the two figures make together.

Put the larger number against the company's own estimate that American firms spend more than $20 trillion a year on supply chain, and the labour it wants to replace is about 1.74 percent of the flow it sits inside. That is a small share of an enormous number, which is exactly the shape of market a growth investor likes.

Read as pricing rather than as rhetoric, the sentence says something specific. Freehand is not competing for the software line. It is competing for the headcount line, and it has raised $100 million to do it.

Why it matters: the budget line decides which review you get

Why it matters. In most companies the approval path is chosen by the budget the money leaves, not by what the thing actually does. Software goes through a vendor security review, a data protection impact assessment where personal data is in scope, an architecture look and a procurement checklist. Headcount and outsourced services travel a different road, one built to check references and rates rather than data flows and access.

So an agent that reads your contracts, negotiates with your suppliers and moves your money can enter the business through the door that has no software review on it. None of that is a trick by the vendor, and the point is not that this particular product is unsafe. It is that the governance you apply is set by an accounting decision taken before anyone technical sees the thing.

The obligations do not follow your accounting. For an in-scope European entity, NIS2 supply chain security duties and, in financial services, DORA's third-party ICT rules already reach an arrangement of this kind whatever budget bought it. The internal classification does not travel with the risk.

Four verbs, two duties that are meant to stay apart

Yes, but. The company's own description of the product spans four actions: read contracts, negotiate with suppliers, catch leakage, process payments. Negotiating the rate is what creates the obligation. Processing the payment is what settles it. Internal control keeps those two in different hands for a reason that has nothing to do with artificial intelligence, which is that a single actor able to both commit the company and release the money is the oldest weakness in a finance function.

Neither the announcement nor the independent coverage names a value above which a person signs, an exception queue for contested invoices, or an audit trail a reviewer could follow after the fact. That is not evidence the controls are absent. Early enterprise deployments at groups the size of Pfizer and Unilever will have had controls negotiated into them. It is evidence that the controls are unspecified in public, which means the buyer supplies them or the buyer inherits whatever the default happens to be.

The published results point the same way. Customers recovered 5 to 10 percent of spend in complex categories, completed workflows 5 to 7 times faster, and cut procure-to-pay cycles by more than 70 percent. Every one of those measures throughput or savings. Not one measures a payment correctly refused, and that is the number an auditor will eventually ask for.

Write these down before the pilot starts

Set a value threshold above which a named person approves, and put it in the contract rather than in the configuration, because configuration changes without a signature and a contract does not. Agree it before the pilot, when you still have leverage, rather than after the savings figure has been presented to your board.

Require an exportable audit trail that shows what the agent read, what it proposed and what it executed, timestamped and retained on your side rather than the vendor's. Then settle who the counterparty is on an agreement the agent negotiates, because a supplier that later disputes a rate will address the company whose name is on the contract, not the system that typed it.

Name the exit while you are still buying. Ask what happens to the negotiated terms, the supplier relationships and the payment history if you stop in eighteen months, and get the answer in writing. A tool you can switch off is a procurement decision. A function you have dismantled your own team to buy is something else.

Frequently asked questions

Is this a software purchase or a services purchase?

Commercially it is sold as work performed rather than as a tool licensed, which is the whole point of the pitch. Operationally it reads your contracts, holds credentials in your systems and moves money, which is software behaviour. The mismatch matters because most approval processes are triggered by the first description and designed for the second.

Does the EU AI Act settle how this has to be controlled?

Not on its own. AI Act duties attach according to the role you hold and the risk category a system falls into, and an agent negotiating supplier terms is not automatically in a high-risk category. For most buyers the binding constraints are the ones already carried: internal control and segregation of duties, NIS2 supply chain security for in-scope entities, DORA third-party ICT rules in financial services, and whatever the contract says.

What does 5 to 10 percent of spend recovered actually mean?

It is a savings figure measured on complex categories, not on total spend, and it is the vendor's own number. On a European group with 400 million euros of addressable category spend it implies 20 to 40 million euros. Ask which categories were in the denominator, over what period, and who verified the baseline, because a recovery rate is only as good as the spend it was measured against.

The interesting disclosure in this round is not the amount raised. It is that a founder who spent years selling supply chain software has concluded the money is in the budget line next to it, and has priced his company against payroll rather than against licences. He is probably right about where the spending is. Owners should take the same sentence and read it the other way round: if this is bought as labour, it will be reviewed as labour, and the review that would have asked where your contracts are going will never be scheduled. Decide which budget it comes out of deliberately, because that choice, and not the technology, determines what gets checked.

FreehandBattery VenturesProcurementAI agentsSegregation of dutiesVendor riskSeries B

A Sorting Belt Camera Now Counts as Evidence

Greyparrot raised 27 million dollars on 28 July. The larger fact sits inside the same announcement: a UK regulator has accepted AI-generated waste composition data for statutory reporting, weeks before two fee regimes start pricing recyclability.

5 min read

Capital

16,000 Firms Pooled Their Supplier Risk. You Didn't

Risk Ledger raised 24 million pounds in Series B on 15 July 2026 for a network-first supplier-risk platform used by 16,000 organisations. Why NIS2 and DORA make the questionnaire model your problem, and what European owners should change.

4 min read

Law & Regulation

The EA Buyout Now Hinges on Brussels

The largest take-private in history has cleared EA shareholders and US antitrust. The last real gate is the EU - and not classic merger review, but the Foreign Subsidies Regulation, facing its first blockbuster test.

3 min read

← All articles