Marcus Built the Agent That Could Do Everything — Except Close the Deal

--

Press enter or click to view image in full size

# Marcus Built the Agent That Could Do Everything — Except Close the Deal

Marcus runs a boutique logistics consultancy out of Austin.

Marcus runs a boutique logistics consultancy out of Austin. Eleven employees, forty-some active clients, and a procurement operation that had been bleeding time for years. Every quarter, his team spent roughly 200 hours sourcing freight vendors, comparing rates, vetting compliance documents, and cutting purchase orders. It was the kind of work that felt important but wasn’t strategic — the operational glue that held everything together while quietly draining the people who should have been doing something harder.

So Marcus did what a certain kind of builder does when they see a solvable problem: he built an agent.

— -

The agent was genuinely impressive. Over three months, Marcus and a freelance engineer stitched together a workflow that could ingest a shipping brief, query four freight APIs simultaneously, score vendors against a weighted rubric (price, reliability rating, insurance coverage, carbon offset tier), draft a contract using pre-approved legal templates, and surface a ranked recommendation with supporting documentation — all without a human touching a keyboard.

In testing, it cut vendor selection time from four days to about ninety minutes. On a Thursday afternoon in February, Marcus pushed it to production.

By the following Monday, it had processed eleven briefs. Eleven vendor matches. Eleven contracts drafted and approved by Marcus in under ten minutes total.

And then it hit the wall.

— -

The first sign of trouble was subtle. A mid-tier freight carrier — a company the agent had correctly identified as the best fit for a time-sensitive Gulf Coast shipment — had a net-30 invoicing requirement but needed a $1,400 deposit to hold the capacity. Standard stuff. Marcus had seen this a hundred times.

The agent, however, had seen it never.

It knew the deposit was required. It had parsed the vendor’s onboarding terms, flagged the line item, and even calculated that the deposit represented approximately 8.2% of the total contract value — well within Marcus’s pre-approved budget thresholds. Every signal pointed toward: pay the deposit, lock the capacity, move forward.

But it couldn’t pay the deposit.

Not because it lacked the funds. Marcus had earmarked the budget weeks earlier. Not because the vendor was a bad match — the agent had ranked them first for four independent reasons. Not because any human had said no.

Because there was no mechanism. No wallet. No payment rail. No identity layer that would let an autonomous system send $1,400 to a vendor’s bank account without a human physically logging into a payment portal and doing it manually.

The agent did exactly what it was designed to do in this scenario: it paused and generated a task for Marcus. “Deposit required: $1,400 to [Vendor]. Approve and initiate payment to proceed.”

Marcus approved it in forty seconds. Then he logged into his bank’s ACH portal, navigated three authentication screens, entered the routing number the agent had already found, and submitted the transfer.

Eleven minutes. For a $1,400 payment the agent had already decided, verified, and contextualized completely.

— -

This happened six more times over the next three weeks. The amounts varied — $670 for a compliance document refresh with a carrier in Memphis, $3,200 as a retainer for a new drayage partner in Houston, $88 for an API subscription that one of the data-enrichment tools the agent relied on had quietly moved to a paid tier. The pattern was always the same: agent identifies action, agent prepares action, agent is incapable of executing action, human is interrupted.

Marcus started tracking the interruptions. In March alone, he was pulled back into the payment loop nineteen times. Total dollar value across all nineteen transactions: $12,400. Average time per interruption: nine minutes. Total time lost: just under three hours.

Three hours doesn’t sound catastrophic. But Marcus had built this agent specifically to reclaim his team’s time. The 200-hour quarterly drain had become an 80-hour drain — a real win. The remaining friction, though, was entirely concentrated in the payment execution layer. The agent was fully autonomous up to the moment money needed to move. Then it became a very sophisticated request form.

“I’ve essentially built a very expensive assistant,” Marcus told me. “It does all the thinking. I still do all the paying.”

— -

This is the scenario that rarely makes it into the demos. The AI agent showcase videos tend to stop at the decision — the recommendation surfaces, the contract generates, the workflow completes. What happens next, when the workflow requires actual financial execution, is left as an exercise for the operator.

The gap isn’t theoretical. It shows up in real production environments with real dollar figures and real time costs. And it compounds: every time a human has to step in to execute a payment the agent already decided on, the promise of autonomous operation retreats a little further.

The frustrating part is that the intelligence problem is largely solved. Marcus’s agent wasn’t making bad financial decisions. It was making good ones — quickly, consistently, with documented reasoning. The failure wasn’t cognition. It was plumbing.

AI agents, as currently deployed in most production environments, have no native financial identity. They can’t hold funds. They can’t initiate transfers. They can’t establish the kind of verifiable, accountable payment presence that financial systems require before they’ll move money. Every payment API on the market was architected around a human principal — someone who can be KYC’d, who can accept terms of service, who can respond to a fraud alert at 2am.

Agents are none of those things. Not yet, anyway.

— -

Marcus’s story ends the way most of these stories end: not with a catastrophic failure, but with a persistent, low-grade drain that erodes the ROI case for autonomous systems. His agent is still running. It’s still valuable. But it will never be fully autonomous until the financial execution layer catches up to the reasoning layer.

He’s not waiting for a better model. GPT-5 doesn’t solve the ACH problem. Gemini 2.0 doesn’t give an agent a wallet.

What solves it is infrastructure purpose-built for non-human principals — payment rails where the agent itself is the accountable entity, with programmable controls, spend limits, and settlement logic baked in at the protocol layer rather than bolted on as an afterthought.

Until that exists at scale, builders like Marcus will keep doing the same thing: designing workflows that hand off to humans at the exact moment they should be closing.

The agent can do everything. It just can’t finish.

— -

BananaCrystal is building payment infrastructure for AI agents — wallets, liquidity rails, and settlement logic designed for autonomous systems. If you’re hitting this wall in production, the research brief is at agentpayments.substack.com (https://agentpayments.substack.com).