Ramp's Karim Atiyeh on AP automation agents, the Jolt AI acquisition, and why AI adoption is harder than it looks

Summary

Ramp co-founder Karim Atiyeh makes the case that accounts payable is the natural next frontier for AI agents in finance — and that most companies have no idea how badly they're managing it.

Ramp's earlier agent launch focused on expense policy enforcement, where the logic was clean: a written policy acts as a ready-made instruction set. AP is messier. Most companies don't have a bill-payment policy — they have a person, or a vague instruction to escalate anything over a thousand dollars to the founder. Atiyeh argues that gap is where fraud lives, where fat-finger errors happen, and where Ramp's new AP agents are designed to operate.

What the AP agents actually do

The agents handle three tasks. First, they process incoming invoices and infer classification preferences from historical behavior — how the company has split similar bills before, how it handles taxes, which cost categories it prefers. Second, they run fraud detection, flagging doctored invoices, unfamiliar bank accounts, and other anomalous signals. Third, they navigate payment itself, which Atiyeh notes is less trivial than it sounds: vendors in 2025 still send invoices with no payment instructions, or expect buyers to find a link on a website or fill a PDF form.

The agent has access to email, Slack, and calendar to gather context, and can browse the web, fill forms, and make calls. Guardrails sit at the payment-authorization and payment-method level, sitting above whatever the agent is permitted to do — meaning a prompt injection attack or a fabricated approval chain can't override the underlying controls.

Atiyeh draws the autonomy analogy explicitly: the agents are performing well on the highway, and the goal is to extend that to city roads as trust and capability build. On the data advantage, he points to Tesla's parallel — years of customers paying bills through Ramp means the system knows not just which invoices get paid, but how they're coded, which ones get rejected, and how buyer-vendor relationships evolve over time. That behavioral history is what makes the inference layer work, and it's something a bank logging into a portal can't replicate.

Model agnosticism under pressure

Ramp remains model-agnostic and sits in what Atiyeh describes as the trillion-token-per-month club — though he implies the actual figure is higher. Maintaining that agnosticism is getting harder, not because models are converging, but because the release cadence is relentless. By the time a team optimizes a cheaper model for a specific workflow, a new frontier release resets the comparison. The bigger performance gains, Atiyeh argues, have come from the agent tooling ecosystem — web browsing, form-filling, email integration, contextual memory — rather than from raw model improvements. Moving from a one-shot LLM inference to an agent running in a loop with a full tool suite is a larger leap than the gap between GPT-4 and its successors.

Why enterprise AI adoption fails

On Fortune 500 AI pilots that quietly died, Atiyeh is direct: garbage in, garbage out. The failure mode isn't the technology — it's that companies bring agents context that is outdated, inaccurate, or self-contradictory, then blame the model when the output is wrong. The analogy he uses internally is asking a designer to "make it pop" and being surprised when the result doesn't land. The underlying problem for large incumbents is years of tech debt and fragmented context that makes it structurally difficult to give an agent anything reliable to work with.

He also flags a category of waste that predates AI entirely: companies paying six-figure annual contracts for software that does something close to a for-loop, like splitting bills proportionally across three legal entities. The AI conversation is surfacing how much of that legacy spend was never justified.

Ramp's own AI adoption was frictionless precisely because the company was already resource-constrained and looking for leverage. Atiyeh's read is that large companies adopt more slowly not just because of complexity, but because they don't feel the same pressure — neither the competitive threat nor the resource scarcity that forces a startup to extract more from less.

Jolt AI acquisition

Ramp acquired Jolt AI, an agentic coding assistant built to help engineers adopt AI-assisted development. The deal came together through shared investors and moved quickly after Atiyeh met the team. The rationale is two-layered: Jolt's tooling improves Ramp's own engineering productivity internally, but the bigger bet is that the UX and adoption playbook Jolt developed for software engineers is a template for what needs to happen in finance. Atiyeh's view is that the transformation already underway in software engineering — where agents handle meaningful portions of code authorship — is about to repeat itself across every other function, and Ramp wants to be the company that figures out what that looks like for finance teams.