In week three, the number went the wrong way. An AI employee reviewing purchase requests at a global biopharma business cost $14.76 per request, according to Zamp’s published account. It had started at $11.57. More experience had, briefly, produced more expense. Anyone expecting automation to arrive fully formed should spend a moment with that little bump.
- Zamp gives AI employees recurring jobs across business systems.
- Its operating procedures capture rules, exceptions, and human authority.
- One reported rollout got cheaper after the work was redesigned.
By week seven, the reported figure was $1.21. Weekly volume had grown from 205 requests in week one to 2,712. The interesting part is the sequence: the team reorganized the work, eliminated redundant reasoning, and then moved to a smaller model. The cheaper model became useful because the job around it had changed.
01 The job behind the chatbot
Zamp builds software it calls AI employees. The phrase has an obvious marketing appeal, but it also describes a particular ambition: assign an ongoing responsibility, connect the relevant systems, and have the software carry the case through to completion or a useful escalation. Procurement, accounting, financial analysis, recruiting, and customer operations all appear in its public materials.
The customer is an enterprise team with work spread across applications and people. An invoice arrives in one place. The purchase order lives elsewhere. Someone knows whether this supplier’s peculiar formatting is acceptable. Zamp wants to put that operating knowledge within reach of the agent. Its architecture describes a runtime that assembles context, controls tools, handles execution, and records what happened.
That places Zamp among enterprise agent platforms, alongside overlapping alternatives such as Ema, Kore.ai, and Decagon. Its pitch centers on sustained operational responsibility and company-specific procedures. The useful comparison is the actual job: which systems it touches, what evidence it produces, and how much supervision remains. A vendor category cannot answer those questions.
02 The rulebook that does the work
Zamp calls its executable rulebook an Agent Operating Procedure, or AOP. The process owner supplies the objective, authoritative information, permissions, approval limits, and escalation paths. Reviewed corrections can change that procedure. Shared company context, which Zamp calls Company Brain, gives subsequent work somewhere to begin.
There is a subtle management problem here. A supervisor approving one unusual invoice has not necessarily changed company policy. Zamp’s procedure design distinguishes explicit policy, reviewed precedent, and the model’s inference. Otherwise, a useful exception could quietly become a general permission. Learning needs boundaries as much as it needs memory.
“We just have to chime in when it asks us a question.”
Kristin Tumulak, Mindbody · Zamp customer testimonial
In the biopharma rollout, the early constraint was human review. The agent initially referred 64% of requests to people. The team studied disagreements with reviewers and broke rules and exceptions into more predictable steps. Zamp says manual review later fell to 15%, while explanations appeared inside SAP Ariba, where reviewers already worked. That made expanding volume easier to justify.
Reported processing cost per request
03 A familiar window matters
Another pharmaceutical deployment supplies a wonderfully ordinary detail. A tagging team checked whether website analytics fired correctly, using a browser extension called Omnibug. Zamp’s October case study describes an estate of roughly 1,000 web properties and thousands of hours of annual manual checking.
The team wanted to keep seeing Omnibug. Zamp built a split screen: the website occupied 60%, the familiar decoding tool 40%. Its agent read the tracking specification, mapped interactions, generated test sessions, checked those instructions, executed them, and assembled pass/fail reports. Trust had an interface. The old tool stayed in view while the surrounding labor changed.
04 Before employees, there was Naivety
Zamp’s history makes the operational obsession more legible. Founded in 2022, it worked on treasury software before widening into AI employees. Its founders are Amit Jain, previously at Sequoia and Uber, and Raghav Saraf. Peak XV lists Zamp as a portfolio company.
An engineering account names the first analytics implementation “Naivety.” Single-threaded Go computations and event-triggered recalculation struggled as data and customers increased. The rebuild used ClickHouse, change-data capture, and separate batch and real-time work. Even then, deduplication and aggregation required care. The lesson for an AI buyer is practical: accurate answers still depend on the machinery supplying the numbers.
05 Count the work that stays with you
The business model follows usage. Zamp’s pricing guide says it meters agent work in Agent Compute Units and scopes quotes around systems, volume, complexity, and controls. The biopharma processing figures describe one deployment; they are not a subscription offer or a complete accounting of implementation and reviewer time.
Zamp also supplies specialists who help teams get agents into production. In banking, an AWS customer story describes an unnamed institution reducing a two-month chargeback backlog to a one-day processing timeframe. These accounts make the proposition concrete, while remaining vendor and partner accounts of particular deployments.
The approach fits a job with accessible evidence, stable policy, a named owner, and a measurable finish. Zamp’s banking guidance itself cautions against open-ended autonomy and consequential actions without proper authority. A queue full of ambiguous cases may simply move work to reviewers. Restricted data or unreliable system access can prevent the agent from doing the job at all.
The part another company can copy is refreshingly unglamorous: choose one queue, define completion, measure exceptions, and examine disagreements before expanding. Watch cost per finished case alongside human effort. Zamp’s week-three bump is a useful reminder to keep watching after the demo ends.