This report puts the numbers from Zamp's published deployments in one place. Every figure links to the case study it came from, with the deployment named, so you can check the context before you reuse a number. For third-party market statistics on AI employees, see State of AI Employees 2026. This page covers only what Zamp's own deployments report.
What do Zamp AI employee benchmarks show?
The deployments differ, so the tables are grouped by metric family rather than ranked. Procurement shows how an agent's decision rate and cost change over weeks. Freight and billing show scale and accuracy on work that crosses many systems. Tag validation and the pharma rollout show coverage and breadth.
How does decision rate change over time?
The clearest time series comes from an AI employee that reviews purchase requests in the procurement division of a global biopharma enterprise. It went live in June 2026, and the biopharma procurement case study reports its first week and its state after the team restructured the agent's rules.
Metric | Week one | After restructuring |
|---|---|---|
Requests handled in the week | 205 | 2,712 (week seven) |
Decision rate (agent recommended a decision) | 36% | 85% |
Sent to manual review | 64% | 15% |
In week one the agent recommended a decision on 36% of requests and sent the other 64% to manual review. The team then studied every case where the agent and the human reviewer disagreed and broke the procurement rules and their exceptions into smaller, more predictable steps. After that change the decision rate rose to 85% and manual review fell to 15%.
Read the two rows as a pair. The manual review figure fell from 64% to 15%, and the decision rate went from 36% to 85%. The 64% is the week-one manual review share. It is not the starting point of the decision rate.
How do cost per request and volume scale together?
The same deployment reports cost and volume for its first seven weeks. These figures are from the biopharma procurement case study.
Week | Purchase requests processed | Cost per request |
|---|---|---|
1 | 205 | $11.57 |
2 | 151 | $11.59 |
3 | 234 | $14.76 |
4 | 1,327 | $7.67 |
5 | 2,389 | $4.62 |
6 | 2,682 | $2.79 |
7 | 2,712 | $1.21 |
Cost per request fell from $11.57 to $1.21, a 90% reduction, while weekly volume grew more than 13x. Cost did not fall at once. It rose in week three, and the large drops came in weeks four to seven, after volume grew and the workflow was restructured.
The case study explains the order of events. First, rules and exceptions were split into small, focused steps. Second, anything predictable, such as checks already run or facts already known, was made deterministic so the model no longer re-derived it on every request. Only then could the workflow move from a frontier model to a smaller, cheaper one without losing accuracy. Switching models first would have cost accuracy instead of saving money.
By this stage the agent also wrote its reasoning for each decision directly inside SAP Ariba, so reviewers could read why a request was approved or flagged.
What do the Zamp freight and billing deployments report?
Two deployments report accuracy, turnaround and hours together. Their processes are different, so the columns are comparable only as reported, not as a ranking.
Deployment | Scale | Accuracy | Turnaround | Hours reclaimed a year |
|---|---|---|---|---|
Freight order management at a major US freight logistics platform | 50,000+ cases a month across 50+ vendor portals with no APIs | More than 95% browser-agent accuracy | Settlement turnaround 40% faster | 20,000+ |
Multi-entity billing at a global IT services and outsourcing provider | Thousands of statements of work across five billing types | More than 70% browser-agent accuracy on the portal-dependent parts of the workflow; 100% compliance with SOPs and SLAs | Billing turnaround 50% faster | 150,000+ |
In the freight deployment the agent logs into each vendor's portal through browser automation, follows that vendor's SOP, records every portal interaction as a session video and escalates mismatches with the recording attached. It unifies more than 50 vendor systems without a single API integration.
In the billing deployment the work is mostly coordination. The agent identifies which contracts are due, requests data from delivery contacts, follows up until it arrives, applies contract-specific rate cards and routes invoice packages for approval with an audit trail attached. The accuracy figure applies only to the portal-dependent steps, which is why it sits next to a separate compliance figure that covers the whole cycle.
What did Zamp tag validation cover before and after?
A global pharmaceutical company with roughly 1,000 web properties used to validate website analytics tags by hand. The tag validation case study reports the manual baseline.
Measure | Manual process before Zamp |
|---|---|
Analyst effort | More than 8,000 hours a year |
Coverage | 20% to 30% of tags validated |
Effort for one site's core journeys | 2 to 3 analysts for at least three days |
The agent replaced that process, and the case study reports 8,000+ hours a year reclaimed. It breaks the job into six deterministic sub-tasks and returns a pass or fail report for each tag. The case study does not publish a new coverage percentage, so we do not quote one here.
What does the Zamp pharma rollout say about scale?
A top global pharmaceutical company started with one process, batch record review in manufacturing operations. Teams in clinical, regulatory, commercial and IT operations then asked for access on their own. The enterprise-wide pharma rollout case study reports hundreds of digital workers deployed or in progress across five functions, an estimated impact in the eight figures, and full auditability with 100% compliance to regulated-process requirements as the starting constraints. In IT operations, vendor onboarding cycle time went from weeks to hours.
This is a breadth benchmark, not an accuracy one. It shows how a deployment grew from one process to many, and the case study attributes that growth to demand from business units rather than a top-down mandate. Our case studies hub lists all of these deployments side by side.
How do we measure these numbers?
The figures above come from different deployments, so the definitions matter.
- Decision rate. The share of requests where the agent recommends a decision instead of sending the case to a person. In the procurement deployment the agent gives a human reviewer its findings and a recommendation.
- Manual review. The share of requests the agent routes to a person because it is not confident. It is the complement of the decision rate in the procurement figures: 64% plus 36% in week one, 15% plus 85% later.
- Straight-through processing. Straight-through processing (STP) means a transaction completes with no human touching it. Decision rate is a different measure. A recommended decision can still reach a reviewer, so read decision rate as the share the agent is willing to decide, not the share that closes untouched.
- Needs Attention. The Needs Attention status is what an AI employee assigns to a case it cannot resolve within its defined scope, such as a tolerance breach, missing data or an unfamiliar pattern. It routes the case to a person with the relevant context already assembled.
Two mechanisms sit behind those counts. Each case runs through a Plan-Execute-Validate loop, and the validate step checks the output against what was expected before the case is marked done, as described in Inside Zamp's self-learning system. Separately, confidence thresholds and guardrails tell the agent when to stop and escalate instead of guessing, as set out in Zamp's trust and security overview.
What is counted: weekly requests processed and cost per request in procurement, cases per month in freight, and hours reclaimed per year in freight, billing and tag validation. Browser-agent accuracy is quoted as each case study publishes it. We have not recalculated any figure.
Why is a high manual-review rate at launch a good sign?
In week one the procurement agent sent 64% of requests to manual review. That was the safety mechanism working. When the agent was not confident about a request, it sent the case to a person instead of approving it, which made it acceptable to run against live purchase requests from day one rather than a sandboxed pilot.
The rate then fell without loosening any policy. Reviewer corrections went into the agent's Agent Operating Procedure, the plain-language document that holds a role's rules, tolerances and known exceptions. Each change goes through a branch, evals against a golden set of behavior the agent already has to get right, review and merge. Only the procedure changes, never model weights, as explained in How Zamp's AI employees work.
The transferable lesson is about what to watch in the first weeks. A low decision rate with a clear escalation path is a sign the guardrails work. A high decision rate with no reviewer visibility on week one would be harder to trust. Ask how a platform behaves when it is unsure, and where the reasoning is written so reviewers can check it.
What do these numbers not tell you?
These benchmarks have limits, and you should read them with those limits in mind.
- A small number of deployments.
They are the published case studies, not a statistical sample. There is no average or range across them, and we have not calculated one. - Different processes.
Purchase request review, portal-based freight settlement, billing coordination and tag validation have different inputs and different definitions of a correct outcome. A 95% figure in one and an 85% figure in another do not rank the deployments. - Different metrics.
Decision rate, browser-agent accuracy and compliance rate measure different things. Match like with like. - Reported results, not guarantees.
Your results depend on your process, your documentation, your systems and your reviewers. Deployment model matters too, and our guide to on-prem, BYOC and SaaS deployment explains the options.
Use the figures to frame questions for your own pilot: what manual review rate you will tolerate in week one, how corrections reach the procedure, and how cost and volume are tracked weekly.
