Where AI works, by function
The analysis of what makes AI work in production, function by function. We read across every documented deployment in each area and surface what recurs — the architecture that works, the outcomes that held up, and the failure that shows up first. Start with your function, or search the record.
Start here
The functions with the deepest documented evidence.
Every other function
Ordered by how much evidence each one carries.
Frequently asked questions
What is The V Ones?
The V Ones is an analysis of 3,128 production AI deployments across 37 business functions. Each function is read across its full record to show the architecture that recurs, the implementation decisions where outcomes split, the failure that appears first, and the patterns that held up in production.
Which business functions have the most production AI?
Back-office operations and customer support lead by a wide margin, followed by quality assurance, marketing operations and e-commerce operations. The thinnest records sit in regulated, high-stakes work such as regulatory reporting and prior authorization, where few teams have published a production deployment at all.
What goes wrong first when companies put AI into production?
Scale, more often than accuracy. Among deployments that disclose a failure, the most frequent report is that pilot behaviour did not survive production volume: throughput, latency and cost surfaced after go-live. Extraction errors come next, then integration, where the model works before the write-back does. Often what failed was the tool the team replaced, not the AI.
Where should a company start with AI automation?
With a high-volume, routine process where the steps are predictable and a single error is cheap: a lookup-and-update, a document classification, a routing decision. Start with one document type or one decision cycle, prove accuracy there, then widen. Where inputs are messy, fix the input before scaling the output.
Can AI automation run unattended?
Rarely, and not durably. Across the functions that address it, the pattern that lasts keeps a person at the checkpoint, clearing exceptions, approving low-confidence output or making the final call. Teams automate the routine majority and route the complex tail to people. Autonomy expands only after the accuracy record has earned it.
Should companies build or buy AI automation?
Mostly buy. Over two thirds of deployments with a known path are vendor-led, though vendors also publish more case studies. In many functions the record holds no successful self-build at all. Build where the process is genuinely your product, you hold data no vendor has, or an engineering team will own it long-term. Incident management, quality assurance and compliance monitoring lean toward build.
Skip the reading. Describe what you're automating and we'll return the best-fit documented cases on your stack and scale.
About this site
This site analyzes the public record of production AI deployments across 37 business functions. Each function guide synthesizes what the documented cases have in common — the recurring architectures, the reported outcomes, and the failure modes teams hit first — with every underlying case attributed to its named public source. Individual case pages preserve source-reported metrics as published; the analysis, patterns, and verdicts are our own.