AI workforce: agents that take over support, QA, engineering, and phone jobs.
An AI workforce, as Aldenebai builds it, is a set of autonomous, guardrailed AI agents that each take over one whole job inside the tools you already run: support tickets, QA cycles, engineering tickets, phone calls, bug backlogs, and documentation. Six such agents, built by Paul Rahme, run in production at a US marina-software company, 2025–2026. Each is tested and auditable, keeps a human on the exceptions, and none is a chatbot.
Open any card to see how it is built and watch a scale model run.
The six systems
AI support agent
Triages, answers, and closes customer tickets end to end, 24/7, escalating to a person only when confidence is low. About six in ten tickets resolved with no human touch.
AI test automation
Reads the product, writes the coverage, runs it on a schedule, heals its own failures, and reports. Full regression before every release with nobody running a test.
AI dev agent
For feature and change tickets: a Linear ticket is triaged, scoped, coded with its tests, pushed through the build gate, and opened as a review-ready pull request. Engineers approve instead of build.
Voice AI agent
Answers the support line, verifies the caller, answers from your documentation, logs the ticket, files the bug, and hands off to a person the moment judgment is needed.
Autonomous bug fixer
For production exceptions and bug reports: the error is reproduced, root-caused, fixed with tests, rebuilt, and opened as a PR for review. Red builds retry themselves before a human is asked.
Documentation automation
Audits your documentation against the product and the support log, finds the genuine gaps, drafts what is missing, and files it for review. The knowledge your agents answer from stays current.
Autonomy is earned with limits
Every system here runs on the same rules. The machine acts on its own where the risk is low and the volume is high. It stops and asks where a decision is irreversible. Every action is logged with the reason behind it, and a human can see, in one place, what it did and why.
That discipline comes from a QA background. The person who designs these systems spent years finding out how software actually breaks, so each one ships with its own tests, its own guardrails, and a kill switch.
Which AI system fits which job?
Start from the work you want gone, not from the technology. Each row is one job, the system that takes it over, and where that system runs today.
| THE JOB YOU WANT GONE | SYSTEM | WHAT IT REPLACES | STATUS |
|---|---|---|---|
| Tier-1 support tickets, the how-to and known-issue repeats | AI support agent | The first-line queue: about 60% of tickets closed with no human touch | In production, the marina-software client |
| Release regression and test maintenance | AI test automation | Manual regression cycles and broken-test triage: about 80% less manual QA | In production, the marina-software client |
| Clear engineering tickets that need setup and boilerplate | AI dev agent | The build step: engineers review a pull request instead of writing it, about 2× faster | In production, the marina-software client |
| The bug backlog and production exceptions | Autonomous bug fixer | Reproduce, root-cause, fix with tests, open the PR | In production, the marina-software client |
| The documentation backlog | Documentation automation | Gap audits against the product and support log, drafts filed for review | In production, the marina-software client |
| The support phone line | Voice AI agent | First response on the phone: verify, answer from the docs, log the ticket, transfer | In production, the marina-software client |
| Something not on this list | The Automation Audit | Mapped, scored, and either scoped or marked "not yet" | Free, 30 minutes |
How an AI workforce differs from AI assistants
An AI assistant helps a person do a job faster: it drafts the reply, suggests the code, summarises the call. The person still owns the job, so the queue is still theirs, and the hours saved are hard to see on a chart. An AI workforce takes the job itself. The system owns the queue, works it around the clock inside a fixed allowed-action list, and hands a person only the exceptions, with its notes attached.
The difference shows up in what you measure. Assistants are measured in minutes saved per task. A workforce is measured in tickets resolved with no human touch, releases shipped with no manual regression, and pull requests reviewed instead of written. Those are the numbers on this page, rounded from production data at a US marina-software company, 2025–2026.
What "in production" means here
A system is listed as in production when it works real traffic for a real company today, with a person on the exceptions. All six systems on this page meet that bar at a US marina-software company, on its marina-management product. Three carry published, measured results on their case studies: the support agent, the test automation pipeline, and the dev agent. The bug fixer and documentation automation run in production without published figures yet, and the voice agent publishes its call volume, around 100 calls a month, but not yet a resolution rate; each page says so rather than borrowing a number from elsewhere.
Every number carries its provenance and a measurement note on the relevant case study, and every system ships with the same limits: a confidence threshold, an allowed-action list, an audit trail, and a kill switch. Where the person belongs in each system is set out in the human-in-the-loop guide.
Where to start
Questions about the AI workforce
What is an AI workforce?
An AI workforce is a set of autonomous, guardrailed AI agents that each take over one whole job, such as support tickets, QA, engineering tickets, or phone calls, with a person handling only the exceptions. It differs from AI assistants, which help a person do a job faster but leave the job with the person.
Which system should we start with?
The one with the highest volume of repetitive, reversible work. For most companies that is the support queue or release regression, which is why those were the first two built. The free Automation Audit ranks the candidates for your company in 30 minutes and writes the ranking down.
Do these systems replace people?
They replace the repetitive part of a job, not the judgment. At the marina-software client the support team now handles the escalations and the customer relationships, the QA team decides what quality means and what blocks a release, and the engineers review pull requests instead of writing boilerplate. Headcount was not the goal; the queue coming off the org chart was.
What happens when a system is wrong?
It stops before the wrong action ships. Below its confidence threshold a system escalates to a person with its notes; irreversible actions such as merging code or issuing a refund are never on its allowed-action list; every decision is logged with its reason; and a kill switch stops any system instantly, after which work routes back to people as before.
Can the systems run on our tools and our cloud?
Yes. Every system runs in the client's own repositories and cloud accounts, with model calls through the client's provider account such as AWS Bedrock, so data is not used to train provider models and the client holds the keys. The tools named on each page are the ones the systems run on today; the adapters for other helpdesks, trackers, and CI tools change, not the design.
If it's repetitive, a system can do it.
Six systems here, and the seventh is whatever repeats inside your company. The free Automation Audit finds it, in 30 minutes, with a written Automation Map.