AI Inside the Business: Dependencies, Documents, Hiring and Support
Agent launches aimed at developers get the headlines, but a steady share of Show HN projects are aimed at the boring middle of a company: the dependency queue nobody owns, the invoices someone retypes, the inbox of applicants nobody has time to read. These tend to be less ambitious and more immediately useful than the platforms, because each one replaces a task somebody is doing by hand today.
Here are eight from the latest batch.
The dependency queue
Almost every repository has a Dependabot backlog that silently grew into a monument. PatchWave is a GitHub app that analyzes, fixes and merges those pull requests: its agent checks each dependency bump against your actual code and repairs what the bump breaks, so the queue drains without a human reading a diff about a patch version. It installs as an app rather than a pipeline rewrite, and there is a free monthly allowance of merged PRs.
Where development happens
On the building side, JetBrains Air is one system for building software with agents. It works with the agents teams already use, including Claude Agent, Codex, Junie, Copilot and OpenCode, plus anything you can connect over ACP, and exposes tasks, automations and projects you can drive from JetBrains IDEs, the web, the CLI or a phone. The bet is that the agent is replaceable and the surrounding task system is not.
For teams whose constraint is confidentiality rather than tooling, Privatemode runs AI with data encrypted before it leaves the device and still protected while the model processes it, including inside Excel, Word and PowerPoint. It is the kind of product that exists because contractual assurances were not enough for the documents in question.
Documents in, structured data out
Document extraction is one of the few AI jobs with an obvious return on investment, and the interesting thing here is the pricing model. Invofox turns invoices, payslips, contracts and other documents into structured JSON through one API and bills only for pages it extracts correctly, with the extraction continuing to learn from corrections. Paying for accuracy instead of volume is a much easier purchase to defend than paying per page regardless of whether it parsed.
Hiring and support
Recruiting is where AI scoring has the most upside and the most risk. Worklicant is a tech hiring platform where AI reads every resume and ranks applicants: you share one apply link, it scores each person against the job, hides spam and fake resumes, and auto-rejects those who fail a deal-breaker you set, with candidates booking their own interviews. It charges per vacancy rather than per seat, which suits teams that hire in bursts.
On the other side of the customer relationship, Tori is an AI support agent from the TalkJS team. It learns your product from the help docs, product pages and company knowledge you already published, answers through a UI you can customize, and hands off to a human with the full conversation when that is what the customer needs. Building on existing messaging infrastructure means delivery, threading and history are solved problems rather than features to rebuild.
Getting answers out of data
Datastory is a no-code platform for data storytellers: upload your own file or search a curated catalog of open datasets, then build interactive charts, websites and reports with its Studio, CMS and AI features. It sits in the gap between a spreadsheet and a dashboard product, for people whose job is to explain a number rather than compute one.
And into the places people talk
Most agent deployments still end up in a sidebar nobody opens. Channels SDK is the open-source SDK for putting any agent into Slack, Microsoft Teams, Discord or Telegram with native interactive UI. Your agent keeps its own model, tools and business logic; Channels connects an AG-UI-compatible agent to the platforms where the work is already discussed. It is a small piece of glue that decides whether an internal agent gets used or ignored.
What to take from this
None of these eight tries to be a platform. Each one picks a queue, a document type or an inbox that a real team is drowning in, and removes the human step. That is a quieter kind of AI product than the ones that make the front page, and it is the kind that tends to survive a budget review.
It is also worth noticing how the pricing follows the risk. Billing only for correctly extracted pages, charging per vacancy instead of per seat, and metering agent work in credits are all attempts to let a buyer start small without a procurement cycle. Where an AI product cannot be tried on one queue or one job posting, it usually cannot be tried at all, and these launches seem to understand that better than the platforms do. The other shared trait is that each one leaves a trail: a pull request with its analysis, a JSON payload you can validate, a ranked shortlist you can override. Auditability, not intelligence, is what gets these tools past the second meeting.
