Introducing Ariftly — Agents that act. Not chat.

Most AI products today give you a chat interface. You type, they reply, and you do the work yourself.

We're building something different.

Ariftly is a Vertical AI Agent Platform — a system of autonomous agents that understand their domain, connect to your existing tools, and take real action. Not chat. Action. And nothing happens without your explicit approval.

Why not another chatbot?

The problem with chat-based AI isn't the AI — it's the interface. When you have to describe your problem to an AI, evaluate its response, copy the output into another tool, and execute the action manually, you've built an elaborate autocomplete system. The AI saved you some typing. Your workload didn't change.

The alternative is agents that initiate work. An agent that knows your domain, monitors your signals, surfaces findings, drafts actions, and asks: "Should I send this?"

That's what Ariftly is.

What we're launching

We're starting with two agents:

AI Readiness Agent — EU AI Act compliance is here. Procurement teams are sending 40-page security and AI readiness questionnaires before signing contracts. Most companies fail them and never know why they lost the deal. The AI Readiness Agent ingests those questionnaires, grounds its answers in your knowledge base and codebase, and drafts complete responses in hours rather than weeks. It also performs full EU AI Act gap analyses and generates board-ready compliance reports.

Every answer the agent produces is grounded in evidence from your actual documentation, architecture, and code — not boilerplate. Procurement teams know the difference. A response that says "Our model inference pipeline logs all decisions to src/logging/audit.py" closes deals. A response that says "We have appropriate logging mechanisms in place" does not.

Sales Agent — Lead discovery and outreach without the research grunt work. Define your ICP, connect Gmail, and the Sales Agent finds contacts, enriches them with verified signals (recent funding, tech stack from GitHub, hiring patterns, recent job postings), and drafts personalized first emails. Every draft comes to you for approval before anything is sent.

The Sales Agent works best for technical founders and sales teams selling into engineering-led organizations. It detects real intent signals — the difference between a company that is theoretically a fit and a company that is actively evaluating AI tooling right now.

The platform beneath the agents

Every Ariftly agent runs on the same control plane. This shared infrastructure is what makes agents useful in practice rather than impressive in demos:

Human-in-the-loop approvals — every external action waits for your explicit review and approval. This is built into the protocol, not bolted on. The agent can draft, prepare, and stage everything — but nothing leaves your accounts without a human decision. This is how AI agents should work, and it is the design principle we will never compromise on.

BYOK by default — bring your own AI provider keys (Anthropic, OpenAI, or any provider we support). We charge for the platform; you pay your AI provider directly. Your data does not go through a shared inference tier. Your usage appears in your own provider dashboard.

Skill Builder — extend any agent's behavior by describing what you want in plain English. The Skill Builder meta-agent generates the configuration YAML, you approve it, and it's live. No engineering work, no deployment, no waiting.

Event-sourced — every state change in Ariftly is an immutable event. Full audit trail. Full replay capability. If you need to reconstruct exactly what happened during a task — what the agent read, what it decided, what it sent — the event log has it.

Remote Agent Protocol — our open, versioned HTTP spec. Any language, any team, any infrastructure. If you want to build a custom agent that runs on your own servers and integrates with the Ariftly control plane, RAP v1 is the contract. We use the same protocol internally for our own agents.

Why vertical agents beat general-purpose ones

General-purpose AI assistants are genuinely useful. But they face a structural problem: they don't know your business, your customers, your integrations, or your domain-specific requirements.

When you ask a general-purpose AI to help with an EU AI Act questionnaire, it gives you generic answers that will not satisfy a procurement team doing diligence. When you ask it to research sales leads, it gives you a list of company names with no verified contact information, no intent signals, and no personalization.

Vertical agents are different because they combine:

• Domain expertise baked into the agent's design (not prompted in at runtime)
• Live data access via real integrations — your GitHub repos, your Gmail, your knowledge base
• Structured outputs with schemas your systems can consume
• Approval workflows that match how actual business processes work

The AI Readiness Agent has EU AI Act, NIST AI RMF, and common procurement questionnaire patterns encoded as part of its knowledge. The Sales Agent understands B2B sales signals, ICP matching, and outreach personalization as core capabilities. Neither is a prompted version of a general model — they are purpose-built for their domain.

What's next

More agents. Finance Agent (invoice reconciliation, spend anomaly detection, cash flow forecasting), DevOps Agent (incident triage, deployment risk scoring, post-incident report generation), Founder Agent (competitive intelligence, board pack generation, investor update drafting).

Every agent we add runs on the same control plane, shares the same approval system, uses the same Skill Builder, and can trigger and react to every other agent on the platform. A Sales Agent that detects an inbound reply mentioning compliance requirements can automatically trigger the AI Readiness Agent to pull the relevant evidence — before the meeting happens.

That cross-agent intelligence is the real product. Individual agents are useful. A system of agents that coordinate is transformative.

Early Access is open. → app.ariftly.io