7 questions about AI agents — answered with drawings, not jargon.
You don't need to understand AI to use it safely. Read for 3 minutes. If it makes sense, one message does the rest.
Your message reaches the founder directly.
1. What does ContactLab do, in one sentence?
It runs tasks with AI agents in an isolated environment — and hands you the proof of what was done.
Invoice reconciliation. Batch release. Traceability and recall. Work that repeats — and needs proof.
And it does not stop at operations: the catalog brings hundreds of starting points — agents, skills and templates, from finance to marketing, from healthcare to engineering. If a code agent with a good model can do it, here it comes out with proof.
2. Isn't this just another AI chat?
No. In a chat, you operate: ask, read, copy, paste. Here, an agent assembled from the catalog runs the task in an isolated environment — and the work comes back with a receipt, from the first command to the final result.
Under the hood, they are code agents: the technology engineering teams worldwide run in production every day (Claude Code, OpenAI Codex, Qwen Code, Google Antigravity). The bet here is not an experimental agent — it is reusing the most production-run agent class there is, with governance on top.
Your team follows the outcome on a panel — no terminal, no programming.
The chat stays your thinking tool. ContactLab is your execution line.
3. How do I know what the agent did?
Every run leaves a complete record: the commands, the files created, the accesses made and the result. You read the report like an audit report.
And while the agent works — for many minutes — you follow the events and approve what matters. Nobody waits staring at a screen.
4. What if it errs? What if it does too much?
Three lines of control — all defined by you:
Consequential actions — sending, paying, publishing — only happen with recorded approval. The catalog is a starting point, not an endorsement: reviewing the content and every step is on you. Autonomy is earned with evidence, not granted.
5. What does it cost? What if it blows the budget?
Cost shows per run: tokens, time and the model used. You see the spend before it becomes an invoice. And the 30-day pilot only starts with success criteria agreed in writing.
What happens if it does not work is agreed in writing too: if the workflow does not run end to end with an evidence trail, the documented decision is to End — with no commitment to continue.
6. I already use ChatGPT and Claude. Why one more tool?
You replace nothing — and today's chat is powerful: it thinks, answers, drafts and, increasingly, acts.
ChatGPT itself now advertises itself as an agent that "thinks and acts". The company's question is not whether the agent acts — it is who answers for it.
ContactLab takes care of the work that cannot go without a receipt: what touches money, clients or company systems.
What today's chat already does
Far more than talking: ChatGPT and Claude browse, run code, use connectors and complete multi-step tasks — in the vendors' own environments, with you operating each task.
What the company needs around it
Execution in an isolated environment — with the tools on your own AWS account. Network and credentials with scope. Recorded approval on consequential actions. The same work repeatable without rewriting the prompt, with cost and outcome measured per run — and the work becoming a company asset, not one person's conversation.
And the right model changes with the task: the same task can cost many times more — or err more — in another combination of agent and model. ContactLab measures each combination (Claude Code, OpenAI Codex, Qwen Code, Google Antigravity). And you choose based on the evidence of your own work, not generic rankings.
And the AI use that is personal today becomes the company's: isolated execution, shared memory and an auditable history.
And you become the person who brought AI with control into the company — not the person who imported a risk.
What's inside
For the skeptic on the team — the spec in ten lines:
Isolated execution per task
Temporary environment, destroyed at the end. Network with allowed destinations — off-policy attempts get logged. Credentials with scope.
Complete evidence per run
Inputs, outputs, events, decisions, time, cost, agent and model. When audit asks, the answer exists.
Human review, progressive autonomy
Every run passes through a person. Autonomy grows with evidence — by design, not by promise.
Reusable work catalog
Hundreds of starting points, skills and templates — one-click environment and direct GitHub import.
MCP: public and on your AWS account
Public connectors (context7) and custom MCP via Bedrock AgentCore with SSO. Your tools, your account, your control.
Company memory
Hindsight at runtime: approved procedures and corrections become context. The second run beats the first.
Multi-agent, multi-model
Claude Code, Codex, Qwen and Antigravity — with Anthropic, OpenAI/Azure, Bedrock, Gemini/Vertex, Alibaba or local (vLLM/Ollama).
Governance and privacy
SSO and role permissions, private by default, encryption in transit and at rest, measured consumption on the panel.
Analytics and FinOps
Technical completion, approval latency (p50/p95), adoption under governance, cost per model and tokens. Even cost is visible.
Granular access
Visibility per resource — private, shared or organization — and revocable consents at any time.
Your tools, your AWS account
Your data and files are private by default — and can persist in AWS resources you control, with controlled cross-account access. MCP runs via AWS Bedrock AgentCore: an ephemeral token per run, authorized against your SSO.
7. How do I start?
A 30-minute conversation, with a real execution in front of you. If it makes sense, an entry diagnostic on your first workflow — and a 30-day pilot with success agreed before running.
The diagnostic — a read of your current workflow, with an execution plan — becomes credit toward the pilot. And the pilot includes architecture and security documentation ready for your IT team — the person who approves the pilot is often someone else.
Then the path is gradual:
A question left? It becomes an answer in one message.
Not us saying it. The market is.
“AI use by Brazilian companies almost doubled in one year: from 29% to 45%.”
FGV — survey of 3,000 companies“ChatGPT now thinks and acts — using tools to complete tasks.”
OpenAI — ChatGPT agent launch“If you sell the tool, you're in a race against the model.”
Sequoia Capital — Services: The New Software“Over 40% of agentic AI projects will be canceled by 2027 — inadequate risk controls among the causes.”
Gartner — 2025“Why AI is going vertical (again)”
“82% of Brazilian companies have AI in use without approval — and only 18% use it in an integrated way.”
Exame — corporate AI use in BrazilWhen models commoditize, value moves to the applications.
Mistral — Arthur Mensch, CEOAgents and cheaper reasoning expand what software can do.
Sarah Guo — Conviction, AI Engineer 2025AI demands a new operating model: from artisan experimentation to production at scale.
Marco Iansiti — Harvard Business SchoolAgents in production need session isolation, identity, memory and observability.
AWS — Bedrock AgentCore“Only 7% of agent turns need a frontier model.”
LangChain — Switchyard benchmarkOpportunities in AI: the map of value — from tasks to systems that do the work.
Andrew Ng — Stanford, 2023“AI use by Brazilian companies almost doubled in one year: from 29% to 45%.”
FGV — survey of 3,000 companies“ChatGPT now thinks and acts — using tools to complete tasks.”
OpenAI — ChatGPT agent launch“If you sell the tool, you're in a race against the model.”
Sequoia Capital — Services: The New Software“Over 40% of agentic AI projects will be canceled by 2027 — inadequate risk controls among the causes.”
Gartner — 2025“Why AI is going vertical (again)”
“82% of Brazilian companies have AI in use without approval — and only 18% use it in an integrated way.”
Exame — corporate AI use in BrazilWhen models commoditize, value moves to the applications.
Mistral — Arthur Mensch, CEOAgents and cheaper reasoning expand what software can do.
Sarah Guo — Conviction, AI Engineer 2025AI demands a new operating model: from artisan experimentation to production at scale.
Marco Iansiti — Harvard Business SchoolAgents in production need session isolation, identity, memory and observability.
AWS — Bedrock AgentCore“Only 7% of agent turns need a frontier model.”
LangChain — Switchyard benchmarkOpportunities in AI: the map of value — from tasks to systems that do the work.
Andrew Ng — Stanford, 2023