
Hey everyone!
We’re AJ and Miguel from Grep AI. We’re building self-improving agents for tedious, repetitive work that’s slowing down your business. IG.com, a UK-based online brokerage, used our agents to automate 80% of onboarding, unlocking a record year of growth that led to their graduation to the FTSE 100.
Launch video: https://youtu.be/s9NEm6rRcCA
The problem: repetitive work that requires judgment and expertise is still too hard to automate and scale.
At Brex, we had 2,000 businesses waiting to onboard because we needed to manually review each one to verify they had a physical address. Fast-growing companies are full of tedious, repetitive tasks like this that become a bottleneck to growth, but require research, judgment, and some expertise.
But getting an agent to accurately and efficiently complete a task that requires automation, investigation, and some judgment thousands of times is still a challenge.
Hard-coded LLM workflows like LangGraph and n8n are cheaper to scale, but can’t adapt to edge cases. E.g., the business has a physical address on Google Maps, but it’s actually a mailbox in a co-working facility.
General-purpose agents like Codex, Cowork, and Copilot are much better at handling exceptions, but when performing repetitive tasks, they waste expensive tokens reasoning through the workflow as if they were a new employee. Enterprises often offshore these types of tasks at $6-$12/hr, so if the agent costs $4/tasks they are not going to use it.
Grep AI: Self-improving agents that get smarter and cost less with every run.
Grep AI is a managed agent platform that lets you create, optimize, and deploy self-improving agents to automate repetitive work in minutes without writing any code. Our agents learn from every run, getting smarter and costing less as you scale them.
How Grep AI works:
Create an agent in minutes with just a prompt and any useful docs. Grep uses deep research, a library of more than 300 built-in skills, and over 100 data integrations to make each agent an expert out of the box.
Frontier or open-source models: Unlike OpenAI’s Agent API or Claude Managed Agents, you aren’t locked into one model provider.
Self-improving harness: Our harness learns as it repeats a task, moving repeatable steps from frontier LLM calls into workflows and code, so it gets smarter and cheaper with every run.
Secure sandbox: Agents run in a secure sandbox with a file system, audit logs, access controls, and cost caps included.
Run from anywhere: Use agents manually in our app, or via schedules and monitors, batches, API, MCP, and TypeScript and Python SDKs.
Stay in control: Manage your agent workforce with a central dashboard to monitor cost, accuracy, and auditability.
How companies use Grep AI
Companies like Airwallex, Citizens Bank, Worldpay, Global Payments, and Shopmonkey use Grep for business due diligence, underwriting, document verification, HR operations, GTM ops, competitive intelligence, and social and market research.
In six months we’ve saved these companies over 150,000 hours of work
Why us?
AJ led AI/ML products at Coinbase and Brex. His first startup was acquired by Google, where he worked on Google Photos and Android.
Miguel led Platform Engineering at Brex across ML, data, financial infrastructure, and compliance operations tooling. Before that, he led data science at Twitter.
We met at Brex in 2020, where we scaled the customer base 10x in a year with automation across risk, compliance, customer support, and operations, before AI agents existed. That experience inspired us to start working on AI agents for compliance teams in 2023. Then we took everything we learned from building agents and turned it into a managed agent platform to scale any repetitive work.
Do you have any repetitive operations/back-office work that’s slowing down your company?
We met at Brex in 2020, where we worked on the platform team and saw how manual operational work could become a bottleneck to growth. Even straightforward onboarding tasks required people to research information, handle exceptions, and make decisions.
We started Parcha in 2023 to build AI agents for compliance workflows. That gave us firsthand experience with repetitive, high-stakes work where accuracy and control matter. We built Grep AI to make those capabilities available to teams automating repetitive knowledge work across their business.
We started in 2023 as Parcha, building AI agents for compliance teams. Working on those workflows taught us how to handle complex research, exceptions, and judgment while keeping people in control.
We also saw the same need beyond compliance. In 2026, we expanded into Grep AI, a managed agent platform for repetitive, high-stakes knowledge work across operations, research, and other business functions. We are now building self-improving agents that learn from repeated work, with the goal of improving accuracy and reducing cost as they scale.
Businesses are full of repetitive work that still requires research, judgment, and context. Reviewing documents, investigating a business, or researching a prospect can look similar from one case to the next, but exceptions make rigid automation difficult.
The people who understand these jobs often have to do them manually or wait for engineering help. Grep AI lets them describe the work and create managed agents to repeat it, with tools, shared context, and controls. Our focus is work that needs to be accurate and economical when it runs thousands of times.
At IG Group, our agents review screening alerts during customer onboarding, clearing false positives and escalating cases that need manual review. The IG case study reports that activation time for flagged customers fell from 2–3 days to under 10 minutes.
We want to rid businesses of tedious, repetitive work so their employees can focus on the highest-leverage tasks. The people who know a workflow best should be able to turn it into a managed agent without having to build and maintain the infrastructure themselves.
As those agents repeat the work, they should become more useful and less expensive to run. Success means teams can grow without manual operational work becoming the bottleneck.