How Block Got 12,000 Employees Using Ai Agents In Two Months

RALEIGH, N.C. — While astir companies are still figuring retired really to get their developers to usage AI coding devices effectively, fintech institution Block conscionable deployed AI agents to its full 12,000-person workforce successful 8 weeks.

At nan All Things Open 2025 convention present past week, Angie Jones, Block’s vice president of engineering, explained really nan institution did it.

It Started Small

Like truthful galore tech occurrence stories, this 1 began pinch a azygous disappointment engineer. “Bradley,” a main machine learning (ML) technologist astatine Block, said he was tired of AI devices that could only make codification snippets, Jones said successful her presentation. He wanted thing that could really do things — automate analyzable improvement tasks, not conscionable propose nan adjacent statement of code, she said.

When OpenAI introduced function calling, Bradley and a mini squad started experimenting pinch automating parts of their improvement workflow. The experiments worked, but they deed a scaling problem fast. Without immoderate standards, building integrations for different APIs became a headache. Each caller instrumentality meant civilization codification and attraction issues.

Then Anthropic reached retired to Block astir thing called nan Model Context Protocol (MCP) — an unfastened modular for connecting AI agents to devices and information sources. Block loved nan idea, became a motorboat partner, and rewrote their soul supplier (“Goose“) to activity pinch MCP.

The Accidental Discovery

Yet, while building Goose for engineers, nan squad realized they could make it activity for everyone astatine nan company.

“We started thinking, what if we could make this activity for each 12,000 employees?” Jones said. It sounded ambitious. Most of nan workforce had ne'er seen a bid line, fto unsocial managed API keys.

The first attempts were misses. Nontechnical labor couldn’t fig retired really to instal nan software. Those who managed that portion sewage stuck connected authentication flows. The full acquisition was built for developers, Jones stated.

Making It Work for Real People

However, nan squad stopped reasoning for illustration engineers and started reasoning for illustration their users.

First, they made nan package easy to install. The full process of downloading a file, unzipping and dragging it to nan Applications files was unfamiliar to galore employees. So, they added Goose to nan soul package center. Now it autoinstalls and autoupdates connected each institution laptop.

“So, if you person a Block laptop, you person Goose. It conscionable works,” Jones said.

Next, they tackled nan “which exemplary should I use” problem. Instead of forcing everyone to usage 1 AI model, they fto labor choose. Marketing folks thin to for illustration ChatGPT for writing. Engineers instrumentality pinch Claude for method tasks. The elasticity made take overmuch easier crossed different teams.

“We let labor to take from OpenAI, from Anthropic, from Google, from Meta providers,” Jones said. “And this gives group that powerfulness to take which makes take overmuch easier crossed nan various disciplines.”

Perhaps nan biggest situation was astir security. As Jones noted, “Block is simply a fintech company, meaning Square and Cash App, right? And so, we woody pinch a batch of regulations. We person to beryllium very secure.” Instead of letting group instal random tools, “We spun up a mini task squad to build soul servers that group needed.”

The squad sewage busy.

“By nan extremity of that week, we had complete 60 soul MCP servers fresh to go, and today, we’re complete 100.”

These soul MCP servers see connections to devices for illustration Slack, Google Calendar, their information storage and soul systems. All preapproved, each automatically available.

The Problems Nobody Talks About

Yet, moreover pinch amended software, caller problems emerged. Employees started enabling each disposable instrumentality “just successful lawsuit they needed it,” Jones said. With dozens of devices active, nan AI supplier became slow and sometimes unusable.

Block’s solution was to instrumentality automatic instrumentality management. If a personification asks astir almanac scheduling and Asana tasks, nan strategy automatically enables Google Calendar and Asana while turning everything other off. Users don’t person to deliberation astir it, Jones said.

They besides built speech summarization. When chats get excessively long, nan strategy automatically summarizes what’s happened and continues pinch a caller context.

Building a Support System

Moreover, Block invested heavy successful helping group usage these tools.

They created 2 Slack channels: 1 for getting help, different for sharing things group built. The inspiration transmission became 1 of their champion take drivers. Employees would spot a workfellow automate thing they struggled pinch and instantly want to effort it themselves.

“One [Slack channel] is simply a spot wherever group tin inquire questions, they tin report bugs, they tin petition features aliases conscionable get unstuck successful existent clip and everyone conscionable benignant of pitches successful here,” Jones said. “I can’t overemphasize really captious this has been to adoption, particularly for our nontech users.”

The 2nd transmission became Block’s concealed weapon.

“It’s a abstraction wherever labor tin driblet their occurrence stories aliases cool workflows aliases clever prompts that they travel up with, aliases chaotic things that nan supplier was capable to propulsion off,” Jones said. “So, this transmission quickly became 1 of nan champion integrated drivers of adoption. People scroll through, they spot what others are doing, they instantly want to effort it retired themselves.”

Block besides runs play workshops, agency hours and training sessions astir Goose. Not conscionable “here’s really to usage AI” sessions, but targeted workshops for circumstantial teams. Sales operations sewage different training than customer support.

What People Actually Built

The results were surprising, Jones noted. Employees crossed nan institution started automating tasks that would person been unthinkable before, she detailed:

• A information expert now detects fraud by asking their information storage questions successful plain English alternatively of penning analyzable SQL queries.
• Someone successful income operations analyzed 80,000 income records and figured retired really to redistribute them crossed different programs. What utilized to return days of spreadsheet activity took an hour.
• An technologist utilized their incident guidance integration to spot patterns successful strategy outages, starring to fixes that made their infrastructure much reliable.

Jones shared her ain experience: When she needed to sponsor an arena but didn’t person clip for their accustomed procurement process, she asked Goose for help. “It felt for illustration Goose was whispering successful my ear, ‘Let maine show you who to talk to. Want maine to conscionable grip this for you?'”

Jones said nan astir telling communicative is 1 regarding an worker who built her ain MCP server, past asked wherever to taxable it. When they gave her a GitHub link, she said, “Oh, I don’t person a GitHub account. I’m not a developer.” A nonengineer had built a instrumentality that solved her circumstantial problem, Jones explained.

The Bigger Picture

Block’s acquisition reveals thing important astir endeavor AI adoption: nan exertion exists. What’s missing is everything astir it — installation, authentication, security, training and community, Jones indicated.

“This isn’t a broadside task anymore,” she concluded. “We’re putting AI into believe astatine standard and figuring retired what this looks for illustration for a existent company.”

Other companies trying to standard AI tin study from Block’s approach. Start pinch standards (MCP helped them debar integration chaos). Make things easy to use, not conscionable technically possible. Build information and compliance successful from nan beginning. And put successful acquisition and organization support, Jones explained.

Most importantly: Let users thrust innovation. Some of nan astir imaginative applications came from departments that had thing to do pinch technology, she stated.

Block proved that endeavor AI take does not person to beryllium slow, constricted to developers aliases focused connected replacing workers. When you make AI devices genuinely accessible, labor find ways to activity smarter, faster and much autonomously.

The early of activity mightiness not beryllium astir AI replacing humans. It whitethorn beryllium much astir giving each quality their ain AI-powered capabilities. Block has shown what that looks for illustration astatine scale.

Group Created pinch Sketch.