Integration consolidates 4.5 petabytes of knowledge on a platform with up to 200 AI agents 

The consultancy took five months to turn three decades of scattered files into Int.AI, which responds through Teams and WhatsApp.

Pamela Sousa
Integration and Microsoft
 

Integration, a global strategy and management consultancy founded in São Paulo in 1995, has put Int.AI into operation, an artificial intelligence platform that gives access to 4.5 petabytes of knowledge accumulated by the firm and can activate up to 200 agents at once. The archive covers around 4,000 projects carried out in more than 85 countries. Developed in partnership with Microsoft, the platform took five months to build. By the third month, a working version was already in use by a restricted group of consultants. 

The project began with an access problem. Over 30 years, the consultancy accumulated studies, market research, methodologies and project data in formats ranging from Excel spreadsheets to videos, images and PDFs. According to Integration, searches that once took days can now be completed in up to 30 seconds. 

“We were always organized in how we stored things, but the difficulty was in access,” says Integration partner Carolina Flores. “Every time someone needed a piece of information, a person had to go into the database, work out where the data was and interpret what the requester needed. It took hours of work to extract a number, rather than produce an analysis.” 

An archaeological exercise 

The first stage consisted of consolidating files stored across distinct layers, from frequently consulted material to content the team rarely opened. Carolina Flores describes the phase as “almost a form of archaeology” and acknowledges that the quality of the records varied depending on when they were created. 

“Back when we still didn’t save files online, we created V1, V2, V3, V final, actual final version,” says the partner. The cleanup produced a useful side effect: the consultancy revised its own filing and taxonomy criteria so it would not have to repeat the curation process at the end of every delivery. 

Before defining the technology, Integration set three requirements: scalability, cost and data governance. The last of these was the deciding criterion. 

“Part of the information belongs to Integration, but a large part belongs to the clients we serve, and we cannot put that at risk,” says Carolina Flores. “Governance is the number one factor in any solution. It has to be born alongside the technology.” 

The firm assessed market options but considered the investment too high for an uncertain outcome, given that previous attempts to solve the problem had not worked. Microsoft already supplied the productivity applications used day to day, so Integration brought the requirement to its existing partner. The team chose Foundry, an agent development platform launched as Azure AI Foundry and now renamed Microsoft Foundry. 

The agent orchestrator 

In practice, the consultant asks a question in natural language and an orchestrator distributes the task among agents specialized in analysis, validation and producing outputs. “There is an agent that searches, one that extracts the information, one that responds to the user, one that compares and one that summarizes,” Carolina Flores explains. “It’s like speaking to a group of analysts: find this information in the project folder, tell me the number and compare it with the market.” 

Andrea Cerqueira, Sales Director at Microsoft Brazil, says the orchestration is invisible to the user. “There are up to 200 agents operating behind the scenes, and everything depends on the question the user asks,” she says. 

The architecture is multimodel. The Foundry catalog contains more than 11,000 models from suppliers such as OpenAI, Anthropic, Meta and Moonshot AI, and the routing determines which of them handles each request. “Depending on the question, the platform may use a frontier model or a much cheaper model that can deliver the same quality of response without a high cost,” says the director. As an example of how quickly the catalog incorporates new releases, she cites Kimi K2.7 Code, from the Chinese company Moonshot AI, available since the beginning of July. 

Consultants use Int.AI through the channels they already use: Teams, WhatsApp and email. The system accepts text, image and audio. 

Permissions inherited from the internal structure 

All Integration employees have access to Int.AI, but the content visible to them depends on their level in the organization: the tool replicates the confidentiality policies that already applied to the files. 

“As a partner, if I want to access a project I wasn’t involved in, I’m not allowed to do so,” says Carolina Flores. “I have to follow Integration’s confidentiality rules.” 

Andrea Cerqueira attributes the pace of the rollout to this recognition of context built into the platform. “In the same way that I would not let Carol open a Word document in the company without permission, I’m not going to let her access information she is not entitled to,” she says. The mechanism also reduces token consumption, the executive adds, because the system does not need to receive the permission rules with every query. 

Microsoft classifies companies that adopt AI at scale as frontier firms, a concept it introduced in the 2025 Work Trend Index. “Because they adopt it with governance, they achieve a return on investment at least three times higher,” says Andrea Cerqueira. 

Routine and plans 

Carolina Flores uses Int.AI to prepare proposals, review lessons from previous work and assemble teams when she needs to check each consultant’s delivery history. 

“I’m driving to see a client and need to know what projects we’ve already done there and who I’ve spoken to. I can sort that out with a WhatsApp voice note and arrive with the information already in hand,” she says. “Before, I’d have had to prepare weeks in advance.” She sums up the result as “another colleague, and a very efficient one.” 

The next step in the partnership is to run multiple agents in parallel, with the autonomy to carry out activities that still depend on a person today. The consultancy is already testing that approach. 

“We don’t believe AI is going to fully replace people,” says the partner. “We want the two to work together, combining human supervision with gains in productivity and scale.” 

Read the original Portuguese version here.