45 minutes to a live integration

How Authenti.ca builds enterprise AI agents on Paragon’s SDK

By Garrett Scott, Head of Customer Marketing

The morning we spoke with Michael Borg, he had already shipped an integration before the call started.

Borg is the CEO and co-founder of Authenti.ca. A few hours before we got on the phone, he had stood up a NetSuite connection for one of the company’s AI agents. Part of it came off the shelf. Part of it was custom to that specific customer. Start to finish it took him about forty-five minutes, and by the time we were talking it was already running in production.

For most companies, a partly custom ERP integration is a project with a ticket, an owner, and a week on the calendar. At Authenti.ca it happened between meetings. Why it happened that fast says a lot about how an AI-native company actually picks the tools it builds on.

01 Why an AI company buys an integration platform for the SDK

Authenti.ca calls itself “Palantir for the rest of us.” The company sells AI Forward Deployed Engineers, or AI FDEs. The Forward Deployed Engineer is the role Palantir made famous: a person who embeds inside a customer’s operation, learns how the work really runs, and builds automation against it. Authenti.ca turns that role into software. An AI FDE is not a chatbot bolted onto a help desk or a copilot waiting to be prompted. It is an agent that learns a customer’s supply chain from their own documents and conversations and then acts on it, auditing trade invoices, planning sales and order flow, and carrying operational load that used to take headcount. The wedge is affordability: a company that could never sign a ten-million-dollar Palantir contract can get that caliber of embedded engineering delivered as software.

That product has one hard requirement baked into it. An AI FDE is only useful if it can read and write across the systems the customer already runs, their ERP, their CRM, their inventory, their purchase and sales orders, their email. Integrations are not a feature Authenti.ca planned to add later. They are the floor the whole product stands on. Without them, the agents have nothing to act on.

So when Borg went looking for an integration platform, he was not shopping for the thing most vendors put on the homepage. He looked at Prismatic, weighed building the layer himself, and picked Paragon. What decided it was the part of the product that usually gets the least attention in a demo.

“The thing for me that initially sold me on Paragon was the SDK. I love that you guys have that no-code UI. I don’t really use it.”

Michael Borg, CEO and co-founder, Authenti.ca

This inverts how the category normally sells itself. Integration platforms are marketed on the visual builder, the promise that a business user can wire up a connector without an engineer. For an AI-native team, that promise solves a problem they do not have. Nobody at Authenti.ca is clicking through a builder, because nobody there is writing integration code by hand in the first place.

“We don’t write code, we just review it these days.”

Michael Borg, CEO and co-founder, Authenti.ca

A stable, well-shaped SDK is what makes that possible. It gives AI code generation a consistent target to aim at, so the model produces integration code that actually works instead of code someone has to rewrite. The no-code builder is the layer Borg skips. The SDK is the layer he bought. The forty-five-minute NetSuite integration is what that decision looks like in practice. By his account the engineering was the easy part, and what took the time was the security review, not the building.

02 What the easy version replaces

To see why that matters, it helps to look at the version without it.

Borg had recently set up a QuickBooks Enterprise Desktop integration for a customer. On-prem, hand-rolled, no managed sync to lean on. He is blunt about how that goes. It is a slog of auth issues and token refreshing, and the worst part is when the customer has to re-authenticate. A single expired grant can cost days, sometimes weeks, and there is no way to fix it without pulling the customer back in.

None of that work is technically hard. It is just relentless, and it never shows up on a roadmap. OAuth flows expire. Refresh tokens cycle. A customer admin changes a permission and every downstream connection quietly breaks until someone notices. For a normal SaaS product that is an annoyance. For an agent that is supposed to be acting inside a customer’s systems on its own, it is worse than an annoyance. An auth failure does not just slow Authenti.ca’s engineers down. It stalls the customer’s AI, in front of the customer.

03 Treating every agent as a user

The way Authenti.ca closes that gap is the piece of the architecture worth slowing down on, because it is what lets the agents run unattended in production.

Every Authenti.ca agent is treated as a real user across the systems it touches. Each one gets its own credentials, scoped to exactly what that user is allowed to do, and it calls the same Paragon-managed action a human would call, inside the same permission model. The token refreshing, the expiring grants, the credential upkeep that ate Borg’s QuickBooks week all sit below the line where his engineers have to think about them.

“We treat an agent as a user across all of these systems. That user-level auth maintenance is really helpful.”

Michael Borg, CEO and co-founder, Authenti.ca

Architecture diagram placeholder

Diagram should show customer systems, including ERP, CRM, inventory, and file storage, flowing into Paragon for Managed Sync, custom SDK integrations, and user-level auth. From there, data flows into the Authenti.ca ontology layer, which constrains agent tool calls, then into AI FDE workflow runs with a human review gate for the known failure category.

That is the detail that makes autonomous behavior safe to ship. An agent that can reach anything is a liability in an enterprise account. An agent that operates inside the customer’s own permission structure, and never around it, is something a security team can sign off on. Authenti.ca reinforces that with a knowledge graph it builds for each customer, an ontology it describes as a digital twin of the business. The agents do not roam. They call only the actions the ontology has scoped to them for a given task, and in every case the action underneath is a Paragon integration. Borg’s team runs evals across those workflows and shares the results with customers, so the roughly five percent of runs that fall into a known failure category get human review while the rest pass on their own.

Managed Sync is the other half of that discipline. Turned loose on live customer systems, exploring agents can burn through API limits inside a day. Reading against Managed Sync instead gives them a normalized, rate-friendly copy of the data to learn from before they ever take an action, which is exactly what those first weeks with a new customer call for. Borg maxed out his Managed Sync allocation in a demo environment and asked for more capacity.

04 Walking into the room with a digital twin

All of this adds up to something Authenti.ca can do that most agent companies cannot. They show up to a first sales meeting with a working model of the prospect’s own business.

Before a deal is signed, before Authenti.ca has touched a single one of the prospect’s systems, it assembles an ontology from public information and prior conversation transcripts and pulls it in through Paragon connectors. The demo runs against that digital twin. The agents have not seen the customer’s real data yet, but they have seen enough of the customer’s world to be convincing. As the conversation continues and more material comes in, the ontology sharpens and the agents get better. When the customer signs, Authenti.ca connects the live systems and the AI FDEs go to work for real.

Security, which you might expect to be the sticking point given that these are defense, manufacturing, and industrial customers, tends not to be. Because the platform underneath is SOC 2 Type 2, that conversation is short. One small moment from the interview captures how far the integration story carries on its own: Borg mentioned, half amused, that he had copied the logos straight off Paragon’s integrations page and mirrored them onto his own site, because showing a customer that breadth of coverage off the shelf does more selling than anything he could say about it.

05 What it adds up to

Five months after launching, Authenti.ca has agents in production across defense, manufacturing, industrials, and food, with retail next, and it got there without ever building an integration platform of its own. The engineering time the team would have poured into OAuth flows and token lifecycles went into the agent intelligence instead, which is the only part of the product a customer is actually paying for.

Borg’s advice to other teams building on agents lands on the same point.

“Integrations are likely one of the biggest bottlenecks you’ll face in providing ROI to the customer. The more you can connect your agents to the edges of their domain or even across the rest of the organization, the better the solution you can provide.”

Michael Borg, CEO and co-founder, Authenti.ca

Michael Borg

CEO and Co-Founder, Authenti.ca

Authenti.ca builds agentic AI for enterprise supply chain operations and ships AI Forward Deployed Engineers into production at customers across defense, manufacturing, industrials, and food. Paragon is the integration infrastructure behind their AI FDEs.

Building agentic AI on enterprise data?

Start with the Paragon Connect SDK that Authenti.ca points its codegen at, backed by Managed Sync for authentication, schema mapping, and sync infrastructure, so your team can stay focused on the agent.

Book a demo: https://www.useparagon.com/request-a-demo