How a Full Day of Portfolio Risk Review Became 10 Minutes

An asset manager at Triten opens ChatGPT and types one sentence: "Tell me the top 20 assets I should be worried about, why, and who manages them."

A few minutes later there is a ranked list. Each asset comes with the reasons it made the list, the figures behind those reasons, the source records they came from, and the asset and project managers responsible for it. Nobody exported a spreadsheet. Nobody chased a colleague for the latest board.

The same review used to take an employee a full working day: 8+ hours across SharePoint, Monday.com and Yardi. Now it takes about 10 minutes.^E1 This is how that changed, and what it means for any fund that runs on more than one system.

What did the review look like before?

Triten's portfolio data lives where most firms keep it. Financials and leases sit in Yardi. Project work and ownership sit on Monday.com boards. Loan files, reports and the documents that explain everything else sit in SharePoint.

A risk review meant one person opening all three. They pulled the numbers, read the documents, compared one source against another, and then did the slowest part: working out who was actually on each asset. Boards drift. Documents go stale. Two systems can name the same property differently.

By the owner's account, that took an employee 8+ hours each time.^E1 It also produced an answer that was hard to check. If an executive asked "why is this one on the list?", the honest answer was "because I read it somewhere this morning."

What changed?

Two things were built, in order.

First, a governed data platform, Triten OS. It reads Yardi, Monday.com and SharePoint and turns their records into facts. Each fact keeps where it came from, when it was true, who is allowed to see it and how it was derived.^E2 When two sources disagree, the platform records the conflict and applies written rules about which source wins and why. Those rules are deterministic; the AI model does not decide what is true.^E2

Second, a read-only MCP server on top of it. MCP is the standard that lets assistants like ChatGPT and Claude call outside tools. Triten's server offers 22 tools, every one of them read-only, and every one of them goes through the same governed API the web app uses.^E2 There is no write tool. The assistant can ask; it cannot change.

Among those tools are investigations built for questions like the one above. They rank exposure across the portfolio by property, attach the responsible asset managers and cite the underlying records. Before release, the investigations were proven read-only against production data.^E4

How does ChatGPT answer a question like "which assets should I worry about"?

The assistant does the reasoning; the platform supplies the facts.

When the asset manager asks for the top 20, ChatGPT calls the investigation tools, which return ranked properties with the evidence behind each one. It pulls the financial and operational facts it needs, joins each asset to its project and its managers, and writes the answer in plain language.^E3

Three properties of that answer matter more than speed:

• It is cited. Every figure points back to a governed fact, and every fact points back to its source record.^E2
• It respects access. The model sees only what the person asking is allowed to see, because security labels are enforced before data leaves the API.^E2
• It names owners. The list ends with people to call, not just problems to worry about.^E3

You can try the same experience on synthetic data in the portfolio risk demo. It shows the question, the ranked list, the reasons and the owners, without any real Triten data.

Why read-only, and why does governance come first?

Executives ask the right question about AI and their data: what stops it from being wrong, or from doing something it should not?

The answer here is architecture, not prompts. The assistant never touches Yardi, Monday.com or SharePoint directly. It has no credentials for them. It reaches data only through tools that can read and nothing else.^E2 If a document contains text that looks like an instruction, it is treated as data.

Governance is what makes the speed safe. A fast answer from messy data is a liability: it sounds confident and nobody can check it. A fast answer from governed facts, with sources attached, is something an asset manager can act on and an auditor can trace.

That ordering is the main lesson for anyone planning a similar project. Connecting ChatGPT to a database takes a week. Making the answers trustworthy is the actual work, and it is most of the work.

What does the time saving actually mean?

The headline is a day of work turned into minutes.^E1 The bigger change is what happens to the question itself.

When a review costs a full day, it happens on a schedule, maybe monthly, maybe before a committee meeting. When it costs about 10 minutes, it can happen whenever someone is curious.^E1 An asset manager can ask a follow-up. A COO can ask the same question about a different slice of the portfolio before a call. The review stops being a report and becomes a conversation.

It also changes who can do it. The old process depended on one person who knew where everything lived. Now anyone with access can ask, and the answer shows its work.

Could this work for my fund?

If your portfolio data is spread across a property management system, a project tool and a document store, you have the same problem Triten had. The systems do not need to change. What needs to exist is a governed layer that reads them, reconciles them and exposes them safely to the assistants your team already uses.

The pattern is repeatable:

1. Read each source and keep its records as evidence.
2. Turn them into facts with clear rules for conflicts and access.
3. Expose those facts through a narrow, read-only set of tools.
4. Let the assistant reason over them and cite what it used.

See the full project record at Triten OS, try the synthetic demo, or book a call to talk through what your version would look like.