# Choosing an AI Platform for an Investment Office

> An investment office with a small tech team had to pick an AI platform and defend the choice. I turned its use cases into workflows a vendor could be tested on, held six vendors to the same seven gates, and laid out two paths. The executives chose the one our deck didn't lead with.

By Ben Shapiro · 2026-08-17
Web version: https://shapyro.com/work/investment-office-ai-platform

- Organization: Investment office of a large institutional investor
- Sector: Investment Management
- My role: Delivery lead from my first call to the handoff. A colleague sold the first phase.
- Team: A Cadre sales lead on contract terms, two colleagues on schedules and cost analysis, and the office's own engineers, who built every tool they run.
- Timeline: Feb to Aug 2026
- Tools: A workflow workbook, a gated vendor scorecard, demos of six platforms including Claude
- Outcome: The executives chose Claude alone. Two follow-on contracts signed.

## In numbers

- **20 workflows** specified from 18 loosely written use cases
- **6 vendors** held to the same 7 pass-or-fail gates, Claude included
- **2 contracts** of follow-on work signed after the decision

## The ask

In February 2026 the investment office of a large institutional investor asked Cadre, the AI consultancy I co-founded, which AI platform to bet on.

They had a list of 18 use cases, a shortlist of vendors and a Claude pilot already running. Their engineers had built a prototype agent of their own. The tech team was small, with about two and a half people's worth of time to spare, and it was partway through choosing a data platform.

Nothing on the list could be compared yet. Many of the 18 use cases were two or three builds sharing one name, so there was nothing to test a vendor against. The office wanted a decision it could defend to its executives.

## What I did

The evaluation ran from February to April. I built the workbook, drafted the scorecard, organized the vendor demos and presented both reviews.

- **Workflows.** I split the use cases into single workflows. Each got a trigger, inputs, outputs, the steps between, the data it must have and where that data lives. The final review had 20, after six items turned out not to be workflows and two fell out of scope. Their tech team liked the format and built their own tracker in it, with an owner on each workflow.
- **A map of the data.** A second sheet listed 20 data sources and how an AI tool could reach each one: an API, an MCP connector, or nothing yet.
- **Gates.** The workflows surfaced the must-haves, and their tech lead added one: a vendor had to say which models it ran on. By the final review there were seven pass-or-fail gates: CRM connection, model transparency, security, depth in finance, company viability, where data is stored, and clear pricing. A vendor that failed one was out, so a strong demo couldn't make up for a miss. Every vendor claim was labeled demoed, claimed or inferred.

## What the gates showed

| Vendor | Result |
|---|---|
| Finance platform 1 | Passed all seven. Its CRM connection was claimed, not shown. |
| Finance platforms 2 and 3 | Out on model transparency |
| Finance platform 4 | Out on company viability and pricing |
| General research tool | Out on CRM connection and depth in finance |
| Claude alone | Missed one gate: depth in finance |

Claude went through the same gates as everyone else and missed one. Nothing in it was built for an office that invests through outside fund managers. It passed the CRM gate through a connector the office's own engineers had already built, because the CRM didn't come with one.

## The recommendation

That left two paths: buy the finance depth, or build it.

- **Path A:** finance platform 1 as the main tool, with Claude beside it. By our rough estimate it covered 19 or 20 of the 20 workflows in the first 90 days.
- **Path B:** Claude alone, across chat, Cowork, Claude Code and the API. Our estimate was 15 or 16 of the 20. Their engineers would build the finance-specific parts and keep full control.

If speed mattered most, A. If long-term control mattered most, B. The office's leads leaned A, and so did our deck.

## What happened

In early July the executives chose Path B. Their reason wasn't on the scorecard. The wider organization was standardizing on Claude, and they wanted the investment office on the same platform.

In June we had drafted the follow-on plan in two versions, one for each path, so the Claude version was ready to finish. It covered Claude training for the technical and investment teams and advice on data design. The data platform choice stayed with their own team. A second contract covered training for department champions and executives across the wider organization. I scoped both, and both were signed in August.

I wrote the handover guide for the colleagues taking over and joined their first calls. I left Cadre before any training was delivered, so I have no results to report from it.

## What I learned

The scorecard didn't make this decision. A choice made above the investment office did. The work still held up, because both paths had been scoped before anyone chose.
