The right screen beats the right sentence

Jacob Koenig — Thu, 03 Sep 2026 19:14:08 GMT

Buy-side firms are spending heavily on AI, and many aren't seeing the ROI they expect.
This comes from months spent building an industry initiative we call Project Catalyst,
sitting with trading desks, execution venues, data providers and order management
vendors. The same gap kept coming up: the AI often can't reach the systems where the
work actually happens. Project Catalyst is that initiative, and the firms building it with us
decide what gets built first.

A reasonable place to start
The most advanced firms are taking a similar approach: land all accessible data in one
place, wire agents to it, let them reason across it. The experience is a chat interface, text
back with tables and charts generated on the fly.
That's the right shape for some questions. Ask which of your positions carry earnings risk
this week and a sentence answers it fine. But most of what happens on a trading desk
isn't a question with a sentence for an answer, and that's where this approach runs out.
Traders still need to see things
A chat interface mostly hands back a wall of text, and a summary of your own data is
something you now have to decide whether to trust. That decision often costs more than
the answer saved, especially given the auditability this business requires.
A blotter, a curve, a depth ladder carry a level of complexity text can't, and the trust that
comes from seeing the real thing rather than a description of it. Even a generated chart is
a reconstruction, drawn from the same sources the text would've used, and it still can't
take an action.
And the back half is harder than it sounds
The deeper issue: most AI can't reach the systems where the work lives. In market-facing
work the data can run to trillions of rows, too large to copy anywhere else, and an agent
without your current context has the least to offer exactly when you need it most.
SimCorp's 2026 InvestOps survey puts 70% of buy-side firms now running AI in the front
office, up from roughly 10% a year ago, but only 20% have a unified platform delivering
real-time decision support. Operations adopted first with reconciliation and reporting.
Research followed with filings and transcripts. The trading desk is still mostly untouched,
and MCP, the emerging standard for letting AI reach into applications, only closes part of
that gap today: coverage is early, and most vendors haven't built a server yet.
The tractable problem
Getting the right screen in front of you is easier than getting the right number into a
sentence. If an agent gets a number wrong in a chat window, you've lost money and
maybe have a compliance issue too.
Your vendors already compute the number correctly, under your entitlements. We don't
need to rebuild that, we need to open it. Many providers license a live feed and restrict
onward distribution, which stops centralisation projects before they start. None of that
matters if the answer is read in the vendor's own window instead of copied out of it.
So the AI opens the application that already holds the answer, with the right thing already
on it. We call this invocation. What you read is the vendor's own screen, not the model's
account of it.
That's the gap Catalyst closes. Your vendors' applications run in one surface, your AI
sees what you see, takes action and builds what's missing. Nobody rips out what works,
and nobody has to trust the model's recall, since you're looking at the same screen you'd
have opened yourself.
That also solves the audit problem. Auditing AI is usually reconstruction: the model acts,
and someone assembles logs afterward into an account of what it probably did. Under
invocation, what the user was looking at when they made the call is the record. The
evidence and the decision are the same artifact.
What that looks like on a desk
Take a dealer's axe list. Twenty-plus names land on it and only a handful matter to what
you're holding. Checking means going name by name through your own stack, the OMS,
your internal research, wherever else you'd normally go.
Invoke an agent into that stack instead and it opens each screen for you, checks every
name against your positions and your PMs' mandates, and surfaces what's relevant in
one dashboard. You checked all twenty, not the six you had time for, and once you verify,
it pushes your selection back to the dealer.
Moving the data to the AI is one route. Bringing the AI to the screen is the other, and it's
far shorter.
Vendors fear becoming a dumb pipe behind someone else's screen. Invocation puts their
own interface in front of the user at the moment of the decision, which is why they're
willing to open up. Neither side gets there alone, which is why this has to be built on
neutral ground, shaped by the firms and vendors who live with the result.
Get involved
We're running pilots now, inside firms' own environments, on their own models. If you
want to see it on your stack, or you're building something a desk should be able to reach,
talk to us.
Details are live at here.io/project-catalyst. What gets built first is decided by the desks we
build it with. Email us at catalyst@here.io

HERE_Project Catalyst

The right screen beats the right sentence