How to Choose Your Tech Stack: Data, Then Agents, Then Screens

At a sync meeting last week, the operations lead at a licensed apparel company told me what she liked about how we were building: the flexibility, having all of their data in one place, and not being tied to any one vendor.

That last point explains the whole shape of the build. When people ask what we are building, I give them one line. Data at the bottom. Agents in the middle. Screens on top.

Where the data lived

When we started discovery there, the company ran on a handful of systems: an online store, an operations platform that holds orders and invoices, an accounting package, a CRM that also carries the art requests, and a shared drive full of live documents. In between them sat spreadsheets, what felt like hundreds, one or more for nearly every person. That is how a capable team keeps a growing business moving when its systems do not talk to each other.

Each spreadsheet worked for the person who kept it. None of them worked for anyone else.

The first request we heard was for reporting, one place to see the business. Then came a long list of jobs people wanted off their plate. It would have been easy to answer each request separately, with a tool apiece. That is how a company ends up with a dozen logins and a dozen copies of the customer list, each slightly different.

The bottom layer: one platform the company owns

So the foundation is a data platform. It pulls from the systems the company already runs and keeps one copy of the data, on infrastructure the company controls. People sign in with the work accounts they already use. A portal shows each person the apps their role needs, and nothing else.

It is not finished. We are partway through, and it will be the longest-running piece of the work.

This is the layer that answers her point about vendors. When the data lives in one place that belongs to the company, any tool feeding it or reading from it can be changed later without starting over. Replace any one of those systems in two years and you rewire one connection, not every report and agent that ever touched it. That is what I mean by Stack Don't Scatter: each new piece stacks on the same foundation instead of arriving with its own.

The middle layer: agents that do one job each

On top of the data sit agents. Each one does a piece of work that a person used to assemble by hand from three or four places.

Meeting prep. Before a sales meeting, a brief lands in the rep's calendar invite: who the account is, and what they have bought this year and across the whole relationship.

Pricing decks. A proposal and pricing deck generator that puts the product on a model in the deck, for a few cents a run. It shipped in June.

Collections. Consolidated customer statements, a reminder cadence, and credits for short payments. In build now.

Royalties. The monthly accrual, calculated per licensor, with an export to the general ledger. In build now, alongside the accountant who does it every month.

Each of these is small, and that is deliberate. The expensive part of any agent is getting clean data to it. When the platform does that once, each new agent is mostly the logic of the job.

Now the part that did not go in order. The meeting-prep agent and some of the early automations were built before the platform existed and read straight from the source systems, because they were worth having before the platform was ready. That was the right call for speed. It also means some of that wiring will be redone as the platform catches up, and I would rather say so now than discover it in a year.

The top layer: simple screens for the people who look

Screens are the top layer, and there are fewer of them than people expect. An executive dashboard for the chief executive, who asked for one view of the business. A sales dashboard, with email activity, for the director of sales.

A screen is the right answer when looking is somebody's job. For everyone else, the output should arrive inside something they already open. That is why the meeting brief sits in the calendar invite and not on a dashboard.

Draw yours on one page

You can do this before you talk to anyone. Take a sheet of paper and draw three bands.

In the bottom band, list every system that holds data you care about, spreadsheets included. Put a mark next to each one your company does not control.

In the middle band, list the jobs you want done, the ones a person assembles by hand from several of those systems. Draw a line from each job down to every system it needs.

In the top band, list the people whose job includes looking at something, and what they look at.

If the lines from the middle band cross all over each other on the way down, you do not need more tools yet. You need the bottom band. And when a vendor offers you an agent, ask one question before any other: where does its copy of my data live, and who owns it?