04 · AI platform strategy

They wanted to build a new platform. I got them onto the one they already had.

A research institute came with six agent ideas and a plan to build its own platform to run them, without an engineering team to build it. I showed them that the firm's existing data platform and AI gateway already did what they needed, so we could test and learn right away instead of waiting on a new build.

6 → 1agent requests, one plan on platforms the firm already runs

My roleDiscovery lead; made the case to reuse
WhenMay – Sep 2026
Domaina research institute's data and tooling
StatusShared credential retired; on the firm's gateway; data move underway

A research institute wanted to be "largely agentified": agents for a public microsite, a question-answering assistant over its knowledge base, data apps, newsletter and research automation, and a knowledge graph. Six requests, six sponsors, and a plan to build a brand-new platform to run them all. But they had no engineering team to build or maintain one. And underneath the requests sat one shared credential, no way to track usage, and a legacy data platform that was hard to find. Only a dozen or so of about three thousand datasets were actually used.

I led discovery across all six requests and treated them as one piece of work instead of six. I wrote the prioritization and the recommendation to the sponsors. Most of my time went into the conversations that changed the plan: showing each sponsor that what they wanted already existed, and that reusing it would get their agents live sooner.

Use what already exists, then test and learn. A new platform would have needed an engineering team they didn't have, and a lot more time. The firm's existing data platform already met their needs, and the firm already ran an AI gateway that handles sign-in, routing, and usage tracking. So instead of building, we moved them onto both, in three steps: give each agent its own account in place of the one shared credential, connect the agents to the existing gateway, and move the data off the legacy platform.

That took influence more than architecture. Each sponsor came in attached to a build plan, so I framed reuse as the fastest way to get their agents in front of people, not as a compromise. The build-or-buy call stayed theirs: I wrote down the few facts that would flip the answer, and they made it.

Everyone was excited to build more AI. Underneath was one credential and no way to see who used what. We fixed that first.What I told the sponsors
  1. six agentsmicrosite, Q&A assistant, data apps, newsletters, knowledge graph
  2. own accountsone per agent, in place of one shared credential
  3. the firm's gatewaysign-in, routing, and usage tracking, already running
  4. the firm's data platformafter the move off the legacy one

Nothing new to build or staff: every step joins something the firm already runs.

The plan they had, building a new platform that needed an engineering team they didn't have, crossed out. The plan we ran instead, in three steps on things the firm already runs: give each agent its own account, join the firm's AI gateway, and move the data to the firm's platform. the plan they had build a new platform needs an engineering teamthey didn't have reuse the plan we ran 1 · give each agent its own account2 · join the firm's AI gateway3 · move the data to the firm's platform the one shared credential is goneusage tracked from day one, nothing new to buildoff the legacy system no new platform, no new gateway, no team to hire
Instead of a build they couldn't staff, three steps onto things the firm already runs, in the order the risks demanded.
Built withWorkflow mapsRICE + DFVLeast privilegeThe firm's AI gatewayBuild-or-buy, decided by the teamSecurity review
6 → 1
requests, one plan on platforms the firm already runs
0
new platforms or gateways built; the agents moved onto ones the firm already runs
7 figures
a year, documented in the institute's brief, from retiring the legacy platformnet of the migration, it's lower in year one

The risky part was fixed before any new agent went live. The shared credential was replaced with separate accounts first. The agents now run through the firm's existing gateway, so usage is tracked from day one, and moving the data is underway. The team made the build-or-buy call themselves: the firm's platform for the data, and only the app as their own.

The team owns the call. One plan replaced six separate projects, six sponsors follow one sequence, and the build-or-buy decision is theirs instead of inherited from me.

  • Start with what already exists. The fastest platform was the one the firm already ran. Convincing people of that was the real work.
  • Leading sometimes means saying not yet. The hard decision was the deliverable.