Rhea Morales runs Brand Ops at a Fortune 100 consumer brand whose name is not on this page for procurement reasons. The brand spent most of 2024 and the first half of 2025 running seven concurrent AI pilots across the marketing organization — a writing assistant for the social team, a campaign-brief drafter for the brand team, a translation tool for the global team, two competing creative-production tools that ended up in different business units for political reasons, a content-summarization tool for the leadership team, and an internal chatbot built on top of all of the above by the IT organization. By week eight of 2026 she had killed all seven and replaced them with one architecture install. Her team's output is up. Her team's tooling cost is down. The board, predictably, has questions about both numbers.
We sat down with her for thirty-two minutes during the second week of April. She had just finished the QBR where she presented the consolidation results. The conversation has been edited for length and to remove the parts where she names specific vendors.
On the moment she decided to consolidate.
Knyte: What was the meeting that changed your mind?
Rhea Morales: It was not one meeting. It was a procurement review where we were renewing two of the seven, and the procurement person asked me a question I could not answer, which was whether the two we were renewing did the same thing. I said "sort of" and she said "is it sort of, or is it yes," and I went back to my desk to find out. The answer was yes. The reason we had two was that one of them had been bought by a director who had since left and the other had been bought by a different director who did not know about the first one.
Knyte: And then you started looking at the other five.
Rhea Morales: I started looking at all seven. I made a spreadsheet. I had used spreadsheets before to do this kind of thing, and I knew it would not be honest unless I made the columns specific. So I made the columns specific. "What workflow does this tool support? Who is the editor of last resort? When this tool goes down, what happens? When this tool's contract ends, what happens to the work it produced?" When I filled in those columns honestly, four of the seven were doing overlapping work, two had no editor of last resort, and six of them produced no asset that we would still have if we cancelled.
Knyte: Six of seven.
Rhea Morales: Six of seven. The seventh had a vendor who had given us a way to export, but the export was so locked to their format that it was not going to be useful to anybody else. So really, seven of seven, if I am being honest.
On the architecture-call frame.
Knyte: You ran an architecture call before the consolidation. What did that change for you?
Rhea Morales: It changed the question. Before the architecture call, I was asking which of these seven tools we should keep. After the architecture call, I was asking what the right architecture was for the work my team actually did, and which of those seven tools fit inside it. The answer was zero. The architecture I needed had four properties — a brand-trained model that was mine, a corpus my team could query, an editor-in-the-loop runtime that defaulted to review, and an audit trail that survived our SOC review. Two of the seven tools had one of those properties. None of them had all four.
Knyte: Did the call surface anything you had not considered?
Rhea Morales: Two things. The first was that the cost of running seven shallow tools was higher than the cost of running one deep architecture by a factor I had not believed until I saw the spreadsheet. The math is in the replacement-math article you linked me to. I had assumed the seven tools were collectively cheap because each one was individually cheap. They were collectively expensive in a way that did not show up on any single procurement decision.
Rhea Morales: The second was that the asset from the seven tools was zero. I knew this intuitively but I had not made it explicit. When I sat down with the architecture-call frame, I realized that everything my team had taught those seven tools — every brand-voice tweak, every campaign-brief template adjustment, every translation memory entry — would evaporate the day we stopped paying. We had built nothing that was ours. We had been renting a capability and pretending we were building one.
On the install.
Knyte: Walk me through the eight weeks.
Rhea Morales: Week one was data ingest. We pulled three years of campaign briefs, brand guidelines, customer research, and shipped creative into the corpus. The corpus indexer took about two days. We thought it would take two weeks, which I want to flag because I think most procurement teams over-budget that step.
Rhea Morales: Weeks two through four were the brand-voice fine-tune. We had editors review side-by-side outputs of the base model and the fine-tuned model on a hundred and twenty representative briefs. The fine-tuned model started winning at about brief sixty. By brief one-twenty it was winning consistently. The editors signed off.
Rhea Morales: Weeks five and six were the workflow build. We picked three workflows to migrate first — campaign-brief drafting, social-copy generation, and competitive-positioning research. Three workflows, not seven. I want to be clear about this because the urge was to migrate everything immediately. The team that runs Knyte refused to let us. They said depth on three would beat breadth on seven. They were right. We have since added a fourth workflow.
Rhea Morales: Weeks seven and eight were the cutover. We turned on the new workflows in shadow mode for a week, with the old tools still running, and compared outputs. The new workflows were producing more, getting accepted at a higher rate, and citing the corpus correctly. The editors were happier. We turned off the seven tools at the end of week eight, in two batches over two days, with a written rollback plan that we ended up not needing.
On the metrics.
Knyte: What does the QBR slide say.
Rhea Morales: The headline is replacement math. We replaced seven contracts. The annualized contract spend was a number I am not going to say on the record but it was substantial. The Knyte install costs less than half of that, and that includes the install fee. So immediately we are net-positive on the spend line.
Rhea Morales: The second number is compounding output. The brief-drafting workflow is producing about three-and-a-half times as many editor-accepted briefs as the previous combined tool stack was producing, with the same headcount. The social-copy workflow is at about two-point-eight. The research workflow we cannot compare because we did not have a previous tool stack for it; that is new.
Rhea Morales: The third number, which is the one the board cares about most, is editor satisfaction. We did the same survey we had been doing for eighteen months. The team's reported satisfaction with their AI tooling was up substantially. The thing the team likes most is that they no longer have to context-switch between seven tools. There is one place to do the work.
On what surprised her.
Knyte: What did you not expect?
Rhea Morales: I did not expect the audit benefit. We had been on the SOC 2 audit prep cycle the whole time the seven tools were in flight. The audit prep was painful because we had to chase down data-handling addenda for each of them, get attestations on data classification, and explain what was being processed where. The Knyte install consolidates that into one trust-center conversation. The audit prep cycle this year is going to be hours instead of weeks.
Rhea Morales: I also did not expect the speed. The architecture call was forty-five minutes. The install was eight weeks. I had budgeted six months. The thing that took the time, in retrospect, was the political work to get the seven directors aligned on the consolidation. The technical work was bounded.
On what she would do differently.
Knyte: If you were starting over.
Rhea Morales: I would do the architecture call before the first procurement decision, not after the seventh. The cost of the procurement spread was not the contracts; it was the political work to consolidate. If I had started with the architecture frame I would have headed off five of the seven contracts. The other two might still have happened — there were genuine reasons in those cases — and we would have wrapped them into the architecture install at the appropriate phase.
Rhea Morales: I would also have built the editor-in-the-loop discipline earlier. We had been treating editor review as a special case for the seven tools — only the most consequential outputs got reviewed. The Knyte default is the opposite, and the difference in the quality of the model's training signal has been visible. I wish I had operated my team that way for the prior three years.
On what is next.
Knyte: What is the next twelve months.
Rhea Morales: Depth on the existing pillars. We are not adding new workflows for at least two quarters. We are concentrating the editorial attention on the three we have, watching the compounding curves, and planning the fourth and fifth workflows for the second half of the year. The discipline is to expand depth before expanding breadth, which is exactly the lesson of the seven-tool spread.
Rhea Morales: I will say one other thing. The reason this conversation exists is that I made the consolidation public. I told peers in other organizations how it went. About a third of them have either started the architecture conversation or scheduled their own. The pattern is replicable. The hard part is the political work. The technical part — the install, the cutover — is bounded and survivable.