Meet the Agents: How Noor and Vera Are Accelerating Competitive Intelligence at TRM
In this “Meet the Agents” post, we’re introducing you to two new members of our team, because they were designed to work as a pair.
Meet Noor and Vera — the newest members of TRM's sales enablement team.
- Noor is the researcher: She gathers, structures, and publishes competitive intelligence briefs every Friday afternoon
- Vera is the QA reviewer: She shows up 22 minutes later, reads everything Noor wrote, and catches accuracy issues before any of it reaches a battlecard
They’re named after two women from WWII-era intelligence history. Noor Inayat Khan was Britain's first female wireless operator sent into Nazi-occupied France. Vera Atkins was the intelligence officer who recruited, briefed, and tracked every agent she deployed into the field. Naming them after Noor and Vera felt right; two exceptional women who deserved to be remembered.
The problem they were built to solve: Competitive intel had become a monthly luxury
Before Noor and Vera, competitive intelligence updates happened roughly monthly, when someone on the enablement team had the bandwidth to manually pull together a brief on what competitors were doing, cross-reference it against our positioning, and update the battlecards that sellers were using in deal conversations.
To most, a monthly cadence likely sounds reasonable — until you consider that in our market, a competitor can launch a new product, shift pricing, or land a marquee customer in the span of a week. By the time a monthly update reached a seller's hands, it was already stale. And stale competitive intel in a revenue conversation is, at best, unhelpful — and at worst, actively dangerous. A seller quoting outdated competitor pricing or missing a new feature launch loses credibility in the room.
The real cost wasn't just the hours of manual research this process was taking (when we could dedicate time to it). It was the growing gap between what we knew and what our sellers knew, compounded every week we didn't create an intel update.
The architecture: Research agent + QA agent + one human checkpoint
We made a deliberate choice to split the work between two agents instead of building one that does everything because competitive intel flows directly into battlecards, which flow directly into revenue conversations. The need for accuracy in these deliverables is extremely high. We didn’t want an agent that was both researching and self-validating — essentially grading its own homework.
So we built a two-agent pipeline:
Noor (the researcher) pulls from a defined set of competitive sources, structures findings into a consistent brief format in Notion (including what changed, what it means for our positioning, and recommended battlecard updates) and publishes it.
Vera (the QA reviewer) triggers after Noor publishes. Vera reads the full brief, cross-references claims against source material, and flags anything that looks inaccurate, unsupported, or misleading. On her very first run, Vera caught an issue: a competitor feature claim that was technically true for one product tier, but not the tier most relevant to our deals.
That kind of nuance is exactly why Vera, our separate QA agent, was necessary. A single agent without a structured reviewer might have let that slide as "correct enough" (Vera didn't).
The human checkpoint: After Vera's review, a product manager (PM) does a final spot-check before approved edits hit the battlecards. We kept this human-led step in deliberately. Given how directly this intel lands in revenue conversations, a human gut-check is critical — and not because we don't trust the agents, but because the cost of a bad edit reaching a seller mid-deal is too high to fully automate away.
Now, what used to take hours of human research and updating requires less than five minutes of human involvement per cycle. And it happens on a weekly cadence instead of monthly.
The surprise: Same model, dramatically different quality
We knew having a second agent in the mix was a good idea; but what we didn't expect was just how much stronger that decisions made the deliverables. A separate QA agent dramatically increased output quality, despite Noor and Vera running on the same underlying AI models.
It mirrors something we've observed with human teams: the quality of work changes when you know someone else is going to review it. That dynamic holds for AI agents too.
Here's an example of the efficacy (and criticality) of having both agents working in tandem: A competitor announced an integration in an area where we were actually first to market. Noor surfaced the announcement and flagged it as a potential differentiator, meaning our sellers could have walked into calls positioning us as behind in a category we'd pioneered. But Vera caught it, cross-referenced against our own product documentation, and pulled it from the battlecard before it ever shipped. If that incorrect intel had reached a seller mid-deal, they may have inadvertently downplayed one of our strongest talking points without even realizing it.
What we got wrong (and quickly fixed)
Building two AI agents that work as a team sounds clean in theory. In practice, the first few iterations taught us a lot about what "good enough" actually isn't.
Stale sources slipping through
Early on, neither Noor nor Vera were verifying publication dates on the external sources they pulled from, including blog posts, press releases, and product announcements. That meant Noor could surface a three-month-old blog post and drop it into that week's competitive briefing as if it were breaking news, simply because it was the first time Noor had encountered it. The fix was straightforward: we added mandatory date verification to every polling cycle so nothing could make it into a brief without a recency check.
Grading our own homework — again
In one early run, Noor pulled in a competitor's product announcement and positioned it as a gap in our offering. The problem? We already had that capability, and had shipped it first. Noor had checked the competitor's press materials but hadn't cross-referenced against our own website, blog, or product docs. That's how we landed on what we internally call "Step 8" in Vera's verification workflow: any time Noor's brief positions a competitor as ahead of TRM on something, Vera now runs a detailed check against all of our internal resources before that claim goes any further.
Too many battlecard updates, not enough signal
Noor and Vera turned out to be overachievers, and recommended battlecard changes on nearly every run. That sounds productive until you realize that most of those suggestions were time-specific details that would go stale within a week. We had to retrain the agents to distinguish between evergreen differentiators (the kind of positioning statements that belong on a battlecard) and weekly intel (the kind of context that belongs in the Friday briefing but not in a living sales asset). Teaching them to be more selective about when a battlecard actually deserved an edit was one of the bigger programming iterations we went through.
What Noor and Vera are not
Noor and Vera don't replace the enablement team's judgment about which competitive shifts actually matter for our current deal landscape. They can surface that a competitor launched a new integration. But they can't tell us whether that integration matters to the three enterprise deals we’re closing this month. That prioritization still lives with the humans who know the deals inside and out.
Welcome to the team, Noor and Vera. The women they're named after spent their careers making sure nothing slipped through the cracks. These two are holding up their end.
{{horizontal-line}}
This post is part of a series where we introduce the AI agents TRMers have built. Each one has a name, a role, and a story. If building systems like this sounds like your kind of work, we're hiring.
