What We’re Learning as We Build AI Agents Across Our Portfolio

AI agents are having a moment defined by two opposite reactions at once: excitement about what they might do, and anxiety about what they might replace.

As Microsoft's 2026 Work Trend Index put it, "the anxiety around AI at work is real," running from job loss fears to the pressure of keeping up with quickly advancing technology. At the same time, the report found a growing number of workers are already "using AI in advanced, resourceful ways."

The tools are outpacing most organizations' ability to absorb them.

We didn't want to write another piece vaguely weighing in on that tension. So instead, we sat down with Nik Gauvreau, a software engineer who joined SureSwift in March to help our software businesses put AI agents to work, to talk about what's happening inside our portfolio right now — not what agents could theoretically do someday, but what they're doing today, in real businesses, with real customers and limitations.

Table of Contents

• From experimenting with AI to relying on it
• Finding where AI earns its place
• How Ghost Inspector and MeetEdgar are putting AI to work
• What makes an AI agent trustworthy enough to use
• The advantages of using AI across our portfolio
• AI is getting better. What does that mean for our teams?

From experimenting with AI to relying on it

Gauvreau has spent the last three years working almost exclusively with AI for software development, and before joining SureSwift, he ran his own businesses, which shapes how he thinks about AI: less like a technologist chasing a trend, more like an operator looking for leverage.

"One of my first lightning rod moments," he told us, was an early experiment with ChatGPT, version 3 or 3.5, on a side project — a graphical editor for a game he was building. Out of curiosity, he asked the AI to add a new feature.

"Not only did it do what I asked, but it actually did it better than I would've. But I didn't have to do it myself, and it just automatically generated it for me. That was a big moment for me."

Today, Gauvreau uses that same filter — handing something to AI to see if it comes back better than he could make it — to run SureSwift's AI initiatives through.

Finding where AI earns its place

Ask Gauvreau where AI agents don't belong in a software business, and he'll tell you almost nowhere. "As far as what people should be using AI for, in my opinion, it should be everything," he asserted.

That conclusion comes from solid experience in an experiment he ran on himself. In his first three days at a previous company, Gauvreau put together a report on everything the business should be doing to modernize its operations. Afterward, purely as a test, he fed the AI the same context he'd absorbed over those three days, without his own report, just to see what it would come up with.

"It managed to pick out everything I'd said. It didn't miss anything. And it was able to expand on some of those things even further."

That doesn't mean every idea is worth building, though. The questions Gauvreau asks include: does this save real time, improve quality or consistency, or change a customer outcome — and how can you tell?

Scoping tasks is key

A good illustration is running research to find new marketing angles for a portfolio company. Rather than rely on his own experience, he had AI investigate how competitors and other startups have used unconventional, low-budget tactics to get attention — the kind of scrappy, guerrilla marketing approach bootstrapped startups lean on when conventional advertising channels are out of reach.

The output was pretty useful — 450 distinct tactics — but the process itself became a warning about scope. "It killed my AI credits," Gauvreau laughed. "I have the largest plan Anthropic offers, and I actually have more than one, because of this problem." The idea was good, but the execution needed guardrails around how deep to let an open-ended research task run before it becomes its own cost center.

That's the pattern SureSwift is learning to watch for: AI is rarely the wrong tool, but an unscoped task is often the wrong use of it.

How Ghost Inspector and MeetEdgar are putting AI to work

Gauvreau currently works across three SureSwift companies — Ghost Inspector, an end-to-end website and app testing tool; MeetEdgar, a social media automation and scheduling platform; and LeadDyno, an affiliate marketing platform.

Ghost Inspector and MeetEdgar are where the AI-driven rebuild has gone deepest. Both started in what Gauvreau called a state of 'disarray.'

Ghost Inspector

At Ghost Inspector, the priority coming in was development velocity and modernization. The product was in maintenance mode and needed new features and a broader refresh to keep pace with where the market was heading, while a fairly heavy monthly infrastructure cost sat mostly unaddressed.

"A bunch of low-hanging fruit was discovered through AI," Gauvreau explained, and the team has been working through it every couple of weeks, steadily making the project leaner.

MeetEdgar

MeetEdgar's starting point was more foundational. The engineering process was in disarray — most notably, the ad-hoc continuous integration and deployment (CI/CD) pipeline had gone completely unused.

In practice, that meant developers were pushing code changes straight to production servers rather than routing them through automated tests first, because the test suite was broken. Fixing that wasn't a quick pass — it took AI roughly four hours just to repair the existing tests, followed by weeks of fallout as issues surfaced, before the team had a deployment pipeline they could trust again.

A structured workflow and improved QA

With that foundation in place, Gauvreau built two structured, human-gated workflows that now run most feature and bug-fix work across both businesses.

Each stage is segmented based on a set of deliverables the AI needs to provide for users to approve before moving on to the next workflow stage. They look roughly like this:

Work Ticket (5 stages):

1. Investigate, reproduce, scope: Reproduce the bug (or scope the feature), write a failing test or acceptance criteria, research root cause/similar existing work, decide if mockups are needed. Deliverable: findings brief + failing test/screenshots.
2. Plan: Turn the confirmed scope into a concrete implementation plan — files touched, edge cases, test strategy, rollout considerations. Mockups go here if flagged. No code yet.
3. Implement and pull request: Build the approved plan, make the test pass, open a pull request for review (but don't merge yet), capture before/after evidence.
4. Deploy to staging and draft QA handoff: merge to preproduction, verify end-to-end, draft a plain-language QA handoff comment for the ticket.
5. QA handoff and wait: Post the QA comment, move the ticket to "needs QA," notify the reporter/QA team, then wait.

Deploy-to-Production (3 stages):

1. Pre-deploy analysis (go/no-go): Reconcile what's actually shipping (code changes in the repository vs. tracker vs. cohort), verify QA sign-off per ticket by reading comments (not statuses), audit pending database changes (migrations) and manual steps, classify each as blocking or follow-up, produce a go/no-go report, open the release PR.
2. Execute deployment: Re-verify nothing's shifted, confirm what merging triggers, wait for the automated tests and checks to finish, merge, watch deployment (new code) land in production, verify health, capture the exact revert procedure.
3. Announce deployment: Post the release notice to the team Slack channel with an @mention, once the user has signed off.

The throughline: "It always ends with a deliverable that the human reviews before it goes on to the next stage." But between those human checkpoints, AI is doing the investigation, the drafting, and increasingly, its own first-pass testing.

That last part turned into one of the bigger operational lessons of the whole rebuild. Gauvreau started having the AI test its own changes on the staging server — launching a browser, walking through the new feature, and returning screenshots — before handing anything to the human QA team.

"It's actually improved our QA throughput, because I'm no longer relying on QA to tell me what's wrong with it. Hopefully it's already good, and then it's a smoother process."

When the human touch is still needed

There are still cases where a person has to step in directly: live customer issues that require deep, specific product knowledge, and anything touching the production database.

When a MeetEdgar customer's social integration breaks, for instance, Gauvreau still relies on two part-time developers who previously worked on these products and know the platform's quirks well enough to often diagnose the issue immediately — faster than he could investigate it from scratch with AI.

"That's freed up my time. I can concentrate more on the stuff AI does really well, like implementing a big new feature, or just tweaking things, rather than debugging deeply ingrained issues with existing customers."

What makes an AI agent trustworthy enough to use

The clearest guardrail Gauvreau has put in place across the portfolio is also the simplest: he doesn't give himself write access to production databases. "The reason for that is I want to be able to move fast, so I don't want to be checking everything. It's about speed."

But if a database change is needed, it gets delegated to someone else on the team with that access — someone who "approaches the problems differently," as Gauvreau describes. "They're being more careful about it. They're verifying everything the AI does before it executes."

It's an intentionally artificial silo, but it means that even in the worst case, a bad deploy can be reverted quickly, and the riskiest category of mistake — irreversible data loss — is walled off from the fastest-moving part of the workflow.

Slack ops-agents

That distinction between internal, reversible work and anything touching live customer data shows up again in how SureSwift's Slack-based operations agents are scoped.

Every business unit has its own agent setup, generally split between an operations agent anyone on the team can talk to, and a more restricted general-manager agent with access to sensitive financial and business data.

The operations agents can draft work tickets, pull reports, and even run read-only investigations against the database to explain why a customer might be having an issue — but the access is deliberately one-directional. They can look, but they can't change anything without a person in the loop.

'QA is the new bottleneck' — and what we're doing about it

Where trust breaks down fastest, in Gauvreau's experience, is quality control on the output itself — which is also why QA has become the real constraint on how fast his teams can ship.

"Software development is no longer the bottleneck," he said. That's not a SureSwift-specific observation — industry data backs it up. One recent analysis of AI-assisted development found that AI now writes code faster than humans can review it, so the bottleneck has shifted from creation to verification, with average monthly code output per developer up sharply as AI coding tools have spread.

Gauvreau's read is the same: "It doesn't really matter what company you're in — I think this has resonated throughout the development community, that QA is the new bottleneck."

His response has been twofold: push QA earlier into the AI's own workflow rather than treat it purely as a downstream human function, and bring in a dedicated, specialized QA team (it's based in Uganda) for changes that touch deeper logic and need a more thorough, holistic pass than the portfolios' lean teams are set up to handle.

"They've been very good for us," he said, noting the team settled on assigning one QA specialist per product rather than spreading people across all three, because the teams needed people who knew one product well, not generalists.

His broader point, though, is less about any single guardrail and more about a consistent mindset and approach:

• don't extend trust in AI output faster than your ability to verify it
• internal, reversible workflows, like ticket drafting, reporting, and research, can move fast with light review
• anything customer-facing or touching production data needs a slower, more deliberate check — even if the same AI agent is doing the underlying work.

The advantages of using AI across our portfolio

One direct, quantifiable advantage of testing AI agents across three different businesses at once, rather than just one, showed up somewhere Gauvreau didn't initially expect: cost.

Because Ghost Inspector, MeetEdgar, and LeadDyno share access to pooled AI subscription plans rather than each paying for API usage independently, the portfolio's total AI spend is a fraction of what it would otherwise amount to.

"We could be using API costs, which would amount to thousands of dollars per month," Gauvreau noted, "but we end up only spending maybe $300–$600 per month instead."

That's the concrete, dollars-and-cents benefit Gauvreau pointed to directly. Beyond that, in SureSwift's own view, running the same kind of AI experimentation across several different software businesses at once has a second, less measurable advantage: pattern recognition.

A lesson learned solving a QA bottleneck in one product, or deciding how to scope database access for an AI agent in another, doesn't have to be relearned from scratch the next time a different portfolio company hits a similar wall.

AI is getting better. What does that mean for our teams?

As for what's shifted over the past year, Gauvreau's more confident than he expected to be, and less anxious about that confidence than he expected to be, too. "AI, especially when it comes to intelligence and development, seems to be continuing to improve at a pace that I wasn't really expecting. It's able to take on much bigger tasks, it's messing things up less, and I'm able to trust it a lot more," he notes.

That trend lines up with what's showing up more broadly in software engineering circles, where the conversation has shifted from whether AI can generate usable code to how teams verify what it generates fast enough to keep up.

On job retention — the question around most of the anxiety about AI agents — Gauvreau doesn't dodge it. "It's definitely replacing people's jobs," he admits. Where six developers previously worked across Ghost Inspector, MeetEdgar, and LeadDyno, there are now three, two of them part-time. But he doesn't think that's the whole story.

He reaches for a historical comparison to the automobile replacing horse-drawn transportation: an entire economy built around one technology got upended, and many people in that sector were out of work — but what replaced it, from manufacturing to maintenance to infrastructure, ultimately created more jobs than existed before, just different ones.

Still, Gauvreau acknowledges that transition isn't painless. "There will be an awkward period with a lot of transitioning, people moving around, training, doing different things," he predicts.

But his bet is that the people who treat AI as something to work alongside, not around, come out ahead of it. "If people aren't interested in using AI to augment their work, no matter what it is, I think they'll be left behind."

That's the takeaway that matters more than any specific workflow: SureSwift isn't building AI agents to remove people from the businesses it owns. It's using them to find out, business by business, which parts of the work truly need less human time, and which parts still need someone with judgment, context, and the willingness to say "no, that's not right yet" before it ships.