In the first two parts of this series, we covered the systems, the goals, the context, and the process: the machinery of getting agentic automation to actually work. One thread runs through all of it: humans. Earlier generations of automation were built as if people were the problem to be resolved, and that's a big part of why they never caught up. The goal and the judgment never lived in the system. They lived in someone's head. Designing the human's place in agentic automation is what decides whether you built it right.

‍

So this one is about us; what’s more, the one crucial element that binds our work: trust.

‍

What people’s role actually is

When Stanford audited 1,500 workers across 844 tasks, the top reason they wanted automation, cited in nearly 70 percent of responses, was freeing time for higher-value work. People aren't asking to be replaced. They're asking for their time back.

That tracks with what a person is actually for. It's easy to file people under removable cost and forget what they were hired for: to understand the goal, make the judgment calls, and move the work forward with full context. Agentic automation should multiply that capacity, not shrink it, and that belief is what drives how we build. Our agents work inside the organization like any colleague: talking in Slack, cc'd on email, added to shared accounts, showing up as teammates in the tools your people already use.

The manual process was never the valuable part of anyone's job. Good automation takes the rote work people wanted to hand off anyway, and gives them room for the part that made them irreplaceable: connecting with the people behind the business, and the judgment that belongs to their seat.

‍

Trust: how it's actually earned

When asked what blocks AI adoption, most people point to cost or technology. We think the real blocker sits before either of those - it’s trust. Trust builds naturally between people who work together, but most agentic software is designed around the task, not the people meant to benefit from it. So the trust that would let an organization hand over mission-critical work never forms.

‍

Getting work done was never the whole job. Teams run on trust: you trust your teammates to make the right calls, with integrity and the purpose of the business in mind. And right now, that connection is missing between people and agents. In a 2025 study, KPMG found that 70 percent of US workers are eager for AI's benefits while 75 percent remain concerned about what could go wrong. Even the best product demo can’t close that gap if it wasn’t designed so in the first place.

‍

Where does trust come from? Two things. First, a shared belief that you're aligned on the goal. Second, doing the work in a way where you feel as good about the process as you do about the outcome. Trust gets built by doing hard things together, facing the errors and challenges that come up, handling them gracefully, and watching the other person stretch and grow along the way. Some of the work I'm most proud of with my co-founder Gautam was the hardest we've done, and it built the relationship as much as the result.

‍

We design our agents the same way: as flexible, growing systems with codified processes at their core. Automat lets you build reliable, deterministic automations you can count on. And as we roll out Workforce agents, we're constantly learning their bounds: where they succeed on their own, where they need more training, where they need guidance from our forward-deployed team or from the customer. What makes it work is that the humans alongside the agent are genuinely invested in its development, the same way you'd invest in a new teammate.

‍

Design with humans and trust in mind: Supervised autonomy that you can verify

We know that trust is earned through visibility and shared work, which changes how we design.

‍

It means real intention around onboarding an agent. On our side, a forward-deployed engineer works closely with each deployment. Think of them as a talent development lead for the agent. 

‍

There are two sides to the relationship: making the request, and giving the feedback that lets the agent learn from its mistakes and the open questions that surface as it works toward the goal. And all of it has to be legible inside the platform, so the people working with the agent can see how it's improving and learn from it too.

‍

The design goal is simple to state and hard to earn: a person managing outcomes like a manager, rather than rubber-stamping a queue.

‍

Records, trackability, reasoning, and ownership

Flip it to the other side: If you're the customer, how do you know the trust is real once agents are doing your work every day?

‍

We handle it very practically. I like to joke that we keep a CSI-grade audit trail for every action an agent takes. You can open the sessions view and watch a video of the agent moving through the interfaces, then drill into each action, each step of its reasoning, and each piece of information it read or entered. For our more agentic work, like Workforce, there are live traces of explainability: you can interrogate the agent while it works and it will tell you which tools it used and why. My favorite thing is asking it, after it does something genuinely clever, why it chose that path. The answers come back robust and specific.

‍

The research says this is the right place to spend the effort: transparency measurably improves calibrated trust in AI systems. It also carries a warning we take seriously: a smooth explanation can lull a reviewer into deferring too much. That's why the deeper safeguard is escalation by consequence rather than eloquence, which brings me to the questions below.

‍

If you're evaluating any agent vendor, us, or anyone else, here are the questions worth demanding straight answers to:

‍

Can I see why, not just what? Ask to see the agent's reasoning on a real decision it made last week. A log of clicks is a black box with a receipt printer.

‍

Is every action on the record? You should be able to hand an auditor a complete account of what the agent did on any file, with no engineering help. "We can pull that for you" means no.

‍

Does it know when to stop and ask? Small, reversible actions should proceed. Consequential ones should pause for a person. Escalate everything, and you've bought an expensive inbox. Escalate nothing, and you've bought a liability.

‍

Where is it deliberately rigid? Compliance checks and calculations should be fixed in code, giving the same answer every time, with the agent's judgment working around them. "The AI handles all of it" is a red flag, especially in regulated work.

‍

Who owns the mistake? When the agent gets something wrong, a confident vendor puts its own engineers and economics behind the fix. If failures route to a ticket queue you manage, the trust is decorative.

‍

For what it's worth, that last one is where we plant our flag: we own the outcomes. Our forward-deployed team is there to make sure the work gets done and keeps getting done, so you're never left to figure out a fancy tool alone. Anything less wouldn't be good business for us either.

‍

The through-line of this whole series comes down to one idea: automation done right gives the human a better seat. They manage outcomes, make the calls only they can make, and trust the work because they can see it.

‍

That's the standard we built Ace to meet. Ace is our digital teammate: it works inside your tools like any new hire, with the supervised autonomy and the audit trail you just read about, and our team stands behind the outcome. Ace is live now. If you want to see it run on your own work, come meet Ace.

‍