I spent a lunch this week at a roundtable of Perth technical leaders. Different companies, different stages, one subject: AI. Where it belongs in the development lifecycle, and what it takes to put an AI system in front of paying customers. The word that kept surfacing was trust, and it meant something different every time.

The first version was about the output, whether that’s the coding agent opening a pull request or the model answering a customer at two in the morning. People don’t believe it produces the right answer, and the arithmetic there is unforgiving: one confident wrong answer costs more credibility than ten right ones earn. Bad answers arrive with the same fluency as good ones, so anyone who has burned an afternoon on plausible garbage walks away calibrated against the tool itself.

The second was about intent. Not whether the technology works, but whether the vendors selling it and the executives rolling it out have anyone else’s interests at heart. Where does the code pasted into a chat window end up, what does the telemetry fund, and is the enthusiastic rollout the opening act of a headcount story.

The third was about permanence. Nobody in that room wanted to rebuild a delivery pipeline, or ship a customer-facing feature, on top of a model that might cost three times as much in twelve months, or get absorbed into a competitor, or quietly stop being maintained.

Three different objections wearing the same word. They get filed under resistance, and the filing is what kills the rollout, because each one needs a different answer and enthusiasm supplies none of them.

The room isn’t an outlier

Gallup’s Q2 numbers landed in late July: 47% of US employees say their organisation has integrated AI tools, up six points in a quarter. Trust never followed that curve. Stack Overflow’s latest survey puts 45.7% of developers in the actively-distrust-the-accuracy camp against 32.7% who trust it.

On intent, an SHL survey last November found only 27% of workers fully trust their employers to use AI responsibly. No lunch-and-learn moves a number like that.

The permanence worry is the best-founded. In July 2025, Windsurf’s 350-plus enterprise customers watched OpenAI’s $3 billion bid crater, Google hire away the founders, and Cognition buy what was left, inside a single week. Nobody asked the customers.

The flagship pilot antagonises all three

The standard playbook opens on the most critical, highest-status workflow, because that’s where the ROI slide looks best. It’s also the one bet nobody with a working risk sense will take: maximum output trust demanded before any is earned, aimed at the work people are valued for, with a critical path handed to a vendor that may triple its prices or vanish.

When the pilot stalls, leadership files it under change resistance. The team’s risk assessment was better than the steering committee’s.

Start with the work nobody defends

Trust compounds in increments, so pick increments. Inside the lifecycle, aim at the undifferentiated heavy lifting: release notes, triage summaries, test scaffolding, meeting-to-ticket translation, first-draft docs, log spelunking. Nobody derives status from that work, nobody defends it in a performance review, and the failure mode is ten minutes wasted rather than a bug in production.

The same rule holds on the product side. Ship the AI feature where a wrong answer is cheap and visible before you put one anywhere near the checkout path, because the first surface you pick is a trust decision your customers get a vote on.

Leave the human’s judgement step exactly where it was. The moment where a person decides doesn’t move; the drudgery feeding it gets cheaper. People verify while verification is cheap, then stop when it starts to feel unnecessary, and that is the only honest route to output trust.

Reversible has a concrete meaning here: you could still run the workflow by hand tomorrow if the tool vanished tonight. Augmentation on those terms turns a price hike or a deprecated model into an annoyance instead of an outage, so the durability answer comes built in.

Spend the next dollar on people

Most budgets run the other way. Models, platforms, agent frameworks, evals, all funded; the humans expected to trust the output get a login and a launch announcement. Grant Thornton’s survey of 950 leaders found training is the most underfunded area of AI investment, and a telling split on who thinks that’s working: 39% of CIOs and CTOs say the workforce is fully ready to adopt AI, against 7% of COOs. The people buying the technology are five times more confident than the people running the work. Hand someone a tool they were never taught to evaluate and their distrust is competence, not obstruction.

Upskilling is the same increment, applied to the human side. Every low-stakes augmentation doubles as a training rep: the person learns what the tool is good at, where it lies, and how to check it cheaply. That earned calibration is what output trust is made of, and infrastructure spend doesn’t buy it on anyone’s behalf. It answers the intent question too, because money spent making people more capable reads differently from money spent making them more replaceable.

Watch who comes back

Time saved is the wrong first metric. It invites gaming and proves nothing about trust. Watch instead whether people opt in again next week unprompted, spot-check less over time, and start proposing the next increment themselves. Gallup found the same signal from the other end: visible manager support tracks with an 18-point higher engagement rate.

The version you can run this week: pick one task nobody would defend at a performance review, wire the tool in behind the existing judgement step, and let voluntary reuse decide whether it earns a second week. Picking that first task badly is the most common way this goes wrong, and it’s worth a conversation first.

Sources