We have a responsibility to shape how we leverage AI in the workplace.

That responsibility isn’t abstract. Every time we introduce a new tool, automate a workflow, or hand a task to an algorithm, we’re making a decision about how people will work, learn, and grow inside our organizations. AI is forcing that decision into the open faster than any technology we’ve worked with before, and the organizations we serve – nonprofits, public sector agencies, mission-driven teams – cannot afford to get it wrong. Their missions depend on the judgment, creativity, and trust of the people doing the work, not just the speed at which the work gets done.

Most conversations about AI implementation start and end with efficiency. How much faster can we process this. How many hours can we cut. How much can we reduce headcount or reallocate staff time. These are legitimate questions, but when they’re the only questions, we end up optimizing for the wrong outcome.

Technical problem-solving is difficult, messy work. It resists shortcuts. When we design AI implementations purely around speed and automation, we quietly remove the friction that makes that work valuable in the first place – and we remove it for everyone, not just the people doing routine tasks.

That friction matters more than most technology vendors want to admit. The obstacles, dead ends, and moments of confusion that come with solving a hard problem are often exactly where our best ideas surface. They’re where we’re forced to ask better questions, reconsider our assumptions, and notice things we would have missed if the answer had simply been handed to us. Strip that friction away entirely and we lose the conditions that produce insight.

I’ve learned far more in my career from failed projects, mistakes, and errors than I ever have from quick wins or rote tasks. That’s not a comfortable thing to admit, but it’s true for almost anyone who has built something worth building. The projects that didn’t go as planned taught me how systems actually behave under pressure, how teams communicate when things go sideways, and how to rebuild trust with a client after a setback. None of that shows up in a project plan. It shows up in the mess.

If AI is deployed in a way that eliminates messes – and the discomfort, iteration, and problem-solving that come with them – we risk raising a generation of professionals who are excellent at supervising outputs and much less practiced at the underlying judgment that produces good outputs in the first place. That’s a real cost, even if it never appears on a balance sheet.

This is the tension at the center of the framework in the graphic above: conventional AI implementation optimizes for speed and bias toward automation, measures success by ROI and headcount reduction, and treats machines as the dominant actor in a process. Human-centered AI asks a different set of questions. It prioritizes usability and trust over raw velocity. It measures success by whether real problems get solved for real people, with staff and constituents helping define both the problem and the desired outcome. And it keeps people in the loop throughout, not as a supervisory afterthought, but as a design requirement, because the work of implementation is iterative by nature.

At Craftsman Technology Group, we’re building an internal framework that governs how we use AI in our own work before we ever recommend it to a client. We don’t believe in prescribing a human-centered approach to AI that we haven’t tested against our own processes, mistakes, and our clients’ expectations of us.

Concretely, that means asking a consistent set of questions before any AI tool touches a workflow: Who defines the problem we’re trying to solve – is it us, or the people closest to the work? Does this implementation give staff more room to exercise judgment, or does it quietly narrow their role to reviewing what a model produced? Are we measuring success by time saved alone, or by whether the people affected – staff, constituents, community members – actually trust the result and find it usable? Have we built iteration and improvement into the process in a way that allows us to continually assess the trajectory of the technology and any associated risks?

These aren’t rhetorical questions. They shape which AI tools we recommend, how we scope implementation timelines, and where we insist on keeping a human decision point even when a fully automated path is technically available. For organizations built around a public mission, that decision point is often the whole point of the work – the moment where professional judgment, community context, and institutional values get applied to a technical output. Automating past that moment doesn’t just risk a bad outcome; it risks eroding the reason people trusted the organization to begin with. Some may see this as unnecessary friction, we see this as good friction.

Clients come to us skeptical of AI, and often for good reason. They’ve seen technology vendors oversell automation as a replacement for expertise rather than a complement to it. They worry, reasonably, about what gets lost – in judgment, in relationships, in accountability – when decisions move from people to systems. A framework like this is our answer to that skepticism: not a promise that AI is risk-free, but a commitment to implementing it in a way that keeps people, purpose, and problem-solving at the center.

Our technology work should unlock creativity, enable critical thinking, and help develop the next generation of leaders inside the organizations we serve – not quietly replace the conditions that produce those things. That’s a higher bar than efficiency alone, and it’s the one we’re holding ourselves to.

These ideas aren’t a finished product. They’re a working framework, and we expect it to change as we learn more and the technology improves. Getting AI implementation right, for ourselves and for our clients, is iterative work. It’s the kind of messy, human-centered problem-solving we believe in. And it’s how we intend to keep earning the role of trusted, long-term partner, one implementation at a time.

This piece was created with AI assistance; the ideas and final work are our own. See: how we use AI in our work.