Tools, not gods.

I've worked with AI tools for over a decade. They are powerful. In the right application, they're the difference between a project that ships and one that doesn't.

I'm still not buying the full pitch.

Most AI discourse right now is either religious or apocalyptic.

Every year the story escalates. AGI by Christmas. The death of software engineering. Autonomous agents that run your business while you sleep. The framing keeps getting louder; the underlying tools keep doing more or less the same thing, slightly better than last quarter.

From where I sit (running a technical consultancy, shipping AI work for clients every week), the reality is more grounded, less glamorous, and far more useful than the loudest version of the story.

The pattern repeats. And every time, we lose the tool.

We've seen this movie. More than once. #

Big data was going to let every company act on the same insight that powered the platforms. It mostly didn't. The infrastructure consolidated inside cloud vendors and a handful of data brokers, and most companies are still drowning in dashboards they don't trust. The metaverse was going to be a shared, open digital world. It became a Meta product nobody asked for, and serious work on shared virtual environments got defunded in the backlash. Blockchain was going to be decentralized verification of state. It became NFT grifts and exchange collapses, and the useful pieces got buried under the wreckage. The Segway was supposed to rewire cities. It became a mall-cop joke, and the mobility revolution had to wait a decade for shared scooters and e-bikes to relearn the idea.

Every cycle works the same way. The hype machine promises something close to magic. The execution misses the promise. The backlash hits. And here's the part that doesn't get talked about: the tech itself gets contaminated by the hype, and the useful version of it gets harder to do because the room has lost patience.

The tech gets contaminated by the hype.

That's the real cost of overhype. The technology gets stolen by the people selling the fantasy. The grounded, useful version (the one that actually shipped value for real businesses) gets thrown out with the speculative one when the cycle turns. Companies stop investing. Practitioners learn to mock the tool instead of building with it. Infrastructure work gets defunded. By the time the dust clears, the parts that worked have to be rediscovered.

AI is in that window now. The hype is wrong about the timeline and wrong about what it will replace. The underlying tech is still real and still useful. The question is whether enough of us hold the line on the grounded version long enough for it to become infrastructure, instead of letting the disillusionment phase take the useful parts with it.

The honest disruption. #

AI is revolutionary in some ways, but it is still fundamentally an evolution of automation rather than the arrival of an artificial mind.

Entry-level work that can be reduced to predictable patterns is being automated. Basic content creation, support scripts, rudimentary code, repetitive document processing. If a job is mostly predictable, it's at risk.

Mid and senior roles aren't being replaced. They're being reshaped. A skilled strategist using AI is faster than one who isn't. Same goes for developers, marketers, designers, researchers. The right mental model isn't "AI replaces the worker." It's "AI is a fast assistant that never sleeps, gets the texture wrong, and needs supervision."

Productivity gains are real, and they go disproportionately to the people who know how to work with these tools, not just the people who have access to them. That gap widens every quarter.

We're not close to AGI. The agents aren't either. #

What we have today is pattern recognition operating at a scale that imitates intelligence convincingly. That isn't nothing. It's also not the same thing.

The systems we're building have no awareness, no intent, no goals of their own. They don't reason the way humans do. They don't learn across unrelated domains the way a child does. They don't understand what they say. They generate the most statistically plausible continuation of the text you gave them, and a lot of the time that continuation is useful.

The agent story is the 2026 version of the same overclaim. Bounded agents (the kind that operate inside a defined toolset against a specific goal) work. They do real work. Unbounded agents, the kind that supposedly run your business autonomously, are still the same plausible-text machine wearing a different hat. Give one too much rope and it will confidently spend your budget on the wrong thing in a syntactically perfect way.

Artificial general intelligence, as it's classically defined (an autonomous, self-aware mind capable of operating across arbitrary domains), isn't sitting one product cycle away. The fact that today's tools are more capable than last year's tools doesn't mean we're closer. It means today's tools are more capable. Those are different claims.

This isn't a popular position in some rooms. It's also the correct one.

Where AI actually earns its keep. #

The honest list, in 2026:

• Drafting and rewriting structured content (proposals, summaries, outreach, internal documentation)
• Pulling structured data out of unstructured documents at scale
• Translating natural-language intent into search queries, code, configuration, or queries against your own data
• Summarizing long material into different lengths and registers for different audiences
• Acting as a thinking partner against a body of context someone else has already curated
• Running narrow agents that complete bounded multi-step tasks with tool access and human checkpoints
• Surfacing patterns across qualitative data that would take humans weeks to read through

That's a meaningful list. None of those items require AGI. All of them have real business value when applied to the right workflow.

We're working with a research firm right now on the last item. Their qualitative methodology depends on nuanced human interpretation of interview transcripts. We built tooling that processes transcripts automatically, surfaces emotional and symbolic markers, and organizes early insights into structured outputs for analyst review.

The AI doesn't replace the analyst. It gives them a head start. Hours of grunt work compress into minutes, and the analyst spends their time interpreting rather than processing. That's the shape of the win.

Amplification, not autonomy.

The economics nobody talks about. #

There's an economics conversation we mostly skip past. Every model call costs something. Tokens in, tokens out, charged per million. At consumer scale it's invisible; at production scale it shows up fast. A workflow that hits a frontier model on every user action can quietly burn through a developer's salary in compute by year-end. Per-call cost is a design constraint, not a footnote.

The pricing model creates its own incentives. Per-token billing rewards verbosity. Longer outputs make more money for whoever's selling the model. Nothing pushes the other direction unless the buyer is paying attention. The metric that matters in production is tokens per useful answer, not total spend.

Measure tokens per useful answer.

Then there's the energy. Training the largest models is one cost; running them at the scale we're running them now is another. Inference at hyperscale draws meaningful grid load, and the data center buildout has become a real political conversation in the regions where it's happening. None of this stops being true because the abstraction layer is friendly. Every "ask the AI" button has a power draw behind it, and the aggregate is not small.

The implication for how you deploy is simple. Use AI where the answer it gives you is worth more than the call it cost. Skip it where it isn't. Route the small jobs to small models. That sounds obvious. Almost no one designing AI workflows in 2026 is doing the math.

The frontiers that actually matter. #

If you want to watch for breakthroughs that would actually change the shape of this, the ones to track are structural, not parametric. Bigger models alone will not get us there.

In rough order of how much each would change the calculus:

• Agents that can be trusted with consequential, irreversible decisions, not just bounded research and retrieval
• Durable memory across sessions that doesn't quietly hallucinate state
• Interpretability tooling that lets operators verify what a model is actually doing before they ship its output
• Honest evaluations of agent behavior in the real world, not just benchmarks built by the same people selling the agents
• Efficiency gains that drop the cost-per-useful-answer by an order of magnitude without sacrificing reliability
• Embodied systems that handle physical-world feedback loops without falling over

Until those land reliably, we are in the realm of specialized automation that gets steadily more capable. That's a fine place to do real work. It is not sentient software, and the gap between "useful tool" and "trusted autonomous operator" is wider than the marketing suggests.

Tools, not gods. #

At Joust we use AI heavily. We also critique it, because tools should be understood before they're adopted or feared. These tools are genuinely useful when applied with discipline.

The constraint on most teams has never been the supply of ideas. It's time, budget, and bandwidth. The work that needs doing has always exceeded the hours available to do it. What these tools genuinely unlock is the ability to put more of those hours into solving real problems and less into the repetitive scaffolding that surrounds them.

That's why we built the AI Operations Roadmap. Not a transformation pitch, not a platform sale. A focused audit that finds the workflows in your business where AI actually fits, sizes the real-dollar opportunity, and gives you a sequence to execute. The recoverable manual work is usually hiding in plain sight, and it usually adds up to six figures.

Businesses that treat AI as infrastructure instead of ideology will outperform the ones chasing hype.

Stay grounded. Stay curious. Stop pretending autocomplete got smart enough to qualify as a person.

If a digital mind shows up tomorrow in a mountain bunker, I'll update my priors. Until then, I'm sticking with the workflows and the tools we actually have in front of us.

Tools, not gods.