The Missing Human Half of AI

$100M/Month Through WhatsApp Users

At AbleCredit, our AI is used by people whose entire digital life is WhatsApp and Instagram. They are not developers. They are not power users. They are loan officers and credit analysts in small towns across India, and they use our harness and our internally-trained model to disburse $100M in loans every month.

This works because we had to solve a problem that the AI industry is not even talking about: how do you make AI usable for people who will never learn to prompt?

The Harness Consensus

The AI industry has converged on one idea in 2026: the harness matters more than the model. Karpathy called the shift from autoresearch to "agentic engineering." OpenAI published their own framing of harness engineering around Codex. Philipp Schmid at DeepMind laid out why harness engineering is replacing prompt engineering. Aakash Gupta pointed to Meta's $2B acquisition of Manus as the market signal. The Stanford/MIT Meta-Harness paper showed a 6× performance gap on the same benchmark with the same model weights, purely from harness design.

This is all real progress. The model-side harness is getting serious engineering attention, and it shows.

But the discourse is missing half the problem.

Only Half the Harness

The harness, as the industry currently defines it, is the software that wraps the model: what context it retrieves, what tools it can call, how sub-agents are orchestrated, when human approval is triggered, how failures are recovered, how state persists across calls. This is the model-side harness. It answers: how do we make the agent reliable?

There is a second question: how do we make the agent usable?

These are different questions. Reliability is about whether the agent produces correct output. Usability is about whether a human can get that output without doing work that defeats the purpose of having an agent. The first is a systems problem. The second is a human problem. The industry is solving the first and assuming the second takes care of itself.

At AbleCredit, we learned that it does not. Our users cannot prompt. They cannot spec out a task. They cannot evaluate whether the model's output is correct in the abstract. They can make a judgment call on a loan application because that is their domain, and our harness is designed to present the AI's work in a way that supports exactly that judgment, nothing more. We had to build the human-side harness from scratch because nobody had built it for us.

The Power User Blindspot

I also run a fairly elaborate personal AI setup. I burn 10M tokens/day pretty regularly (48M one weekend, cost me about $25 because I route to cheaper models automatically). I get absurd ROI from AI. And I watch my peers, who are smart, technical people, struggle to prevent AI slop.

So the UX gap exists at both ends. My AbleCredit users cannot use AI at all without a human-side harness. My tech peers can use AI, but they do all the integration work themselves, and most of them give up before getting real value. Same problem, different severity.

The UX of AI, for a non-power-user, works like this: you type a sentence into a chat window with zero context, the model gives you a generic answer, you conclude AI isn't ready. The model didn't ask you for better context. It didn't suggest an alternate approach. It didn't break the task into steps and check if you agreed with the plan. It just answered what it got, which was not enough, and the result was slop.

This is not a model failure. This is a harness failure on the human side.

What Does Human-Side Look Like?

What does a human-side harness look like? The same thing you do when you delegate a project to a person on your team.

Before starting the task, you debate the approach with them. You define what success looks like. You break it into steps and agree on the plan. You don't just hand someone a one-sentence brief and disappear.

Midway through, you check in. You look at the plan and adjust if needed. You redirect if the work is going off track. You don't wait until the entire thing is done to discover it went the wrong direction.

After the task, you review the result against the success criteria. If you make changes, those changes are logged so the person learns for next time. You don't have to re-explain the same context every single time you work together.

At AbleCredit, this is not theoretical. Our harness debates the loan structure with the officer before proceeding. It shows its reasoning. The officer makes a go/no-go judgment, not a prompt. When the officer overrides a recommendation, the system logs it and learns. This is what a human-side harness looks like in production.

Don't Make Me Think (Again)

This is how human delegation works. It is also how every successful product in history treats its users. Steve Krug wrote "Don't Make Me Think" in 2000 about web usability. The insight was simple: users scan and guess instead of reading, so design for scanning. GStack just shipped Krug-inspired UX for their AI agent skills, which is one of the few examples of anyone applying this logic to agent design. AI products in 2026 still assume users will think, type in long prompts, and be great at speccing out the task (the early days of OS and search engines had the same assumption, and they grew out of it for the same reason: most people won't).

The Model-Side Is in Good Hands

The AI industry has started building the model-side harness with real rigor. Anthropic's three-agent architecture (orchestrator, subagent, verifier) treats verification as a first-class citizen. The Meta-Harness paper shows that harness optimization can yield 6× improvements without touching the model. OpenAI's Codex harness controls scope, permissions, approval gates, and feedback loops. These are all real advances.

But look at what they optimize for: agent reliability. Agent accuracy. Agent recovery from failure. All measured on benchmarks where the task is already well-defined and the human's job is to approve or reject the output.

The human's job is never just to approve or reject. The human's job is to decide whether to engage at all, and at what level of involvement. The harness needs to support that decision, not skip past it.

AI Fails All Three

Three properties that every successful product has (and AI currently fails on all three):

1. It solves your problem.
2. You don't have to work hard to get good results.
3. It's economical.

The model-side harness is making progress on #1 (agents that actually complete tasks). It is making some progress on #3 (model routing, cost optimization). It is almost entirely ignoring #2, which is where adoption actually lives.

Overload Is Not the Problem

The Forbes piece on AI productivity last week captured the symptom but not the cause. "Employees are experiencing technology overload." No. They are experiencing tools that make the human do all the integration work. The fix is not better training or more intentionality. The fix is tools that don't require expertise to use well.

We have partial proof. $100M/month in loans, processed by people who had never used AI before, through a harness designed around what they can do (make judgment calls) rather than what they cannot (write prompts, manage context, evaluate model output in the abstract).

Two Axes, Not One

The agentic harness needs to be designed around two axes: how reliable the agent is, and how little the human has to do to get value from it. Right now the industry is building on one axis and calling it done.

The user should be able to ignore all steps and just get the result. But they must have the option to intervene, improve, and redirect. The way we have when we delegate any project to our human team members.

That is not a feature request. That is the entire product.