Human-centred tools over novelty

The question nobody is really asking

There is a question that sits underneath most of the AI product activity of 2017, and it almost never gets asked directly: who is this actually for?

Not who is paying for it. Not who is demoing it at the conference. Not which research team published the capability first. But who, in the course of a real day, with real responsibilities and real constraints, is actually better off because this thing exists?

The honest answer, for a meaningful proportion of AI products released this year, is: nobody yet. That is not a criticism of the research. The research is genuinely extraordinary. But there is a gap, wide enough to drive through, between what a system can do in a controlled demonstration and what it does for a person trying to accomplish something.

That gap matters. It is where the design problem lives. And it is, largely, where most AI product teams are not spending their time.

The shape of the 2017 AI product landscape

To be fair to the moment: the field is moving fast, and not all of the technical work is premature. There are genuine capability unlocks happening: in natural language processing, in image recognition, in reinforcement learning. The progress since 2012's ImageNet inflection point has been compounding, and 2017 in particular is producing results that would have seemed implausible even three years ago.

But there is a pattern in how these capabilities get turned into products that deserves scrutiny.

The pattern goes roughly like this: a research team achieves something impressive: a model that can answer questions, generate text, classify images with near-human accuracy. The capability is real. Then the product question becomes: what can we attach this to? What existing workflow can we slot this into, or what new interface can we build around it? The starting point is the capability, and the human need is found afterwards, or assumed.

The alternative, the more difficult, less glamorous path, is to begin with a specific human problem. To understand it precisely: what is hard, why it is hard, what someone actually needs, and what "better" would feel like for them. Then to ask whether AI helps with that, and if so, how, and in what form, and with what tradeoffs.

These two starting points produce very different products. The first tends to produce demonstrations of capability dressed up as tools. The second tends to produce things people quietly rely on.

What "human-centred" actually means in practice

The phrase has been flattened by overuse. "Human-centred design" has become a credential people apply to work that is, in practice, still technology-led. So it is worth being precise about what the distinction means when applied seriously.

Starting from the human problem means, first, accepting that the human's experience of the problem is more important than the elegance of the technical solution. A tool that partially solves a problem in a way that fits naturally into how someone already thinks and works is worth more than a tool that fully solves it in a way that requires the person to restructure their working life around the tool.

It also means being honest about the cost of learning curves. Every new interface imposes a switching cost. If a person has to invest significant time understanding how to use a new AI-powered system before they get value from it, then the system needs to deliver enough value on the other side of that investment to justify it. For most people, most of the time, the threshold is much lower than product builders assume.

It means thinking carefully about failure modes. When an AI system makes a mistake, who bears the consequence? If the person does, the bar for the system's reliability needs to reflect the cost of those failures, not the average performance in the best-case demo environment. This is particularly true in any domain that touches planning, decision-making, or execution: the areas where Mustard Seed Group's work is concentrated.

And perhaps most fundamentally, it means asking whether the tool makes the person more capable, or less so. This is the question that gets skipped most often. There is a version of AI assistance that handles so much of the cognitive work that the person's own capability gradually diminishes. There is another version that handles exactly the parts of a task that are low-value or tedious, while leaving the person's judgement, creativity and expertise intact and engaged. The difference is not always visible in a single session. It becomes visible over months.

The seduction of novelty

Novelty is not the same as usefulness. But novelty is easier to measure, easier to demonstrate, and easier to market. It generates attention: from press, from investors, from peers in the industry. This creates a systematic bias in what gets built.

The AI field in 2017 has a strong novelty gradient. The things that get covered, funded and praised tend to be the things that are most surprising: the chatbot that can hold a coherent conversation, the model that can generate a plausible news article, the system that can beat professionals at a game. These are genuine achievements. But they occupy a different category from the question of what is actually useful to people doing work.

The seduction of novelty is that it can masquerade as innovation. Building the newest thing, using the most recent model, integrating the latest capability: these feel like forward motion. And sometimes they are. But the motion only matters if it moves towards a human need that is actually being met.

This is not an argument against technical ambition. Mustard Seed Group's work, including the research happening at Benediction Lab, is deeply invested in what AI systems can become. The argument is about sequencing and orientation. The technical possibilities should expand what we can offer people. They should not determine what we decide people need.

The frame we work from

The thesis underneath everything built here is capability: increasing what people and organisations can do. Not automating people out of their own work. Not replacing judgement with probabilistic output and calling it intelligence. But genuinely extending reach: the ability to research faster, decide better, execute with more consistency, and create with less friction on the mechanical parts.

That framing has a practical consequence for how products get evaluated. The question is not whether a system uses AI. The question is whether a person is better at something after using it than they were before. And specifically: better in a way that persists, that builds, and that does not create a dependency that quietly narrows their capability over time.

This sounds simple. In practice it is a fairly demanding filter. Most new AI products, applied honestly to this test, do not pass it yet. That is not permanent: the technology is improving quickly enough that tools which are marginal today may be genuinely valuable within twelve months. But the evaluation has to be honest. Enthusiasm for a technology is not the same as evidence that it helps.

On waiting and on not waiting

There is a real tension here. If the principle is to start from human problems and work backwards to technology, then there is a risk of waiting too long: of having a perfect human-centred design methodology and a product that a competitor, less principled but more aggressive, delivers first.

That tension is genuine. The resolution is not to abandon the principle but to understand what it actually requires of the pace of work.

Starting from human problems does not mean moving slowly. It means moving in the right direction. A team that deeply understands a specific human problem and is relentlessly working towards solving it can move very fast, faster in fact than a team chasing capability releases and trying to retrofit problems onto them. The second team has to keep reorienting every time a new model comes out. The first team only has to ask: does this new capability help with what we already know we're trying to do?

That is a significant structural advantage, and it becomes more significant as the pace of AI capability development increases.

What this shapes, month by month

The work done across this portfolio is not determined by what is technically exciting this quarter. It is determined by what a specific person, a founder managing a commercial relationship, someone working through a fitness discipline, a team trying to build a product, actually needs in order to do something better.

Different audiences, different surfaces, different constraints. But the same question underneath: what does a more capable version of this person look like, and what would it take to help them get there?

That question is the centre. The technology is what we use to answer it.