Technology Made Human: What Steve Jobs’ Greatest Product Run Can Teach the AI Age

The next breakthrough will come from making AI fit human life. Apple’s extraordinary run offers a practical design brief—and three products worth building.

Follow the AI headlines and you encounter an industry described in layers: energy, chips, data centres, models and applications. Jensen Huang sets out precisely this architecture in his essay, “AI Is a 5-Layer Cake”. It is a useful account of how the machinery fits together. Huang also explicitly identifies applications as the place where economic value is created.[1]

My question begins there. What makes an application worth using?

For the customer, this is the shop front: where the enormous investment behind the scenes becomes something understandable, desirable and useful. A company can assemble formidable engineering and still disappoint the person buying its product. The application is a crucial lever because it determines whether technical capability becomes a better experience.

But the shop front is more than its appearance. It includes what happens after the customer enters: whether the product works, whether the choices make sense, whether help exists, and whether the promised benefit survives ordinary use.

That is why I keep returning to Steve Jobs.

I am a Steve Jobs fan. I think the world lost him too early. What I regard as the greatest sustained product run in consumer technology—from the iMac through the iPod, iPhone, MacBook Air and iPad—demonstrated an unusually coherent understanding of how technology should fit around people.

That is an editorial judgment, not a measurable historical title. Nor was it one man working alone. Designers, engineers, operations teams, suppliers, developers and commercial partners made those products possible. Jobs’ distinctive contribution was insisting that their work add up to an experience with a clear purpose.

Today’s AI leaders have achieved remarkable technical progress. My challenge to them concerns the completeness of the product: can ordinary people understand its purpose, use it without continual instruction, and recognise when it deserves their trust?

Apple’s history helps us break that challenge into its constituent parts.

1. Begin with an activity people already care about.

The iPod addressed a familiar desire: carry your music and listen wherever you go. The first model combined a 5GB drive with a pocketable enclosure and automatic synchronisation. Its promise could be expressed through songs and a pocket rather than a lecture on storage architecture.[2]

Jobs’ product method, as I interpret it, began with the activity and worked backwards towards the technology. The customer did not need to become interested in miniature hard drives to appreciate the result.

AI needs equally concrete promises: understand this topic, reconcile these invoices, prepare this appointment. “Access a powerful model” describes a capability. It leaves the customer to invent the product.

2. Identify the friction that makes the activity unpleasant.

Having a large music library was useful only if people could move it onto the device and find what they wanted. Transfer speed, navigation and organisation therefore belonged to the same design problem.

The iPod’s original rotating wheel made browsing long lists practical. FireWire and automatic synchronisation reduced the work of transferring music.[2] These were responses to obstacles encountered during use.

An AI team should examine the same sequence: where does someone search repeatedly, re-enter information, lose context, struggle to compare alternatives or spend time correcting the system? Those points are the product opportunity. More generated text may leave every one of them intact.

3. Give the activity its own controls.

The iPhone brought calling, music and internet functions into one device, with a large multi-touch surface whose controls could change with the application.[3] A photograph, a keyboard and a web page did not have to compete for the same permanent physical buttons.

AI should extend that principle. Comparing policies calls for aligned clauses. Planning calls for dates and dependencies. Investigating a claim calls for evidence beside the claim. People should be able to select, rearrange and correct these objects directly.

Conversation is a valuable entrance. It need not be the room in which every task takes place.

4. Allocate complexity to the component best equipped to handle it.

iTunes handled the library on a computer; the iPod concentrated on portable listening. Synchronisation connected the two.[2] Simplicity on the small device depended on work done elsewhere.

For AI, the equivalent is a deliberate division of labour. A language model can interpret a question and propose an analytical approach. A database should retrieve the records. A calculation engine should compute the result. The interface should expose the assumptions and let the person correct them.

Asking a model to generate a plausible number when a calculator can establish it is poor product architecture. Integration means choosing the right component for each responsibility.

5. Make the whole journey work.

The 2003 iTunes Music Store extended the experience into finding and purchasing music, supported by agreements with the major record companies.[4] The device became part of a connected commercial system.

AI needs this completeness. A report assistant must help identify authorised sources, analyse them, assemble a reviewable document, preserve it and hand it into the user’s existing workflow. If the customer must rebuild the process across six disconnected tools, much of the integration work has been transferred to them.

The application layer therefore includes permissions, data connections, delivery and support. A beautiful interface resting on unreliable handoffs is an unfinished product.

6. Balance competing requirements around the person.

The MacBook Air pursued portability while retaining a full-size keyboard and a 13.3-inch display.[5] Its compromises made the intended activity clear: carry a usable computer more easily.

AI products also face competing requirements: quality, response time, cost, privacy, battery life and review effort. Maximising one can damage another. A slow, expensive model is wasteful for a trivial transformation; a fast answer is inadequate when a consequential decision requires careful checking.

The product should make sensible defaults and explain meaningful trade-offs. Its user should not need a technical education merely to choose how thoroughly an invoice is checked.

7. Treat aesthetics as part of usability.

Apple’s best products made hierarchy, proportion, typography, physical form and responsiveness feel considered together. Their appeal involved both pleasure and comprehension.

This is the part of Jobs’ legacy I particularly miss. Design communicates what matters and makes people willing to engage. A crowded AI workspace can exhaust attention before the useful work begins.

Yet AI introduces a complication: polish can make weak evidence look convincing. The appropriate aesthetic must make uncertainty legible. A clear distinction between a source, an interpretation and an unresolved question belongs inside the visual design.

8. Choose the form that suits the setting.

The iPad offered a larger touch surface for reading, browsing, watching and interacting with applications, alongside software adapted to it.[6] Different circumstances justified a different form.

Likewise, a researcher may need a large screen for comparing sources; a technician may need a rugged tablet; someone whose hands are occupied may benefit from voice. Accessibility belongs in this decision from the beginning.

An AI device must earn its place through the activity it improves. An existing phone or laptop may already be the right hardware. Another object to charge and carry requires a compelling reason.

9. Build continuity beyond the first purchase.

The iPad launched into an existing world of applications and content.[6] Its value included what people could do after switching it on, and what developers could subsequently add.

AI products need persistent projects, dependable connections, understandable pricing, maintenance and support. Their usefulness must survive a model update or an interrupted session.

Here we should improve on the costs of closed ecosystems: users should be able to export their work, inspect stored preferences and change providers. Continuity should make them more capable, not dependent on an inaccessible history.

10. Demonstrate the benefit and deliver it repeatedly.

A product promise should be visible in an ordinary task. How quickly can someone find the song, read the page or complete the work? My reading of Apple’s strongest launches is that the demonstration made the benefit concrete.

For AI, the demonstration must include realistic exceptions. Show the missing document, the ambiguous instruction and the failed connection. Show how the person corrects the system and resumes.

A spectacular first answer earns attention. Dependable use earns a place in someone’s life.

These principles translate into a practical design brief:

There is, however, an eleventh requirement that changes the brief.

Selecting a recording on an iPod did not normally cause it to invent a plausible replacement. Generative AI can operate as designed and still produce a false statement. NIST identifies this problem as confabulation in its Generative AI Profile.[7]

Earlier software also had bugs and uncertain outputs. What matters here is how frequently an ordinary user must judge generated content that looks finished.

Creative variation, incomplete evidence and factual error need different treatment. Ten alternative illustrations may be welcome. Ten incompatible invoice totals are not. Repeatability is also distinct from correctness: repeating a mistake does not improve it.

Therefore, AI’s design brief must include evidence, correction and limits on action. A draft, a retrieved statement, a computed result and a completed transaction should have visibly different states. Microsoft’s human–AI interaction research likewise stresses communicating capabilities and enabling correction.[8]

With that addition, what should we build? I would propose three products. These are design concepts, not claims that their individual features are new.

Product one: the Thinking Desk.

Its promise: “Turn a difficult question into something you can understand and inspect.”

It opens onto your projects. Each contains the question, relevant sources, assumptions, working materials and unresolved issues. Its primary surface is an editable canvas, with conversation available alongside it.

Imagine comparing two proposals for your business. The Desk aligns their terms, calculates costs using visible assumptions and identifies missing information. Select a conclusion and its supporting passages appear. Change an assumption and the affected calculations update, while earlier versions remain available.

Three simple controls organise the work: Explore, Learn and Check. Explore develops alternatives. Learn offers examples and graduated hints. Check tests claims against sources and suitable tools, reporting what remains unresolved. None should imply that pressing a button guarantees truth.

The Apple inheritance is the iPod’s navigability combined with the iPhone’s adaptable controls: many materials become manageable through an interface fitted to the task.

At Fahm.uk, I have been developing visual thinking and the idea of AI as a thinking partner. That is the direction I want to pursue: an environment that helps people reason more clearly. “Partner” here describes interaction, not consciousness or human judgment.

Parts of this market already exist. Google’s NotebookLM introduced interactive mind maps and source-based study tools, and its notebook offering continues to emphasise grounded research.[9] The opportunity is to integrate evidence, editable assumptions, calculation, learning and continuity so well that the whole activity becomes easier.

Measure success through completed tasks, correction time and the user’s ability to explain the conclusion independently. For learning, test later recall and unaided performance. More conversation is an inadequate measure of better thinking.

Product two: the Business Desk.

Its promise: “Bring me the exceptions, with the evidence needed to resolve them.”

Start with a narrow customer and workflow: a small business processing supplier invoices. The home screen shows items ready for review, missing information and discrepancies requiring attention.

Open an invoice and see its extracted fields beside the original document. The system matches the purchase order, checks arithmetic and highlights differences. Correct a field directly. Prepare an accounting entry or supplier query from the same view.

Before submission, the product shows exactly what will change, where it will be sent and who authorised it. Payment authority remains separately controlled. A failed connection produces a clear pending state; it must not quietly create a duplicate entry when retried.

The Apple inheritance is the complete iPod–iTunes journey: the customer purchases a functioning process, including the difficult handoffs.

Invoice automation is already a real category. Microsoft’s Dynamics 365 Invoice Capture creates vendor invoices from digital images and supports review and correction.[10] The opportunity is therefore not to announce that AI can read invoices. It is to make deployment, cross-system operation, exception handling, pricing and accountability straightforward for a clearly defined customer.

Measure handling time including review, error rates, duplicate prevention, onboarding effort and cost per completed invoice. Compare against existing accounting tools and rules-based automation. An AI badge is no reason to replace a cheaper process that already works.

Product three: the Household Desk.

Its promise: “Keep everyday administration organised, with you in control.”

This product begins on devices people already own. It connects selected calendars, messages and documents with explicit permission. Its home screen contains commitments, approaching deadlines and decisions awaiting approval.

A school message becomes a proposed calendar entry with the original message attached. A renewal notice becomes a dated comparison of the current arrangement and available alternatives, with missing terms marked. An appointment becomes a preparation card containing the time, location and documents needed.

The crucial design feature is the boundary between organising and acting. Reading an email does not confer permission to send one. Preparing a booking does not confer permission to spend. Sensitive material is shared only with the intended household member; joint planning must not erase individual privacy.

Every action has a receipt: what happened, when, under whose authority and whether it can be reversed. Cancellation and correction are normal controls. Where reversal is impossible, the product must say so before proceeding.

The Apple inheritance is portability and continuity: one intelligible view of activities that currently cross several interfaces. A dedicated household display might eventually help, but useful software should establish the need first.

Calendar, reminder and assistant products already cover parts of this territory. The proposed gap is an experience combining family permissions, source traceability, dependable handoffs and understandable recovery. That is a design opportunity to validate with households, not proof that no competitor offers it.

Measure missed commitments, time spent on administration, mistaken actions and how easily people can understand and revoke permissions. A household assistant should earn trust through restrained usefulness.

What, then, is missing from the market?

My assessment is that the opportunity lies in completeness. The examples above show that sources, diagrams and automation are already available. Claiming that nobody is working on human-centred AI would be unfair and inaccurate.

What deserves more attention is the combination: a clear activity, an appropriate interface, persistent context, inspectable evidence, restricted authority, recovery and a price that makes the whole service worthwhile. Availability of the parts does not establish the quality of the experience.

This is a product thesis, not an exhaustive market census. The test is practical: can a defined group of users complete important work with less effort and fewer consequential errors than their current method? At Erasys Ltd, we have already launched products such as ClearSense, which make working with AI more reliable and ClearSeal, which provides a privacy layer and security layer currently lacking in AI workflows and AI labs don’t have a vested interest in getting it there because part of their business model is to be able to ingest as much data as possible.

The companies building these products should begin with observation, establish a baseline and test difficult cases alongside ordinary ones. Expand only when the whole workflow improves. Model upgrades must be evaluated against that promise, because changing a component can change the experience.

That is the lesson I take from Jobs’ extraordinary run. He and Apple’s teams made interacting with technology feel considered from beginning to end. The engineering, interface and commercial system served an activity a person could recognise.

The AI age requires the same ambition, with a further obligation: make the limits of the machine understandable.

The next great product should leave its user with clearer understanding, completed work and more room for life. That is what it would mean to make AI human-centred.

Sources

1. Jensen Huang, AI Is a 5-Layer Cake, NVIDIA, 10 March 2026.
2. Apple, Apple Presents iPod, 23 October 2001.
3. Apple, Apple Reinvents the Phone with iPhone, 9 January 2007.
4. Apple, Apple Launches the iTunes Music Store, 28 April 2003.
5. Apple, Apple Introduces MacBook Air, 15 January 2008.
6. Apple, Apple Launches iPad, 27 January 2010.
7. NIST, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, 2024.
8. Amershi et al., Guidelines for Human–AI Interaction, CHI 2019.
9. Google, NotebookLM launches features to help students study; current notebook product.
10. Microsoft Learn, Invoice Capture solution overview.

Market discussion checked on 20 October 2026. Product concepts and assessments are the author’s proposals; vendor documentation establishes advertised capabilities, not independently verified outcomes.