OP-ED - Cue ominous film score: is the AI agent apocalypse upon us?
Last week, an AI startup watched its entire value proposition ship as a native feature. This is the story of what that moment reveals — and what some are betting it doesn't.
When Anthropic released Claude Cowork last week, the team behind Eigent publicly admitted their startup was dead and open-sourced the product. Within 48 hours, the post had 1.6 million views.
Some context... Claude Cowork is Anthropic's new desktop automation feature, available to subscribers paying USD 100 to USD 200 per month. It lets Claude take control of your computer to perform multi-step tasks like organising files, conducting web research, and generating reports.
Eigent, founded by Oxford AI researcher Guohao Li, was building essentially the same thing as a standalone product. A multi-agent system where specialised AI workers collaborate to automate complex workflows on your machine. Local-first, privacy-focused, open-source roots. The kind of product that looks impressively differentiated right up until the model provider ships it natively.
Li called open-sourcing Eigent "the most rational response" to being erased. The AI Secret newsletter couched it more starkly: "Limited innovation is no longer a viable hedge. If a startup is not creating a new category or redefining the problem, it is exposed by design."
The post captured a fear that has been simmering across the AI agent startup landscape: the model providers will eventually ship what you're building, for free, with superior distribution.
But not everyone is running for the exits.
A US-based, Ghana-built “no-code data platform” called Papermap is watching the same collapse and drawing the opposite conclusion. Their calculated wager: the giants are playing the wrong game entirely.
Papermap was founded in 2025 by Isaac Sarfo and Benedict Quartey, both Ghanaian. Sarfo has a background in physics and data science, spent time in venture capital, and is now based in Brooklyn, New York City. Quartey is a PhD candidate in AI and robotics at Brown University.
The company is currently in pilot with several Series A African startups across logistics, consumer goods, and direct-to-consumer verticals.
Task versus purpose
NVIDIA CEO Jensen Huang frames the future of work as a distinction between tasks and purpose. Tasks are routine, automatable actions. Purpose is the high-level, creative goal of a profession. His advice to developers: focus on purpose; let AI handle the tasks.
By that token, Eigent died because it was automating tasks: thin orchestrations of capabilities Claude already had. The moment Anthropic shipped Cowork, Eigent's value proposition evaporated.
Sarfo, writing on his Medium blog, is betting on purpose. His argument: Silicon Valley is obsessed with Software Code Generation (tools like Lovable and Cursor that help engineers build apps faster). But the real GDP multiplier is not accelerating how fast you build software. It is accelerating how fast businesses make decisions with their own data.
"The 'Real Economy' (the 50% of GDP comprised of SMBs and Mid-Market firms) is not constrained by a lack of software," Sarfo writes. "They are drowning in software. They are constrained by a lack of Decision Velocity."
Full disclosure: I am the “veteran journalist” Sarfo references in that piece. When he first pitched me on "Data Democratisation" as a manifesto, I told him it would not land. The framing was too abstract, too academic. What he has arrived at since (Decision Velocity, Data Code Generation) is sharper. How well it converts in the wild remains to be seen.
To demonstrate the principle, consider Lovable, which reportedly hit USD 200 million annual recurring revenue (ARR) and a USD 2 billion valuation in record time (current valuation: USD 6.6 billion). They help you build a logistics app.
Meanwhile, Papermap helps a logistics company route 500 shipments in three minutes instead of three days by instantly analysing traffic, fuel costs, and driver availability from data they already own. One automates a task. The other enables a purpose.
The wrapper question
Huang describes the AI stack as a five-layer cake: Energy, Chips, Infrastructure, Models, Applications. His framework emphasises that controlling all five layers is essential for building a scalable AI stack.
This is the logic behind the notion of "sovereign AI"—the idea that every nation must own the production of their own intelligence, from data centres down to cultural data codified into LLMs.
Quartey sees it differently. "AI is eventually going to become a commodity," he contended during a live demo of Papermap’s value proposition. "And once that happens, the only thing that truly has value is what you do with it."
His evidence: OpenAI's biggest business is not the model API; it is ChatGPT, the consumer app with 800 million weekly active users. Cursor is not building its own LLM; it is building a system that works reliably on top of many.
Huang himself notes there are now over 1.5 million active AI models worldwide. In a world of abundant models, the value shifts to whoever orchestrates them into real outcomes.
"Everybody is a wrapper of something," Quartey argues. "NVIDIA is a wrapper of semiconductors. Semiconductors are a wrapper of physics. The question is whether you are wrapping at the right layer."
For Papermap, the right layer is the interface between chaotic business data and human decision-making. In Huang's four-stage evolution of AI (Seeing, Creating, Doing, Acting), they are building for Doing: agentic AI that reasons, plans, and uses tools to manage complete workflows.
Their platform connects directly to databases (think accounting software, inventory systems, point-of-sale, payroll) and generates SQL and Python code just-in-time. No data warehouse. No Fivetran to automate moving data from hundreds of different applications and databases into a central cloud warehouse for analysis. No data engineer. Just the question and the answer.
Andile Masuku
The proving ground
The companies Papermap currently claims to serve (among them, a fulfilment company in Ghana and a fintech in Nigeria) are not early adopters by choice. They are the "Missing Middle" that the Modern Data Stack never reached.
They could not afford subscribing to a sophisticated cloud-based data platform like Snowflake to enable the storage, analysis, and sharing of massive amounts of data in real-time. They were not in a position to hire data engineers or wait two weeks for a dashboard that would be stale by the time it shipped.
When you build for resource constraint, you build differently. You do not assume the client has a data team. You do not assume clean schemas. You build for chaos. And in doing so, you build something that scales into markets drowning in software but starving for decision velocity.
The first outside investor to acknowledge the merits of this approach was Jeff Dean: Google's chief scientist, co-creator of TensorFlow (the open-source machine learning framework that underpins much of modern AI development), and co-founder of Google Brain (the deep learning research division that merged with DeepMind in 2023).
Dean has publicly championed AI development in Africa since helping to establish Google AI Ghana in 2018. Apparently, he wrote Papermap a cheque within a week of trying the product, off the back of a prior relationship rather than a formal due diligence process.
What survives
The Eigent story is a warning, but the warning appears specific. It applies to startups automating tasks that model providers can ship natively. It doesn’t appear to apply to startups enabling purposes that giants have no incentive to serve.
Giants want to be the platform, not the specialised tool. Well, at least, until they change their minds. Generally, though, the "cost of focus" is often too high for a trillion-dollar company.
Furthermore, they operate under anti-competitive regulatory constraints, which can offer protective cover for well-pitched and positioned upstarts operating within regulated spaces.
It doesn’t seem Google is coming for the third-party logistics operator in New Jersey any time soon, and Anthropic doesn’t appear to be building for the regional fintech in Lagos. Microsoft is certainly not optimising for the CFO who needs "Revenue" to mean GAAP Recognised whilst the VP/Director of Sales needs it to mean Bookings.
That is the gap. That is what Papermap’s founders and their early backers are betting survives the agent layer cull.
Editorial Note: A version of this opinion editorial was first published by Business Report on 20 January 2026.