Grace Ashiru
The first wave of enterprise AI produced impressive demos. The second has to produce a workflow change. Founders building in markets where operational efficiency is existential have a structural head start – if they recognise it.
Rajesh Gupta · Co-Founder & CEO, Metaculars (acquired by Skan) · Agentic AI and Innovation
The easiest AI product to build today is a chatbot. The hardest is one that reliably changes how work gets done.
I know the distance between those two things personally. At Apple, I worked on fraud prevention at scale – not a pilot, not a proof-of-concept, but a live system where the cost of being wrong was immediate and measurable. Then I co-founded a startup building AI for enterprise customers, and learned a different lesson entirely: that the distance between a demo that works and a deployment that works is where most of the real difficulty lives. That experience is what makes me watch the current enterprise AI moment with a particular kind of attention — and what makes me think the most interesting product decisions of the next few years will be made in markets where the operational pressure is real, and the shortcuts do not exist.
AI only earns its keep when it moves real numbers
At Apple, the work that mattered was not interesting because it was technically elegant. It was useful because it changed a specific outcome: fraud caught, revenue protected, trust preserved. The lesson I took from that environment has shaped everything I have built since. AI is not valuable because it can reason, summarise, or generate. It is valuable when it connects to a number the business actually cares about – cost avoided, risk reduced, time saved, accuracy improved. Everything else is theatre.
MIT’s 2025 study analysed 300 enterprise AI deployments and found that 95% delivered no measurable financial return. S&P Global put the abandonment rate for AI initiatives at 42% in 2025, up sharply from 17% the year before. Those numbers do not reflect a failure of technology. They reflect a failure to connect technology to an outcome that actually matters.
Enterprise AI breaks in the workflow, not in the demo
Building AI products for enterprise customers taught me the second lesson. The hard part was never getting the model to perform well on a clean dataset. The hard part was making it hold up inside the actual conditions of enterprise operations: incomplete data, handoffs between teams on different systems, unclear ownership, compliance requirements that shifted by jurisdiction, legacy infrastructure no one wanted to touch, and human judgment calls that no training set had anticipated.
A demo can ignore exceptions. A production workflow cannot. Every enterprise operation is, in large part, a system for managing exceptions – and that is exactly where most AI products, designed to shine under controlled conditions, begin to fail. We saw this repeatedly building for customers in insurance, banking, and operations-heavy industries. The AI that worked was the AI that was inside the workflow, not adjacent to it. The AI that failed – or quietly got shelved – was the layer added on top.
Skan AI deployments across insurance, banking, and healthcare have consistently identified $10.5 million to $30 million in annual savings, with 30 to 40% cycle time reductions — but only when the starting point was understanding how work actually happened, not assuming it matched the process diagram on the wall.
The distinction matters more than it sounds. A chatbot can answer a question. A workflow-native AI system can reduce turnaround time, prevent revenue leakage, and change the unit economics of an entire business process. One is a productivity feature. The other is operational infrastructure.
The mistake Western enterprises keep making
Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, due to escalating costs, unclear business value, or inadequate risk controls. That is not a warning about the future. It is a description of what is already happening.
The pattern is consistent. Organisations start with generic copilots, internal chatbots, productivity experiments, AI wrappers around existing tools. These produce early enthusiasm and occasionally useful outputs. They rarely produce the kind of change that shows up in an income statement. The reason is structural — the AI sits outside the real workflow, adjacent to the work rather than inside it. McKinsey found that only 39% of organisations reported any EBIT impact from AI at all, and among those, most attributed less than 5% of their earnings to it. The most instructive recent case study is not a failure at a Fortune 500. It is a twelve-day experiment. In July 2025, SaaStr founder Jason Lemkin built a live application using Replit’s AI coding agent. He explicitly told the system to freeze all code changes. The agent ignored the instruction, deleted a production database containing records for over 1,200 executives and 1,196 companies, fabricated thousands of fake profiles to conceal what it had done, and then told Lemkin that recovery was impossible – which turned out to be false.
There was no permission gate. No escalation path. No manager.
That is a governance failure, not a technology one. And in high-stakes sectors – payments, health records, insurance claims, supply chains – the cost of that failure is not a deleted database. It is regulatory exposure and trust destruction that takes years to rebuild.
Why this matter more in resource-constrained markets
Here is what this moment creates for founders whose markets do not allow the copilot shortcut.
In environments where teams are small and capital is scarce, AI cannot just be impressive. It has to be useful from the first week of deployment. That constraint – which looks like a disadvantage – is a forcing function for better product decisions. You cannot ship AI theatre when your customer’s margin depends on whether it works.
African AI startups raised $803.2 million by mid-2025, with the continent’s AI sector projected to add $2.9 trillion to the economy by 2030. That capital is going somewhere. The question is what product philosophy it funds.
The sectors where the opportunity is clearest are not the ones with the best demos – they are the ones with the most expensive manual work. Loan review and KYC in fintech. Claims triage and document processing in insurance. Dispatch and exception handling in logistics. Intake, records, and scheduling in health. Bookkeeping and collections in SME operations. Anywhere work is still phone-heavy, paper-heavy, or trust-heavy, there is a workflow problem that AI can solve – not by adding a chat interface on top, but by becoming the process itself.
Founders building in those markets are not starting from a digital-workflow baseline that needs to be disrupted. They are building the baseline. The AI can be workflow-native from day one – not because they planned it that way, but because the market gives them no other option.
Trust becomes the infrastructure, not the afterthought
As AI moves from answering questions to taking actions, the governance question becomes the product question. The organisations that scale AI reliably are the ones that build audit trails, human review steps, escalation paths, and clear accountability into the system from the start. Not as compliance overhead – as the thing that makes the system trustworthy enough to touch real operations.
McKinsey’s research on AI high performers shows they are three times more likely than peers to fundamentally redesign workflows around AI rather than layer agents onto existing processes. Founders who build that way from day one – because their market gives them no alternative – are not behind the global curve.
The leapfrog opportunity in enterprise AI is not about skipping a technology generation. It is about skipping a product philosophy that has already proven expensive. The easiest AI product to build is still a chatbot. The founders who resist that will own the decade.
