The strongest AI workers will not stand out for raw intelligence. They will stand out because they understand one sector’s revenue chain, its systems, its exceptions and its definition of acceptable work.
For most of 2024, the AI market seemed to be racing towards one universal assistant. One clever system would help every company with almost everything. That picture is fading. The money, the customers and the measurable results are moving to specialised AI agents: systems built to do one kind of work, inside one kind of business, against one clear standard.
The shift matters most for companies that do not have a research lab. A distributor in Lagos, a clinic group in Accra or a property firm in Manchester does not need a machine that can write a sonnet and debug code. It needs something that can answer a delivery question at 9pm, check stock before promising a date, and know when to hand the conversation to a person.
Key takeaways
- Investors and buyers are backing specialised AI agents with narrow, measurable jobs over general-purpose assistants.
- Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, mostly on cost, unclear value and weak risk controls.
- The agents that work share five traits: sector language, revenue-chain knowledge, system access, exception handling and a clear standard of acceptable work.
- Recent results from Octopus Energy and Klarna show the same lesson: scope tightly, label AI output, and keep a human route open.
Why are specialised AI agents attracting the money?
Consider ChipAgents, a Santa Clara company founded in 2024. It builds AI agents for semiconductor design and verification, work that depends on deep technical context and measurable accuracy.
In July 2026 it added $60 million to its Series A, taking the round to $134 million. Backers include Bessemer Venture Partners, Micron, MediaTek and B Capital. The company says more than 120 chip companies have deployed its platform. That figure is self-reported and unaudited, but the direction is clear.
Nobody is paying for ChipAgents because it is clever in general. They pay because it knows how a chip gets verified, which tools engineers already trust, and what a correct result looks like. Its value sits in the narrowness.
The same logic holds far from Silicon Valley. A WhatsApp agent that understands how a Nigerian bank customer checks a balance is more useful than one that can discuss philosophy. Zenith Bank saw this early. In 2021 it launched ZIVA on WhatsApp, handling defined tasks such as balance checks, transfers, airtime top-ups and account blocking.
What separates a useful AI agent from an expensive experiment?
Gartner offered a blunt warning in June 2025. It predicted that over 40% of agentic AI projects will be cancelled by the end of 2027. It named escalating costs, unclear business value and inadequate risk controls.
The firm also flagged “agent washing”. Vendors were relabelling chatbots and automation tools as agents. Gartner estimated that only about 130 of the thousands of agentic AI vendors were real.
Read together, those figures say something simple. A general promise is easy to sell and hard to prove. A narrow job is harder to sell and easy to measure.
In practice, a valuable AI worker understands five things.
- The sector. It speaks the industry’s language. A logistics agent knows what a waybill is. A property agent knows the difference between an inspection and an allocation letter.
- The revenue chain. It knows how an enquiry becomes an order, a delivery, an invoice and a payment. It can tell where a deal stands.
- The systems. It works across the CRM, the finance tool, the messaging channel and the stock records already in use. An agent that cannot read the order history guesses.
- The exceptions. It recognises when the normal process no longer applies: a disputed payment, a vulnerable customer, a refund above a set limit.
- The standard of acceptable work. Its output is checked against evidence, business rules and human expectations, not against how confident it sounds.
What do real deployments of specialised AI agents show?
Two recent cases from Europe point in the same direction.
In London, Octopus Energy ran a three-month trial of Arlo, its AI support assistant. Arlo answered routine emails about tariff renewals, payment dates and account details. It never handled vulnerable customers, sensitive cases or complex complaints. Every AI-written email carried a label, and customers could ask for a human at any time.
The result, published by Octopus in July 2026, was a 76% satisfaction score for Arlo against 72% for comparable human advisers. Arlo handled about 8,000 emails a week, roughly 4% of the company’s UK customer email. These are the company’s own figures.
The interesting number is not 76%. It is 4%. Octopus did not hand its inbox to a machine. It handed over one tightly defined slice and measured it.
Klarna took the broader route first. In February 2024 the Swedish payments company said its AI assistant could do the work of 700 customer service agents. By May 2025 its chief executive, Sebastian Siemiatkowski, was telling Bloomberg that the AI-first approach had produced lower quality. Klarna began recruiting human support staff again, so customers could always reach a person.
One company scoped narrowly and kept humans in the loop. The other scoped broadly and walked part of it back. Both lessons point to specialisation.
How does this apply to businesses in Nigeria, Ghana and the UK?
Most firms in Lagos or Accra are not building agents from scratch. They are deciding what to automate first. The question is where a narrow agent can carry real weight.
In Nigeria, much of that weight sits in WhatsApp. Orders, price checks and payment confirmations often happen in chats rather than on websites. An agent that answers stock and delivery questions, logs each enquiry in the CRM and flags a hot lead to a salesperson does one job well. It does not need to do anything else.
In Ghana, the pattern overlaps but differs. Mobile money confirmations, delivery windows and repeat orders shape many customer conversations. An agent tuned to those moments is more valuable than a general assistant that knows nothing about how a Kumasi retailer gets paid.
For UK firms, the constraint is often regulation and tone. An insurance broker in Birmingham may want an agent that books callbacks and collects policy details, but never gives advice. The boundary is the product.
Across all three markets, the same test applies. If a manager cannot write down what the agent should do, what it must never do, and how a good answer is checked, the agent is not ready to meet customers. That written brief is worth more than any model upgrade.
This is the kind of scoping work Riophany does with clients, mapping where an agent fits into the existing CRM and customer conversations before anything is built. Done in that order, the agent inherits the business’s knowledge instead of inventing its own.
Which jobs should a specialised AI agent take first?
The best first jobs share three features. They are frequent, they follow rules, and a mistake is easy to catch.
| Good first job | Why it suits an agent | Keep a human for |
|---|---|---|
| Order status and delivery questions | High volume, answer sits in your systems | Lost or damaged goods |
| Lead qualification on WhatsApp | Same questions every time, logs to CRM | Large or unusual deals |
| Appointment booking and reminders | Clear rules, calendar-driven | Complaints and cancellations with penalties |
| Payment confirmation and receipts | Checkable against finance records | Disputes and refunds above a limit |
Each row has the same shape. The agent handles the routine case and knows exactly when to stop. That boundary is what turns an experiment into a working system. For a closer look at where enquiries slip today, see our analysis of customer follow-up gaps in Nigeria.
Frequently asked questions
What are specialised AI agents?
They are AI systems built for one defined job in one kind of business, such as answering delivery questions or qualifying leads. They connect to the company’s own systems and follow its rules, instead of trying to handle any task at all.
Are general AI assistants still useful?
Yes, for drafting, research and personal productivity. But for customer-facing work that touches orders, payments or regulated advice, a narrowly scoped agent is easier to test, govern and trust.
Why do so many agentic AI projects fail?
Gartner points to escalating costs, unclear business value and weak risk controls. Projects with a vague brief struggle to show results, while narrow ones can be measured from week one.
Should an AI agent always offer a human?
Both the Octopus Energy trial and Klarna’s reversal suggest yes. Labelling AI replies and keeping a clear route to a person protects quality and customer trust.
Where should a Nigerian business start with AI agents?
Start with a frequent, rule-based conversation that already happens on WhatsApp, such as order status or lead qualification. Connect it to the CRM so every conversation is recorded and followed up.
The first wave of AI promised a colleague who could do anything. The second wave is quieter. It offers a colleague who does one thing, knows where the file is kept, and knows when to knock on the manager’s door.
