When The Agent Enters The Org Chart

For three years, enterprise AI has mostly meant systems that respond: draft this, summarise that, answer the customer. Agentic AI is a different proposition. An agent is given a goal, not a prompt – and it plans the steps, chooses and operates the tools, queries the systems, and executes multi-step work with minimal supervision, escalating to a human only where its mandate ends. The practical unit of adoption shifts accordingly: not a tool per person, but a digital worker per process.

That shift is why the future-of-work conversation has changed registers. The question is no longer how employees use AI. It is how organisations are redesigned when some of the workforce is software.

Where Adoption Actually Stands — Both Halves Of The Truth

C-level readers deserve the sober numbers alongside the exciting ones, because both are true at once.

The excitement is measurable. Gartner’s 2026 CIO and Technology Executive Survey finds only 17 per cent of organisations have deployed AI agents to date, but more than 60 per cent expect to within two years, the most aggressive adoption curve among all emerging technologies the survey measures. The firm forecasts that by the end of this year some 40 per cent of enterprise applications will embed task-specific agents, up from less than 5 per cent in 2025, meaning much of the agent adoption in any organisation will arrive inside software it already licenses, whether or not anyone decided to “adopt agents.” McKinsey’s research finds 62 per cent of organisations experimenting with agents and 23 per cent already scaling them in at least one function, with banking and financial services consistently reported among the furthest ahead — a finding with obvious resonance for this continent’s most digitised industry.

The sobriety is equally well documented. Gartner itself, placing agentic AI at the very peak of its hype cycle, predicts that more than 40 per cent of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls; and MIT research last year found the overwhelming majority of generative-AI pilots produced no measurable financial impact. The pattern is not a contradiction. This is a technology whose capability is real and whose deployment discipline is rare. That is what makes the gap between ambition and execution a management problem, not a model problem.

Follow The Money: The Industry Has Already Voted

The clearest evidence that implementation, not intelligence, is the scarce factor came from the vendors themselves this quarter. Within roughly 60 days, Microsoft committed $2.5 billion and 6,000 embedded engineers to its new Frontier Company; Amazon Web Services put $1 billion into its own embedded-engineering initiative; OpenAI stood up a deployment company backed by more than $4 billion; and Anthropic joined Goldman Sachs, Blackstone and Hellman & Friedman in a $1.5 billion venture to place engineers inside mid-sized firms. Every major provider has concluded that the money now lies in making agents work inside real organisations. The model of forward-deployed engineering pioneered by Palantir two decades ago, industrialised.

Two design signals from that race matter for every IT strategy. First, model-agnosticism is becoming the default: serious deployments assume the ability to switch models by workload rather than lock to one provider. Second, data boundaries have become a competitive pledge. Microsoft’s Satya Nadella framing it as bluntly as anyone: there is no societal permission for an AI future that consumes the intelligence of its customers. When vendors compete on promising not to learn from your data, contract language has become architecture.

What Actually Changes In The Work

Strip away the futurism and the changes agentic systems introduce are specific.

Work decomposes differently. Processes long organised around human hand-offs such as gather, check, reconcile, draft, and route, become sequences an agent can traverse end to end, which is why the organisations reporting real returns talk about redesigning workflows rather than accelerating old ones. As Jackline Mburu, the newly appointed Africa head at Happiest Minds, frames it, the coming era is about enterprises where AI “doesn’t just provide insights but actively collaborates with people, automates decisions, orchestrates workflows.”

Management changes shape. Microsoft’s Frontier Firm research describes human-agent teams in which the employees direct squads of agents. A span-of-control question HR systems are not primed to ask. Judgment, verification and exception-handling rise in value precisely because routine execution falls in cost.

And the talent pipeline question opens. If agents absorb much of the routine work through which junior professionals have always learned, organisations must deliberately rebuild the apprenticeship their career ladders assumed. This then becomes an unresolved design problem the literature flags rather than solves.

Closer To Home

The equation has distinctive terms on both sides.

The opportunity side is real. Thinner legacy estates mean fewer entrenched processes to unwind. “African banks,” as Mburu argues, “have an extraordinary opportunity to become global leaders in digital banking.” The economics have moved too: the efficiency breakthroughs and open-weight models of the past 18 months have collapsed the assumption that frontier-class AI is only for those who can pay frontier-class prices, and the sector globally furthest into production, financial services, is the one where this continent’s digital sophistication is deepest.

Nor is the conversation theoretical at the top of the market. MTN Group President and CEO Ralph Mupita used his MWC Barcelona keynote this year to describe the operator’s transition from conventional telco to technology platform, sending a delegation focused on AI and agentic-AI themes. “Although AI is still in its infancy for us, it’s already enhancing customer experience,” he said, with the group’s next five years built around AI-powered networks and scaled fintech. In banking, the testing is concrete. Last September, Absa Bank became the first institution on the continent to deploy Salesforce’s Agentforce agentic platform, and in commentary published by the bank, former Absa Group’s BB CIO Lindelani Ramukumba describes three autonomous agents in testing – including a co-pilot for relationship managers, who have historically spent as much as three-quarters of the working day assembling customer information across systems rather than engaging customers. That one statistic is the agentic business case in miniature: the work agents absorb first is the work that was never really the job.

The constraint side is equally factual. Capability remains the scarce input. “Technology can always be bought. Capability cannot,” as Vincent Entonu, Managing Director of Westcon Microsoft Sub-Saharan Africa, put it at a Nairobi executive forum this year, and the continent’s skills initiatives, from corporate AI institutes to national programmes, exist because the gap is acknowledged. Infrastructure reality must shape design: this publication’s own analysis of regulator data shows a third of Nigeria’s connections still on 2G and consumption bending under data costs, which means customer-facing agents built for the mass market must work over thin, expensive connectivity or not at all. And the UN’s International Scientific Panel on AI warned this month that most countries lack the technical expertise to govern the most powerful systems, a caution that applies inside enterprises as much as governments.

The Seven Decisions IT Leaders Cannot Delegate

What separates the scaled deployments from the cancelled ones, across the research and the case record, reduces to a set of decisions that belong to technology leadership. Not to vendors, and not to pilots.

1. Accountability: An agent acts; someone answers. Before deployment, define who owns each agent’s actions, which decisions require human approval, what gets logged for audit, and how an agent is suspended. Treat every agent as a hire: a job description, a scope, a manager.
2. Identity and security: Agents are non-human identities holding credentials, permissions and memory. They are also a new attack surface layered onto the old one. The regional threat data is instructive: the attacks dominating this continent’s statistics are credential brute-force and unpatched-system exploitation, and every agent added multiplies exactly what those attacks target. Least-privilege scoping and separate agent identity management are prerequisites, not enhancements.
3. Data readiness and data boundaries: Agents inherit the quality of the plumbing they run on; broken data produces confidently executed mistakes. Equally, contracts must state in plain language whether your data trains anyone’s models. The market now offers that pledge; require it.
4. Architecture before agents: The organisations getting value build the decisioning and orchestration layer first — one place where rules, models and hand-offs are governed — rather than scattering point-solution agents across departments. Preserve the ability to swap models; be deliberate about how much of the pipeline sits with a single vendor.
5. Measurement with teeth: The 40 per cent cancellation prediction is, read correctly, good news: it describes organisations learning to kill what does not pay. Tie every agent to a metric the CFO recognises, and set the shutdown criteria on day one.
6. People, redesigned on purpose: Train the managers of agents; rewrite their roles around judgment and exception-handling; and consciously rebuild how juniors learn when routine work no longer teaches them. Capability built in-house compounds; capability rented does not.
7. Cost governance: Agentic economics are usage-based and compounding, tokens, tool calls, and orchestration overhead. The discipline that cloud spending eventually forced, agent spending will force sooner. Budget it like FinOps from the first pilot.

The Bottom Line

Agency can be delegated. Accountability cannot. The organisations that gain most from agentic AI over the next three years will not be those that deploy the most agents, but those that redesigned work, governance and skills so that delegation is safe. The evidence suggests the winners’ playbook is discipline, not enthusiasm. The leapfrog is genuinely available: less legacy to unwind, cheaper capable models, a young workforce to build around. It is conditional on exactly the seven decisions above, every one of which sits, unavoidably, on the technology leader’s desk.