Beyond the pilot: how AI agents cut costs and multiply productivity

The agentic productivity paradox

Nearly nine out of ten companies have adopted AI in at least one business function. Only one in fifteen claims to have derived "significant" value from it. This is the paradox McKinsey's latest State of AI survey puts on the table of executive committees: AI is everywhere, but its impact on EBIT remains marginal for the majority of organisations.

AI agents — systems capable of planning, reasoning and executing multi-step tasks autonomously, not merely reactive copilots — promise to close that gap. Today 23% of companies have scaled the use of agents in at least one function, and among those who use them two thirds report tangible productivity gains and over half report cost savings. But the relevant message is not the adoption rate: it is the fact that value is not evenly distributed. It concentrates on companies that made a precise choice — and made it quickly.

This article is an operating map for that choice.

Where agents generate value (and where they don't)

Three messages should guide the decision of a CEO or a COO.

First, value is asymmetric. The functions showing the clearest economic impact today are software engineering, IT and manufacturing, with cost reductions in the order of 10–20% reported by those who measure. Marketing, sales and product development generate revenue uplifts above 10%. The practical conclusion: those choosing where to start should begin with documented processes, rich in structured data and with repetitive cycles — not the most visible or "trendy" areas.

Second, pure efficiency is a trap. Data on so-called AI high performers — the 6% of companies that attribute over 5% of their EBIT to AI — show a clear pattern: this group does not chase cost cutting alone. They simultaneously pursue growth and innovation, and are 3.6 times more likely than others to pursue enterprise-wide transformation rather than incremental improvement. Those who set up AI agents as a cost cutting project tend to obtain modest, isolated savings. Those who set them up as work redesign get the multiplier.

Third, the constraint is not technological, it is organisational. Microsoft, in its most recent analysis of digital work, attributes a value impact on AI generated by organisational factors — culture, managerial practices, talent support — double that of individual effort. Agents that remain "boardroom demos" are almost always demos placed next to the process, not inside the process.

Five families of high-return use cases

Based on deployments scaled internationally, the AI agents with the best value-to-complexity ratio concentrate in five families.

1. High-volume repetitive internal operations. IT ticket triage, customer request classification, accounting reconciliation, RFP and bid management. A public case illustrates the principle well: Gelato — a Norwegian software house active in print-on-demand — brought ticket assignment accuracy from 60% to 90% and reduced ML model deployment times from two weeks to one or two days thanks to agent orchestration on Vertex AI. Where there is a structured workflow and sufficient volume, an agent replaces hours of discretionary human work with minutes of supervised execution.

2. Assisted customer-facing. Not chatbots, but agents integrated with transactional systems. JPMorgan Chase uses AI models to scan transactions and identify fraud in real time; American Express personalises offers and recommendations in near real time based on purchase behaviour and risk signals. The operating rule: the agent is useful when it can access customer data, act on back-office systems and close the loop — not when it merely "responds".

3. Supply chain and industrial change management. Coca-Cola Beverages Africa is using agents to manage manufacturing change management processes, reducing approval times from weeks to days. The pattern recurs in manufacturing: agents orchestrate cross-functional workflows and reallocate resources dynamically based on demand.

4. Software development. This is the area where reported cost savings are highest and most documented. A large retail company cited by PwC halved development cycles and reduced production errors by half starting from agents for software, before extending them to HR, finance, supply chain and marketing. Software is the "bridge" use case for those who want to build internal competence before scaling.

5. Specialist knowledge work. Harvey uses advanced models to automate legal document review, reasoning across hundreds of pages to free professionals for strategic work. The same pattern applies to financial due diligence, audits, technical writing and architectural and engineering design.

The decision framework: three horizons, one portfolio choice

Companies that capture value from AI agents work simultaneously across three horizons.

Horizon 1 — Efficiency of existing processes (0–6 months). Insert agents into already-mapped flows, with measurable KPIs: cost-to-serve, cycle time, error rate. Objective: validate the ROI and build the organisational learning curve. Contained investment, fast return, important internal narrative.

Horizon 2 — Flow redesign (6–18 months). Redesign processes assuming the agent is a node of the workflow, not an add-on. This is where you enter high performer territory. Example: instead of automating ticket triage, redesign the entire customer service flow with an agent orchestrating research, decision, action on systems and human escalation only where needed.

Horizon 3 — New service models (12–36 months). Offerings enabled by agentic AI that today would not be economically sustainable: personalised advisory at scale, continuous predictive monitoring, commercial proposals generated in real time. This is where long-term competitive advantage is played.

The typical mistake is staying glued to Horizon 1, mistaking a collection of successful pilots for a strategy. McKinsey's research is explicit: two out of three companies have not yet started scaling AI at the enterprise level, and EBIT value remains concentrated in that 6% that made the leap.

The operating blueprint: 90 days, 6 months, 12 months

First 90 days — Setting the ground

Select two or three use cases with structured data, already-measured KPIs and strong functional sponsors. Avoid areas where you do not know the unit cost of the process: you cannot measure a saving you cannot estimate. Form an agile team of 4–6 people with a business product owner, not just IT. Define success metrics upfront — cost-to-serve, FTEs freed for higher-value activities, error rate, cycle time — because without a baseline there is no ROI. Set up minimum governance: human-in-the-loop on decisions with customer or regulatory impact, complete audit trail, clear ownership of exceptions.

From 90 days to 6 months — Redesigning the flow

Once the pilot is validated, stop before scaling. The question to ask is not "how do we deploy the agent faster" but "how does the process change if we take the agent as the starting point". Reallocate the human time freed up toward activities that generate revenue or reduce risk: if cost savings end up in the P&L without work reconfiguration, the initiative will stop at the first budget review. Build a minimum internal capability — at least an AI product owner, a data engineer and a governance lead — to reduce dependence on external consultants.

From 6 to 12 months — Scaling the platform

Standardise an agentic infrastructure: same security model, same evaluation framework, same library of tools and integrations. Extend to Horizon 2 use cases: end-to-end customer journey, finance close, demand planning. Start measuring impact at EBIT level, not just by project. This is the shift from "sum of pilots" to "transformed line of business".

Three risks that derail the economic case

Data and integration risk. MIT Sloan research is clear: in real deployments, around 80% of the effort is not prompt engineering but data engineering, governance and integration into workflows. Companies that underestimate this cost burn the gains in technical friction.

Governance and reliability risk. Over half of companies have already reported AI incidents. An agent that decides on credit, claims, admissions or payments without traceability and consistently applied standards is a source of reputational and regulatory risk disproportionate to the saving. The principle "human-in-the-loop where it counts, autonomous where it is safe" is not optional.

Organisational risk. About one company in three plans a headcount reduction above 3% in the next twelve months thanks to AI. If cost savings are not communicated and managed as part of a reskilling and role redesign programme, internal rejection blocks scaling regardless of technology quality.

The question to bring to the board

For those leading the company, the question is no longer "are we using AI?". Everyone is. It is: what percentage of our EBIT is reasonably attributable to AI initiatives twelve months from now, and what process redesign makes that plausible?

If the answer is "we don't know" or "a few cost points in call centres", the trajectory is clear — and it is the same as the 94% of companies that today report not seeing significant value from their investments.

If instead the answer includes two or three core processes that will be fundamentally rethought over the next six to twelve months, with explicit EBIT KPIs and a structured team to lead them, the company is positioning itself in the small group that will capture the disproportionate share of value.

AI agents are not a technological solution. They are a management choice. And like all management choices that matter, they reward those who decide early and well — and penalise those who wait for the market to settle.