From Chatbots to Co-Workers

Three questions decide whether your AI investment compounds into operating advantage or evaporates into another round of pilots.

By Yiannis Stavrianos, Senior Manager, Advisory services, PwC Cyprus

This article series began with a fact and a question. The fact: the financial impact of AI at the leaders is real and audited Klarna’s ~$60M in customer-operations savings, JPMorgan’s ~$2B in annual benefit, GitHub Copilot now used by approximately 90% of the Fortune 100 and the leaders are not using fundamentally different models; they are rebuilding the operating model around them. The question: why has capturing that value proved so much harder than the demos suggested? The second article gave the diagnosis in three parts a procurement without a redesign, a chain without checkpoints, a turbocharger on a stock car. AI is a business problem dressed up as a technology problem, and the leaders did the business work first.

This final article translates the diagnosis into a playbook. It is a structured way of answering three questions. WHERE to invest. HOW to govern. HOW to measure. Most organisations that fail at AI fail because they tried to answer the third question before they had honest answers to the first two. The leaders answer them in order.

Question 1: where to invest

The most expensive mistake at the start of an AI programme is to treat the choice of where to deploy as a democratic exercise letting every function nominate a pilot, funding the ones with the most enthusiastic sponsors, and measuring "AI activity" instead of AI impact. PwC’s 2026 Digital Trends in Operations survey found that 85% say they are ahead of most competitors in digital transformation, yet 89% say their technology investments have not fully delivered. Those two numbers matter because they cannot both be true and together they are the signature of investment without prioritisation.

The leaders look for functions with highest P&L impact and lowest operational risk. Four categories tend to score well on both: first-line customer operations (Klarna’s territory); invoice and back-office reconciliation; fraud, risk, and anomaly detection (JPMorgan’s 11% per-unit fraud cost reduction); and engineering and analyst productivity (the Copilot story). Two or three deployments in these categories, properly resourced, will produce more measurable financial impact in year one than fifteen exploratory pilots scattered across the organisation. This is a CEO and CFO decision, not an IT decision.

Question 2: how to govern

Before any agentic system touches customers, money, or sensitive data, eight things should be answered in writing by the executive sponsor. Verification agents that check each other’s work, rather than one model doing everything. Human-in-the-loop checkpoints at the moments where the consequences of getting it wrong are most expensive money movement, external communication, hard-to-reverse decisions. Bounded permissions the email-drafting agent does not have the ability to move money. Deterministic guardrails wrapped around the probabilistic core the AI generates the action, traditional code enforces the rules.

Logging and observability every decision traceable months later. Continuous evaluation against historical cases, alerting when accuracy slips. Escalation and fallback paths a defined route back to a human when the agent is uncertain. Red-teaming by someone whose job is to break it, before production. PwC’s 2025 Responsible AI Survey found that 60% of executives now report that responsible AI practices boost ROI and efficiency, and 55% report improved customer experience. That finding matters because it retires the oldest objection in the room: the discipline of governance is itself a driver of the financial outcomes not a brake on them.

Question 3: how to measure 90 days, five moves

1. Commission an honest audit of every AI deployment running in the organisation today including the shadow AI in use without IT approval. You almost certainly do not know the real number.
2. Name a single accountable executive for AI reporting to you or to the COO, not buried inside IT with authority over both opportunity capture and risk.
3. Adopt the eight non-negotiables as a gating mechanism. Before any new agentic deployment touches customers, money, or sensitive data, require each item to be answered in writing. No exceptions for "pilots."
4. Pick two or three quick wins in the categories above, with bounded scope and clear baseline-versus-AI metrics. Resource them properly.
5. In parallel, stand up the governance scaffolding framework, prioritisation model, roadmap, operating cadence so that by the time the quick wins prove themselves, you have a platform to scale.

On day 91, three signals separate organisations making real progress from those simply talking about AI.

• Visibility: a complete AI inventory available in one meeting.
• Throughput: every quick win is in production with measured results, or formally killed with a written reason.
• Gating in practice: at least one proposed deployment has been delayed or redesigned because it failed the checklist. If everything sails through, the checklist is being rubber-stamped.

The call to action is deliberately compact. This week: commission the audit and name the accountable executive. This month: put the gating checklist in force and select the two or three quick wins. None of it requires a transformation programme to begin. It requires a decision.

The opportunity in front of you is one of the largest of your career. The question is no longer whether agentic AI will reshape your industry. It is whether you will be one of the organisations doing the reshaping, or one of the ones explaining to your board, eighteen months from now, why a competitor is suddenly two steps ahead.

The work starts now.