Author: IRPA AI Senior Analyst, Kieran Gilmurray
Artificial intelligence just hit its third wave. The first wave was characterised by predictive, smart software that analysed data and identified trends. The second wave was generative; tools that created text, images, and code. Now, the third wave is agentic; autonomous systems that act, coordinate, and learn.
Boards are taking this seriously. PwC predicts that 68% of large enterprises will pilot agent-based systems by 2026. Firms with strong governance and a plan for reskilling will win. Others are at risk of being overwhelmed by complexity and cost.
From lab to ledger
On 14 April 2025, at 08:08, a European logistics firm rerouted 4,000 parcels when a motorway closed. The delay impact dropped by 35% without a single dispatcher. Three days later, a hospital in Singapore updated 600 patient plans overnight using agentic care bots. This isn’t theory anymore. It’s ROI.
What makes agentic AI different?
Unlike predictive models that wait for action or generative models that need prompts, agentic systems close the loop. They read intent, break down the task, coordinate with other agents, and finish the job. All without a human intermediary.
They do three things:
- Act independently. A cash agent rebalances liquidity on its own.
- Learn continuously. A customer bot refines replies from last night’s complaints.
- Aim for goals. A warehouse agent maintains a pick accuracy rate of 98%, continually tuning routes to ensure it remains there.
These aren’t tools. They’re teammates.
Brains, pipes, and power
Large language models (LLMs) like GPT-4o and Gemini bring judgement to code. Say “cut retail stockouts in London”, and they pull live data, chat with a replenishment agent, and issue POs in minutes. They don’t just know. They decide.
The execution layer matters too. Modern platforms give agents secure access to APIs, data streams, and vector databases. That means they can book trades, open support tickets, or restart machines. All governed by encryption, audit trails, and access controls.
Why multi-agent matters
One agent is clever. Many together are transformative.
In banking, a fraud agent flags a dodgy payment. Another calls the customer. A third files compliance paperwork. Carnegie Mellon found that task-specialised agent teams solve problems 41% faster than monolithic models.
Coordination protocols enable them to split tasks, monitor progress, and address any hiccups. No boss needed.
Sector impact, proven
- Healthcare: Theatre scheduling agents lifted utilisation by 15% in a US pilot.
- Finance: One bank cut overnight market risk by 22% in Q1 using risk-mitigation agents.
- Retail: A UK grocer increased its margin by 3% during peak season by utilising pricing agents.
- Supply chain: A courier cut empty miles by 12% using route optimisation agents.
Case study: logistics
A parcel carrier in Europe replaced its dispatch desk with three agents. One pulled the weather and GPS data. Another rerouted vans. A third reconciled fuel expenses. In six weeks:
- On-time delivery rose from 87% to 94%
- Refunds fell 28%
- Operating margin improved 2.3 points
- Token costs? €3,500/month. Less than four dispatchers’ salaries
What enables this?
Cloud vendors now offer orchestration frameworks that manage agent identity, memory, and handshakes. Vector databases make search instant. Observability tools track errors, cost per call, and latency. So, firms can tune performance before customers notice.
Risks that need action
- Black boxes: Agents in medicine and finance need traceable decisions and bias metrics.
- Jobs: Coordination roles shrink. But new ones arise in oversight and training. Gartner predicts that by 2028, half of the enterprise staff will work alongside digital agents.
- Compliance: EU, UK, and US rules now mandate overrides, logs, and explainability.
Real incident: data breach
In May 2024, a retail bank’s help bot leaked masked account numbers via prompt injection. Response layers kicked in fast. Access scopes narrowed data, filters blocked strings, and a kill switch took over in 20 seconds. Damage was limited. Still, the bank paid £1.2m in compensation. Guardrails must start at design.
Value trumps cost
Yes, tokens cost money. But results matter more. A service bot that clears 10,000 emails at £0.002 each saves 400 hours. At GPT-4o rates, the margin gain is significant. Early movers secure structural advantages that latecomers struggle to replicate.
Leadership playbook
- Pick a business goal: “Cut invoice cycle from 12 to 8 days.”
- Build 3 agent prototypes. Run them for 14 days.
- Track latency (ms), accuracy (%), cost (USD), failures (ppm).
- Pick the best. Lock the version. Licence it.
- Run risk drills for bias, security, and escalation.
- Write a digital contract: data rights, decision scope, human override.
- Train users in prompt craft, risk signs, and kill commands.
- Review quarterly. Retire poor agents. Scale high performers.
What’s next
By 2027, agents will run locally on devices and adjust goals independently. Imagine a turbine adjusting blade pitch mid-storm to boost energy yield. The edge isn’t computed anymore. It’s orchestration culture and proprietary data.
Conclusion
Agentic AI reallocates labour, cuts waste and enables new growth. Boards that move fast and govern well will lead. Those who wait will pay twice: in cost and trust.
Start small, start focused, but start now. Build the skills, craft the guardrails, and define the outcomes you want. The businesses that embed these systems today will run faster, leaner, and smarter tomorrow. Agentic systems are here to stay. It’s time to harness them.
The third wave is here. Surf it or sink.
About the Author: Kieran Gilmurray
Senior IRPA AI Analyst & Advisor, Kieran Gilmurray is a certified executive coach, intelligent automation & digital transformation thought leader & content guru focused on helping solve complicated problems others can't. For the past 25+ years, he has driven business digital transformation programs across a range of industries like digital technologies, intelligent automation, data analytics, social media and robotic process automation, having generated millions of dollars of value.
CLICK HERE TO SCHEDULE AN ANALYST BRIEFING
Links:
Originally posted in the IRPA AI Network — Announcements & Updates