Agentic AI is growing up.

Companies are moving from chatbots that answer to agents that act. Some are saving millions. Many are stuck. Here’s what the case studies and surveys really say.

5 case studies 6 charts Readiness quiz
Scroll ↓

Chatbot. Script. Agent.

A customer disputes a charge. Three tools, three very different outcomes.

Chatbot

Works from a question
  • Explains the refund policy
  • Talks about surprises
  • Rarely acts in systems
  • Human does the follow-through

Automation (RPA)

Works from a fixed script
  • Refunds only if the case matches the rules
  • Breaks on surprises
  • Runs only the steps it was given
  • Human maintains the script

AI agent

Works from a goal
  • Checks the order and the policy, issues the refund
  • Reasons through many surprises
  • Chooses its own steps
  • Human sets the goal and handles exceptions

How an agent works

Better reasoning, standard ways to plug models into tools, and orchestration software that tracks long chains of steps made this possible. Tap a step.

Who’s actually doing it

Most figures come from vendors or secondary write-ups, not audited financials. Treat them as directional. Swipe →

0

Klarna

Customer service

One support agent credited with the work of about 853 full-time staff. High volume, repetitive, easy to measure.

0

JPMorgan Chase

Breadth

Use cases reportedly running in production every day. Agents managed like a portfolio, each tied to one process.

0

General Mills

Supply chain

An autonomous system for shipment and inventory exceptions. Clear rules, overwhelming volume.

0

FM Global

Discovery first

The insurer mapped its processes first and found 10+ automation targets before building any agent.

Rules

Rimini Street

Governance first

Company-wide adoption managed by an AI steering committee, access policies and monitoring.

Klarna, JPMorgan and General Mills: 2026 roundup. FM Global: Skan AI whitepaper. Rimini Street: TechTarget. Bars show savings relative to Klarna where a dollar figure exists.

Deployed, but not trusted

A Forrester study commissioned by Boomi found a wide gap between using agents and believing in them.

0
have deployed agents
0
trust them

Among companies with real control over their agents, 59% reported productivity gains and 45% automated repetitive tasks.

Adoption vs. reality

The surveys disagree because “deployed” means different things in each. A single-team pilot can count as deployed in one and as piloting in another.

Deployed (Boomi)
86%
Piloting
50%
Production-ready
14%
Fully deployed
9%

“Just exploring” is shrinking

Last August
52%
November
30%

Sources: Forrester/Boomi, a separate 2026 report (9% and 50%), Deloitte as cited in one guide (14%), PYMNTS (52% to 30%). One guide citing Gartner says over 40% of agentic projects will be cancelled by the end of 2027.

Winners vs. stalled

Same technology, opposite habits. Switch tabs.

One workflow Clearly defined scope, not “use AI.”
Rules first KPIs, evaluation and governance set early.
Human exit ramp A named person takes over on edge cases.
Narrow, then wide Prove one use case, then expand.
Vague goal “Deploy AI” with no target.
Late controls Governance bolted on after launch.
No escalation The agent freezes or confidently errs.
Big bang Company-wide on day one.

The ROI question

Numbers vary a lot. Ask who paid for the research and how ROI was defined.

Microsoft study
~120%
Global average
171%
U.S. average
192%
$44.5M

benefits against $20.2M costs over three years in Forrester’s Microsoft study.

8.3 mo

median payback claimed by an aggregated, vendor-friendly source.

Fast first

Customer service and document processing tend to pay back fastest; supply chain and legal take longer.

What can go wrong

Blockers named in the surveys and case studies.

🔒

Security

Cybersecurity gaps were the top blocker at 31% in one report.

🗂️

Data quality

Messy data held back reliable scaling for 23%.

🧭

Governance

Weak governance and orchestration keep many projects in pilot mode.

🚧

No exit ramp

Agents with no escalation path stall or answer wrongly with full confidence.

Are you ready?

Tick what’s true for your team.

Tick a box to start.

Where this leaves us

Agentic AI has moved past the lab-curiosity stage. In narrow, well-bounded jobs it delivers measurable results. But the gap between “we use agents” and “we trust agents” is still wide, and that gap, more than any model upgrade, will decide which projects reach 2027.

Know what each agent is for Know how it’s checked Know who steps in