A Manager's Briefing April 2026
Bot./Copilot./Agent.

What has changed, why it matters, and where to start — for the people who decide.

The journey so far 02 / 23
How we got here

Three generations of business software in ten years.

2015 · Software you click

Tools

Excel, Salesforce, SAP. You learn the menus. Software waits for instructions.

2023 · Software that drafts

Copilots

Embedded assistants in email, docs, CRM. They suggest. You approve.

This briefing is about the step from the middle column to the right.

Why this quarter, not next year 03 / 23
The market has already moved

Every second boardroom is writing an agent budget.

88%
of executives plan to increase AI budgets next year — specifically because of agents.
PwC · AI Agent Survey · 2025
79%
already adopting agents. Two-thirds of those report measurable value.
PwC · 2025
33%
of enterprise apps will include agents by 2028 — up from under 1% in 2024.
Gartner
Part One 04 / 23
01
Part One · Education

The three kinds of AI tools.

They look similar. They behave very differently. Knowing which one you're looking at is the first decision.

The three levels 05 / 23
At a glance

Three levels of autonomy.

LEVEL 01

Chatbot

Like a receptionist.
You ask, it answers. Nothing happens outside the window.
In your life
ChatGPT · Gemini · Claude · website support chat.
LEVEL 02

Copilot

Like an intern next to you.
Drafts the email, the slide, the formula. You approve every step.
In your life
Microsoft 365 Copilot · Gmail "Help me write" · GitHub Copilot.
LEVEL 03

Agent

Like a junior employee with a task.
You set a goal. It plans, uses tools, and returns the finished work.
In your life
ChatGPT agent mode · Salesforce Agentforce · Intercom Fin.
Level 01 · Chatbot 06 / 23
Level 01

Chatbot

A receptionist with a lot of knowledge — who never leaves the desk.

What it does
Answers questions. Reactive. No memory between sessions. No action in your systems.
A typical task
"Summarize this policy." "Explain IFRS 16 for leases." "Draft three subject lines."
What you pay
Often free or bundled. Paid tiers $20–30 per person per month.
Level 02 · Copilot 07 / 23
Level 02

Copilot

An intern who drafts — while you stay in control of what goes out.

What it does
Lives inside an app. Suggests the next sentence, row, or slide. Every output passes through you.
A typical task
"Summarize last 20 emails from this client." "Turn this transcript into action items." "Draft the quarterly update."
What you pay
Per seat. $20–30 / month is market — M365 Copilot, ChatGPT Business, Gemini in Workspace.
Level 03 · Agent 08 / 23
Level 03

Agent

A junior employee you hand a task to — and who returns with the task done.

What it does
Plans its own steps. Opens apps, clicks buttons, calls systems. Remembers across the task. Acts without asking at every turn.
A typical task
"Process this week's supplier invoices." "Handle tier-one refund requests under €200." "Research our top five competitors."
What you pay
Per action, per conversation, or per resolved outcome. Budgets look less like salary, more like electricity.
The one line to take home 09 / 23
Keep this sentence
Chatbotsanswer.
Copilotssuggest.
Agentsdo.
Part Two 10 / 23
02
Part Two · The shift

What makes an agent an agent.

Six capabilities that separate a real agent from a copilot in fresh packaging. PwC's definition, not ours.

Six capabilities · PwC 11 / 23
The PwC definition

What separates an agent from everything before.

01

Autonomy

Makes decisions without asking at every step.

02

Goal-driven

Works toward an outcome — not a script.

03

Environment-aware

Reads screens, data, context. Adapts in real time.

04

Learning

Gets better with feedback and history.

05

Workflow orchestration

Runs a whole process, not one step.

06

Multi-agent coordination

Talks to other agents. Hands work off. Specializes.

Source · PwC Agentic AI Executive Playbook

The shift in plain language 12 / 23
The difference, in one sentence
A copilot helps you do the work.
An agent does the work.

That's why the conversation has shifted from seats (price per user) to outcomes (price per resolved customer, processed invoice, closed ticket). You're no longer buying a tool — you're buying a result.

Part Three 13 / 23
03
Part Three · It's working

It's already working.

Three stories from companies you've heard of. No pilots — production systems, real numbers.

Case · JPMorgan Chase 14 / 23
JPMorgan Chase Financial services · Legal
360,000hours / year
of manual legal review, automated by their COiN platform.
The process: interpreting commercial loan agreements. Lawyers and loan officers once read each one by hand. Agents now extract clauses, flag risks, and route edge cases to humans. What changed for managers: the team moved from reading to reviewing.
Case · Siemens 15 / 23
Siemens Industrial · Manufacturing
−20% maintenance cost
across factory floors using predictive agents.
The process: factory equipment fails in ways humans can't predict from a dashboard. Agents watch sensor streams, forecast failures, and schedule technicians before a machine goes down. What changed: uptime rose; emergency call-outs dropped.
Case · Hospitality (PwC client) 16 / 23
Global hospitality franchise Guest services · Back office
90% productivity gain
in selected back-office functions, after embedding agents across the business.
The process: rate management, guest-issue resolution, franchise-owner support, marketing briefs. Agents took the repeatable work; staff focused on guest experience. The caveat: 90% in some areas — not all. Picking the right process matters more than picking the right tool.
The macro signal 17 / 23
Boardroom view

Companies investing in AI beat the market.

+21%
Total shareholder return vs. sector median, for S&P 500 companies investing >0.5% of revenue in AI (2022–2025).
PwC analysis · S&P 500
+3%
Revenue growth above sector median for the same cohort — a real-economy signal, not just market sentiment.
PwC analysis
62%
of leaders expect agents to deliver greater than 100% ROI.
PagerDuty · Wakefield · 2025
Part Four 18 / 23
04
Part Four · Honest

Why 40% will fail.

Three reasons agent projects are being cancelled in 2026 — and how to avoid becoming one of them.

Why a 95% reliable agent still fails 19 / 23
Reason 01 · The math

A 95% reliable agent fails four times in ten over ten steps.

100% 75% 50% 25% 1 5 10 15 20 NUMBER OF STEPS ~60% success ~36%
End-to-end success
1 step 95%
5 steps 77%
10 steps60%
20 steps36%

What managers take away: shorter processes, humans in the loop at checkpoints, and an agent that can say "I'm stuck."

Why projects fail 20 / 23
Reason 02 · Trust and governance

Most failures aren't technical. They're organisational.

>40%
of agent projects will be cancelled by 2027 — for lack of clear value, cost controls, or risk management.
Gartner
28%
of executives rank lack of trust as the single biggest obstacle to getting value from agents.
PwC
50 / 42%
Half believe their operating model will be unrecognisable in two years — only 42% are actually redesigning processes.
PwC

Buying agents is easy. Redesigning work around them is where the value sits.

The security line managers miss 21 / 23
Reason 03 · The security surprise
Agents don't break governance. They show you how broken your governance already is.
Microsoft · on rolling out Copilot across enterprises

Every over-shared folder, every "anyone with the link" document, every inherited group permission — now reachable at machine speed, by an agent running on your behalf. Clean the permissions before turning on the agents.

How to start 22 / 23
A manager's starting playbook

Three moves. In that order.

01 · This month

Pick one process.

High-volume, repetitive, rules-based, measurable. Invoices. Tier-one support. Meeting notes. One team, one process.

02 · This quarter

Run a scoped pilot.

Draft-only first. A human approves every action for 30 days. Measure one number: time saved or resolved rate.

03 · Next quarter

Redesign the work.

Only 42% of companies do this. It's where the compounding value sits. Move people from doing to reviewing.

The vendor with the prettiest demo is not always the right call. The right process always is.

End of briefing 23 / 23
// TAKE HOME
Name the level.
Pick the process.

The market has moved. The technology works. The hard part is organisational — and that's the part you decide.