Field Notes: When Software Starts Showing Up Like Staff

Picture a normal Wednesday.

Invoices. Tickets. Exceptions. A queue that never ends. The kind of work that’s not “hard,” exactly (until it is), but it’s heavy. It’s where process goes to multiply.

Now imagine that work getting assigned, checked, escalated, and documented by something that doesn’t live in your toolbar.

That’s the shift this week is really about: AI agents as digital employees. Not a helper you summon. A system that watches the workflow, makes a plan, takes an action, and leaves a trail.

What Counts as an “Agent” (Not a Chat Box)?

Here’s the thing. A lot of tools are going to call themselves “agentic” because it sells.

A practical definition—the one that matters for your org chart, controls, and risk posture—requires four capabilities:

1. Observe context (systems, documents, state)
2. Decide what to do next (within rules)
3. Act (click, create, update, route, notify)
4. Prove what it did (logs, tracing, audit)

If it can’t do number four, it’s not ready for serious work. It’s a demo.

Story 1: Goldman’s “Digital Co-Workers” (And Why the Wording Matters)

The most important part of the Goldman story isn’t the model name. It’s the vocabulary: “digital co-workers,” “digital employees,” governance that starts sounding like staffing and controls.

That language is a tell.

It means leadership is no longer treating these systems as a feature. They’re treating them as labor that needs:

• Permissioning
• Supervision
• Separation of duties
• Documentation
• Limits (especially in regulated workflows)

If you want to write an agent strategy that survives contact with compliance, that’s the bar—one already reflected in large-scale enterprise rollouts like DXC’s enterprise-wide Amazon Quick deployment.

Story 2: DXC’s 115,000-Person AI Workspace (What “Enterprise Scale” Looks Like)

DXC didn’t just pilot an assistant. They rolled out an agentic AI workspace, Amazon Quick, across 115,000 employees in 70 countries. That number matters because it drags reality into the room.

At that size, you can’t hide behind vibes. You need:

• Access controls that match your identity stack
• A real data boundary (what agents may and may not see)
• Common workflows people will actually use
• Support and change management (yes, still)

DXC also launched a DXC Amazon Quick Practice to help clients do the same thing—a classic “we proved it on ourselves” move.

Where Agents Land First

One detail that stood out: DXC’s release says an “AI Advisor Agent” is used by more than 40,000 engineers. That hints at the adoption shape you should expect: agents land first where the work is already semi-structured, high-volume, and tool-rich.

Story 3: Kore.ai’s Multi-Agent Model (Designing a Small Team, Not One Bot)

Single-agent tools are tempting because they feel simple. One box. One persona.

But real operations aren’t one persona.

Kore.ai’s pitch is basically: treat the system like a small team. Multiple specialized agents, coordinated by an orchestration layer, with built-in memory, tools, and governance.

They also spell out multi-agent orchestration: agents that collaborate, share context, use tools, and move from human-in-the-loop to higher autonomy over time.

The Design Lesson That Holds

No matter which platform you choose, this architecture matters:

• One agent to handle intake and clarification
• One agent to retrieve facts from trusted sources
• One agent to execute changes in systems
• One agent to review and escalate edge cases

That’s not “extra.” That’s how you avoid a single overpowered agent doing sloppy work at scale.

Story 4: Gartner’s 2028 Forecast (Where This Is Heading, Fast)

Gartner’s numbers are blunt:

• By 2028, 33% of enterprise software applications will include agentic AI (up from under 1% in 2024)
• That shift could put about 15% of day-to-day work decisions on an autonomous path

Translation: you won’t “decide to adopt agents” once.

You’ll wake up and find them embedded in your procurement tool, your CRM, your ticketing system, your finance suite. Default features. Default behaviors.

So the real question is: which decisions will you delegate—and how will you supervise them?

Story 5: Microsoft AI Chief’s 12–18 Month Claim (How to Read Bold Timelines)

Mustafa Suleyman (Microsoft AI CEO) has been quoted as saying most—possibly all—white-collar tasks could be automated within 12 to 18 months.

That’s a spicy claim. Also, it’s strategically revealing.

Even if the timeline is aggressive, it tells you what Microsoft is building toward: agents that don’t just draft. They complete.

How to Read It Without Panic

• “Tasks” automate earlier than “jobs”
• The messy middle is oversight, permissions, and exceptions
• Automation expands fastest where decisions are already rule-bound

Mini-Playbook: 6 Decisions to Make Before You “Hire” Your First Agent

If you want agents to feel like employees (and not like chaos), decide these up front:

1. What Counts as Success?

Pick one metric a human actually cares about: cycle time, error rate, backlog age, SLA misses.

2. What Can the Agent Touch?

Read-only first. Then limited write actions. Then broader permissions.

3. What’s the Boundary of Autonomy?

Spell it out in plain language:

• “Agent can submit the form, but can’t approve it.”
• “Agent can draft the email, but a human hits send.”

4. Where Does Truth Live?

Agents fail in predictable ways when they’re forced to guess. Decide the trusted sources (systems of record) and make everything else “helpful, not authoritative.”

5. What’s the Audit Trail?

If you can’t reconstruct what happened, you can’t scale it.

6. Who Is on the Hook?

Not “the AI team.” Name an owner per workflow, like you would for any production system.

A Sane Rollout Path (Pilot Without Getting Stuck in Pilot)

Step 1: Choose One Workflow

Pick something with lots of repetition and a clear finish line. Think onboarding packets, reconciliations, ticket triage, or due diligence checklists.

Step 2: Start with a “Shadow Agent”

It does the work and produces a recommendation, but humans execute.

Step 3: Graduate to Guarded Action

Let the agent take low-risk actions (create a ticket, update a field, request missing info).

Step 4: Add a Second Agent

One executes, one checks. This is where quality starts feeling real.

Step 5: Standardize

If you can’t templatize it, you won’t be able to scale it.

What to Watch Next (Signals, Not Hype)

A few tells that your org is crossing from copilot to coworker:

• People start naming the agent in meetings (annoying, but true)
• The agent has a queue
• Exceptions become the main work humans do
• Audit and access control get discussed before model quality