Coder Legion - Recent posts tagged aimanagement

Your Next Direct Report Might Be an Agent

Mon, 15 Jun 2026 13:00:00 +0000

Not metaphorically. Not eventually. Right now, in some organizations, managers are already responsible for agents the way they're responsible for people — assigning work, reviewing output, correcting errors, and deciding when to escalate.

That shift is happening faster than most job descriptions have caught up to. And it raises questions worth thinking through before you find yourself managing an agent without a framework for doing it.

What It Actually Means to Manage an Agent

Managing a human direct report involves a predictable set of responsibilities. You set expectations. You review work. You give feedback. You decide what they should handle independently and what needs your input. You hold them accountable for outcomes.

Managing an agent involves the same responsibilities — with a few important differences.

Feedback doesn't work the same way. You can't have a conversation with an agent about why its output missed the mark. You adjust its instructions, its scope, or its inputs. The correction mechanism is configuration, not coaching.

Accountability is also different. When an agent makes a mistake, the accountability doesn't sit with the agent. It sits with whoever deployed it, scoped it, and let it run. That's you.

And unlike a human employee who can recognize when a situation is outside their experience and ask for help, an agent will keep going until its scope runs out or someone catches it. The boundaries you set are the only guardrails it has.

The Skills That Transfer — and the Ones That Don't

Most of what makes someone a good manager of people transfers directly to managing agents. Clear communication. Defined expectations. Regular review of output. Knowing when to intervene.

What doesn't transfer cleanly is intuition. A manager with ten years of experience has developed instincts for when something feels off — a team member who seems disengaged, a project that's quietly going sideways. That instinct comes from reading signals that agents don't send.

With agents, you replace intuition with instrumentation. You build in monitoring. You set thresholds. You review logs rather than reading the room. The information is there — it just looks different than what you're used to.

The managers adapting fastest are the ones who recognize this shift early. They stop waiting to feel like something is wrong and start building systems to tell them.

The Organizational Question Nobody Is Asking Yet

Here's the question most organizations haven't gotten to: when an agent is doing the work of a human role, who manages it, and where does that responsibility live on an org chart?

Right now, the answer is usually whoever built the agent, or whoever owns the system it runs in. That's a fine answer for a pilot. It doesn't hold up at scale.

As agents take on more responsibility — handling customer interactions, processing approvals, managing workflows — the question of ownership becomes a governance question. Who reviews agent performance? Who authorizes changes to scope? Who gets called when something goes wrong at 2 a.m.?

These aren't technology decisions. They're organizational design decisions. And the organizations that make them deliberately, before they're forced to, will be in a better position than the ones that figure it out reactively.

What This Means for Developers and Engineers

If you're building the agents, you're already closer to this reality than most. You've seen what agents can and can't do. You know where the edges are.

That puts you in a useful position — not just as the person who builds the system, but as someone who can help the business think clearly about what managing it actually requires. The organizations getting this right tend to have engineers who stay engaged after deployment, not just during build.

The feedback loop between "how it was built" and "how it's being managed" is where most production problems get caught early. That loop only works if the people who built it are still in the conversation.

The Bigger Picture

The idea of a direct report who doesn't sleep, doesn't have bad days, and can handle hundreds of transactions simultaneously sounds appealing in the abstract. In practice, it comes with real management overhead — just different overhead than you're used to.

The managers who thrive in this environment won't be the ones who treat agents like software. They'll be the ones who bring the same discipline to managing agents that they bring to managing people: clear expectations, consistent review, and genuine accountability for outcomes.

That's not a new skill. It's an existing skill applied to a new kind of team member.

The Next Leap in AI Isn’t Bigger Models: It’s Better Interfaces

Fri, 19 Dec 2025 02:49:23 +0000

For the last few years, the AI industry has been fixated on scale.

Bigger models.
More parameters.
Longer context windows.

That focus made sense, for a while.

But after working with AI systems across real products, teams, and day-to-day business workflows, one thing has become very clear to me:

The next real leap in AI will not come from bigger models.
It will come from better interfaces.

The bottleneck has moved.
And most teams are still looking in the wrong place.

Intelligence Isn’t the Problem Anymore

Today’s models are already capable of a lot.

They can reason across domains.
They can work with ambiguity.
They can generate high-quality output consistently.

Yet when people actually try to use AI in real work, the experience often feels:

mentally draining
fragile
overly dependent on perfect prompts
hard to trust at scale
difficult to integrate into existing workflows

That gap isn’t because the models aren’t good enough.

It’s because the way we interact with them hasn’t evolved at the same pace.

Why Chat-Based AI Is Starting to Feel Limiting

Most AI products still rely on a simple interaction loop:

You type something.
The system responds.
You adjust.
You repeat.

That was exciting when AI felt new.

But once AI moves from experimentation to operations, this model starts to break down.

I see the same issues repeatedly:

users are forced to “think in prompts”
context gets lost between sessions
decisions feel powerful but risky
outputs need constant double-checking
workflows fall apart under real usage

At that point, AI stops feeling like leverage and starts feeling like overhead.

This isn’t a productivity issue.
It’s an interface issue.

Interfaces Are Becoming the Real System

The future of AI interaction is not more conversation.

It’s better structure.

The most effective AI systems I’ve seen don’t feel like chatbots.

They feel like well-designed environments for thinking and decision-making.

Over time, a few patterns stand out.

Moving From Prompts to Intent

Most users don’t want to “prompt” an AI.

They want outcomes.

They want to express:

what they’re trying to achieve
what constraints matter
what risks are acceptable

Good interfaces capture intent and translate it into system behavior.

When that happens, prompt engineering disappears from the user’s workload and becomes part of the system design.

That’s exactly where it belongs.

Context That Actually Carries Forward

AI that forgets forces users to start over every time.

And that doesn’t scale.

The systems that work well maintain continuity:

past decisions
preferences
domain rules
business context

When context carries forward, intelligence compounds. When it doesn’t, every interaction feels like déjà vu.

This is the line between a helpful tool and something you can actually rely on.

Automation That Respects Judgment

Blind automation breaks trust.

Strong systems do something more subtle:

they show confidence levels
they surface trade-offs
they allow overrides
they make escalation easy

AI proposes. Humans decide.

Every AI system I’ve seen scale successfully preserves this balance.

Once judgment is removed, trust disappears shortly after.

Why Interfaces Will Matter More Than Models

Models will continue to improve. They’ll also continue to commoditize.

Access to intelligence is no longer rare.

What is rare is an interface that:

reduces cognitive load
fits naturally into how people work
hides complexity instead of exposing it
makes AI feel dependable, not impressive

That’s where the real differentiation is forming.

In the next phase of AI, interfaces—not models—will decide which products people actually adopt and keep using.

What Many Teams Still Miss

When people hear “better interfaces,” they often think:

nicer UI
cleaner dashboards
faster responses

That’s not enough.

The deeper shift is this:

AI interfaces are turning into decision environments.

They shape how people:

think through problems
delegate responsibility
evaluate risk
trust outcomes

This isn’t a UI problem. It’s a systems design problem.

Where This Is Going

Model improvements will continue, but they’ll feel incremental.

Interface improvements will feel transformational.

The teams that recognize this early will stop chasing scale for its own sake and start designing AI that fits naturally into human judgment and real work.

The next leap in AI won’t be louder.
It won’t be flashier.

It will be quieter.
Calmer.
More trustworthy.

And it will happen at the interface layer.

34,000 SAP customers are using AI agents. The hard part isn't deployment. It's what comes after

Thu, 06 Nov 2025 12:30:35 +0000

The Agent Governance Problem: What 34,000 SAP Customers Are Learning

The Number That Masks the Real Story

Brenda Bown, CMO of SAP Business AI Marketing, shared adoption data at TechEd: "34,000 customers are using AI" capabilities from SAP.

That's significant growth less than two years after Joule's launch. But the number doesn't reveal what matters most: what happens when you have dozens of agents running across your enterprise?

SAP introduced 22 new Joule Agents at TechEd, bringing the total to over 40 across finance, HR, supply chain, and other functions. Add partner-built agents (Bharat Sandhu said "40 by end of year") and custom agents companies are building themselves, and the number multiplies quickly.

This creates a governance problem the industry hasn't solved: how do you manage, monitor, and optimize autonomous AI agents at scale?

SAP's Solution: The AI Agent Hub

The SAP LeanIX AI Agent Hub is now generally available. It's the first enterprise-grade attempt to solve agent governance across an organization.

The hub provides a centralized cockpit to discover, govern, manage, and optimize all AI agents—whether built by SAP, third parties, or custom-developed internally.

Technical capabilities include:

Agent Discovery and Inventory: Visibility into every agent running across the enterprise. Integrations with Google Cloud and Microsoft (planned Q4 2025) will extend visibility beyond SAP systems.

Business Capability Mapping: IT managers can map agents to business capabilities, uncover deployment opportunities across functions, and align each agent to enterprise outcomes.

Lifecycle Management: Organizations can manage agents in the context of their enterprise architecture, tracking deployment status, performance, and business impact.

MCP Server Management: Capabilities to discover and manage MCP servers (expected Q1 2026) provide visibility into the tools and data sources agents can access.

The hub addresses a reality most enterprises are just beginning to face: agents are easier to deploy than to govern.

Where Agents Are Actually Being Used

When asked which industries or use cases are seeing the fastest AI adoption, Brenda Bown identified four areas:

HR

"People Intelligence Agents help HR teams address organizational challenges by spotting trends in people data and recommending actions for retention, compensation, and skills distribution."

The Career and Talent Development Agent transforms succession planning by prompting managers for updates, recommending successors, and generating personalized development plans. The Payroll Agent helps employees understand their pay and helps administrators strengthen payroll accuracy.

Service and Support

"Digital Service Agents increase customer contact center efficiency by providing accurate responses based on customer context and knowledge bases."

The HR Service Agent was specifically highlighted as saving "time and cost spent on HR policy inquiries by reasoning over employee questions and providing immediate, accurate answers."

Supply Chain and Manufacturing

"Production Planning and Operations Agents automate production order release by checking material, capacity, and scheduling availability."

The Supplier Onboarding Agent accelerates supplier onboarding by orchestrating invitations, monitoring supplier progress, and handling escalations.

Finance

"Cash Management Agents automate reconciliation tasks and identify potential cash shortages and surpluses."

The International Trade Classification Agent helps ensure regulatory compliance by classifying goods for shipping and proposing customs tariff numbers.

The pattern: agents are being deployed first for operations that combine high volume, repetitive decision-making, and access to structured data.

The Trust Progression

Muhammad Alam emphasized throughout the keynote that autonomous operation requires establishing trust first: "After you earn trust and accuracy, and efficiency, you let the agents run autonomously."

The practical implementation path has three stages:

Stage 1: Assistant Mode

AI suggests actions. Humans review and approve before anything happens. This is where most of the 34,000 customers are today.

Brenda Bown confirmed: "Work with assistants and agents. Always be in the loop. Less prone to errors."

Stage 2: Validation Mode

AI takes actions automatically. Humans review afterwards and can intervene if needed. Some customers are beginning to move into this mode for low-risk operations.

Stage 3: Autonomous Mode

AI operates independently within defined guardrails. Only exceptions get escalated to humans. This is still rare in production environments.

Alam was clear about the progression: "This is the job of Joule. After you earn trust and accuracy, and efficiency, you let the agents run autonomously."

What's Surprising About Agent Use

When asked what's surprising about how customers are using Joule Agents, Bown pointed to two things:

Extensibility

"Extending the agents we ship. Building their own agent."

Customers aren't just using SAP's pre-built agents as-is. They're extending them for company-specific processes. A procurement agent might be extended to check company-specific supplier approval workflows. An HR agent might be modified to incorporate organization-specific policies.

Joule Studio (general availability December 2025) is specifically designed to support this. Developers can extend SAP-delivered Joule Agents and build new agents grounded in SAP business data and context.

Agent Interoperability

"The pace of development has been a pleasant surprise, as well as agent interoperability."

Agents working together across departments represents early movement toward what SAP calls the "agentic enterprise."

The BMW example came up repeatedly. Muhammad Alam mentioned it in the keynote: "BMW Agentic Enterprise. Five BMW Groups already creating agents. Working with agents across departments."

The Agent2Agent Protocol

SAP is contributing to the Agent2Agent (A2A) interoperability protocol, which establishes a foundation for AI agents to securely collaborate across platforms.

The first implementation is expected to be released in Q4 2025.

This matters because agents built by different teams, on different platforms, using different frameworks need to interact. A supply chain agent from SAP needs to coordinate with an inventory agent built by a partner. An HR agent needs to pass information to a finance agent for budget planning.

The A2A protocol provides standardized communication so agents can work together without custom integration for every interaction.

The 22 New Joule Agents

SAP introduced agents across seven business areas at TechEd:

Business Transformation Management (5 agents):

Screen Guide Agent: Provides live on-screen guidance navigating SAP Signavio
Value Case Creation Agent: Converts process mining insights into structured business cases
Dashboard Analyzer Agent: Interprets process mining dashboards to identify trends
Process Content Recommender Agent: Recommends relevant process content from best practices
Workspace Administration Agent: Enables faster user onboarding with correct access levels

Finance (2 agents):

International Trade Classification Agent: Classifies goods for shipping with customs codes
Cash Management Agent: Automates bank statement reconciliation

Spend (2 agents):

Bid Analysis Agent: Generates comprehensive bid analysis summaries
Receipt Analysis Agent: Fills gaps in expense entries and verifies vendor details

Supply Chain (3 agents):

Production Planning and Operations Agent: Automates production order release
Change Record Management Agent: Recommends next steps for engineering changes
Supplier Onboarding Agent: Orchestrates supplier onboarding and monitors progress

Human Capital Management (4 agents):

HR Service Agent: Provides immediate answers to HR policy inquiries
People Intelligence Agent: Spots trends in people data and recommends actions
Career and Talent Development Agent: Generates personalized development plans
Payroll Agent: Helps employees understand pay and detects payroll issues

Customer Experience (1 agent):

Digital Service Agent: Provides accurate customer service responses with context

Cloud ERP and Industries (1 agent):

Utilities Customer Self-Service Agent: Delivers tailored answers for utility customers

General availability for all agents is expected Q1 2026.

When asked how customers decide which agents to deploy first, Bown said it "depends on the business function" but emphasized solving "complex business process - HR, supply chain."

The ROI Question

When asked about average ROI from implementing agents, Bown gave the honest answer: "It depends on the customer."

While not satisfying, it reflects reality. ROI varies based on:

How many people the agent affects
What manual processes it eliminates
How much time it saves per interaction
Whether it enables new capabilities or just accelerates existing ones

SAP has updated its ROI estimator for Business AI to help customers model returns. The tool calculates projected ROI based on specific AI features and Joule Agents customers plan to use, explains assumptions, and lets customers request customized quotes.

The tool is generally available now.

But the data from Muhammad Alam's fireside chat suggests significant potential: "7x to 12x value with AI in sprints as learning increases."

That's productivity improvement, not cost savings. Alam framed it as acceleration: "40,000 workers with enough backlog for 200,000. AI is going to be about acceleration."

The Complexity Perception

Bown noted that the biggest misconception customers have about implementing SAP's AI tools is "complexity" and "perception."

This is partly self-inflicted by SAP and other enterprise software vendors. When you announce 22 new agents, multiple deployment paths, evolving architectures, and dozens of partner integrations at a single event, you create genuine confusion.

Bharat Sandhu acknowledged this: "While it's complicated to learn, gen AI will help with learning, adoption, and ROI."

The irony: generative AI is supposed to simplify learning SAP's platform. But that assumes customers have already deployed the AI tools to help them learn. It's a chicken-and-egg problem.

SAP's practical response: free training through the Coursera partnership. SAP committed to equipping 12 million people worldwide with AI skills by 2030.

Brenda Bown explained the rationale: "Equip customers, and the population in general, with AI skills. Upskill the workforce of the future."

The commitment builds on SAP's success training over 4 million people since 2022.

What's Available Now vs. What's Coming

This timeline matters for organizations planning agent deployments:

Available Now:

40+ Joule Agents across business functions (most in production)
SAP LeanIX AI Agent Hub for governance
Agent builder in Joule Studio (general availability December 2025)
Pro-code agent development on SAP BTP

Q4 2025:

Agent2Agent protocol implementation
Integrations with Google Cloud and Microsoft for broader agent discovery
Additional Joule Agents reaching general availability

Q1 2026:

22 new Joule Agents announced at TechEd
System-triggered agents in Joule Studio
Extensibility of SAP-delivered Joule Agents
MCP server management in AI Agent Hub
Centralized agent monitoring

H1 2026:

Additional governance and monitoring capabilities
Expanded partner-built agent ecosystem

When asked what customers can do if they want to start building agents this week, Bharat Sandhu said: "Go to build. Everything is generally available in a couple of weeks."

The Differentiation Question

SAP faces competition from Microsoft (Copilot), Salesforce (Einstein), Oracle, and others. When customers ask "why SAP AI vs. building our own with ChatGPT," Bown's answer focuses on three factors:

Trust: SAP provides governed, secure AI that meets enterprise compliance requirements.

Governance: Centralized management through the AI Agent Hub provides visibility and control that's difficult to achieve with custom-built agents.

Data Breadth and Depth: "Breadth and depth of the data" that comes from being embedded in business applications. Agents built on SAP have immediate access to business context that would take significant effort to replicate elsewhere.

The implicit argument: building agents on generic LLMs might be faster initially, but enterprises eventually need governance, security, and business context. SAP provides that foundation.

What Governance Actually Looks Like

For organizations deploying agents now, SAP's governance approach has several components:

Inventory Management: Track every agent (SAP, partner, custom) in one place through the AI Agent Hub.

Business Alignment: Map agents to business capabilities so you understand which processes are augmented by AI.

Performance Monitoring: Track agent behavior, success rates, and business impact.

Access Control: Define what data and systems each agent can access through MCP gateways and integration governance.

Lifecycle Management: Manage agent versions, updates, and retirement in the context of enterprise architecture.

The governance tooling is maturing, but it's available now. Organizations don't need to wait for perfect governance to start deploying agents.

What Comes Next

The 34,000 customer adoption number indicates SAP's AI capabilities have moved beyond pilot projects for a meaningful portion of the customer base.

The next phase—moving from assisted to autonomous operation, scaling from a few agents to dozens, proving measurable business value—is where the industry will learn whether agentic AI delivers on its promise.

SAP's governance approach through the AI Agent Hub provides infrastructure to support that scale. Whether it's sufficient will become clear as more customers move from testing to production deployments over the next six months.

For developers and architects, the practical question is: how many agents can you realistically deploy and manage before governance complexity overwhelms the productivity gains?

The answer will vary by organization, but the fact that SAP built governance infrastructure before fully scaling agent deployments suggests they're taking the question seriously.