Showing posts with label AI Workflow Design. Show all posts
Showing posts with label AI Workflow Design. Show all posts

Wednesday, September 16, 2026

AI That Can See, Think and Act: The Convergence of Multimodal and Agentic AI

 

AI That Can See, Think and Act: The Convergence of Multimodal and Agentic AI

For years, we interacted with AI mainly through text.

We typed a question.
AI generated an answer.

Then AI learned to understand images, audio, video, documents, and screens. That was the rise of multimodal AI.

Now something even more interesting is happening.

AI is increasingly being designed not just to understand information, but to reason about it and take action.

This is where multimodal AI and agentic AI begin to converge.

The next generation of AI systems won't simply answer:

"What am I looking at?"

They will increasingly be able to answer:

"What is happening, what should happen next, and what action should I take?"

That shift could fundamentally change how we build software and automate business processes.


From Chatbots to Systems That Can Act

Traditional AI applications generally follow a simple pattern:

Input → AI → Response

For example:

A user uploads an invoice and asks:

"What is the total amount?"

The AI reads the invoice and provides the answer.

An agentic system can go much further:

Observe → Understand → Reason → Plan → Act → Verify

For the same invoice, an AI agent might:

  1. Read the invoice.
  2. Identify the supplier.
  3. Extract invoice details.
  4. Compare them with a purchase order.
  5. Check whether the goods were received.
  6. Detect a mismatch.
  7. Ask for human approval if necessary.
  8. Update the accounting system.
  9. Notify the appropriate person.
  10. Verify that the transaction was completed successfully.

The AI is no longer simply answering a question.

It is participating in the workflow.


What Does "Multimodal" Really Mean?

Multimodal AI refers to systems capable of working with multiple types of information.

That can include:

  • Text
  • Images
  • Audio
  • Video
  • Documents
  • Screenshots
  • Tables
  • Code
  • Structured business data
  • Application interfaces

Consider a manufacturing environment.

A traditional AI system might analyze machine sensor data.

A multimodal system could combine:

Sensor data + maintenance history + technician notes + machine images + video

Now the AI has a much richer understanding of the situation.

Instead of seeing one isolated data point, it can connect different forms of evidence.


Agentic AI Adds the "Action" Layer

Multimodal AI gives an AI system more ways to perceive the world.

Agentic AI gives it the ability to pursue a goal through multiple steps.

This distinction is important.

Imagine an IT operations environment.

A monitoring system detects:

"Database CPU utilization is 95%."

That's useful, but limited.

A multimodal agent could potentially examine:

  • Monitoring dashboards
  • SQL Server logs
  • Query execution statistics
  • Recent deployment information
  • Alert history
  • Configuration changes
  • Incident tickets

It could then reason:

"CPU increased shortly after a deployment. Query X is consuming unusually high CPU. The execution plan changed after the deployment."

Depending on the permissions and guardrails, the agent could then:

  • Create an incident.
  • Notify the DBA team.
  • Collect diagnostic information.
  • Recommend a rollback.
  • Request approval.
  • Execute an approved remediation.
  • Monitor the result.

This is a fundamentally different architecture from a chatbot.


The New AI Loop

A useful way to think about this emerging architecture is:

1. See

The system receives information from different modalities.

Images, documents, video, audio, APIs, databases, screens, sensors and text.

2. Understand

The AI converts those inputs into a meaningful representation of the current situation.

3. Think

The reasoning layer determines:

  • What is happening?
  • What is the objective?
  • What information is missing?
  • What options are available?
  • What could go wrong?

4. Plan

The agent breaks the objective into smaller tasks.

For example:

Goal: Resolve a failed database job.

Plan:

Check job history → identify error → inspect dependencies → determine impact → propose remediation → obtain approval → execute → validate.

5. Act

The agent interacts with external systems through tools and APIs.

6. Verify

This step may become one of the most important.

The agent shouldn't simply assume that an action succeeded.

It should check.

Action → Result → Verification → Next action

That feedback loop is what makes agentic systems considerably more powerful than simple automation.


Why Multimodal + Agentic AI Is More Powerful

Individually, both technologies are useful.

Together, they create something much more interesting.

Consider a customer-support scenario.

A customer uploads a photograph showing a damaged product and writes:

"This arrived broken."

A multimodal AI can understand the photograph, read the order information and interpret the customer's message.

An agent can then potentially:

  1. Identify the product.
  2. Retrieve the order.
  3. Check the purchase date.
  4. Review the warranty or return policy.
  5. Determine eligibility.
  6. Create a return request.
  7. Generate a shipping label.
  8. Update the CRM.
  9. Notify the customer.
  10. Track the resolution.

The image provides perception.

The agent provides reasoning and action.

Together they create an end-to-end workflow.


The Enterprise Opportunity

This convergence could be particularly important for enterprises because businesses are full of processes that require people to:

Look → Understand → Decide → Act → Verify

Examples include:

Finance

Invoice → Purchase Order → Goods Receipt → Payment

IT Operations

Alert → Logs → Diagnosis → Remediation → Validation

HR

Resume → Job Requirements → Candidate Screening → Interview → Scheduling

Supply Chain

Order → Shipment → Warehouse → Inventory → Delivery

Customer Service

Message → Account → Product → Policy → Resolution

Many of these workflows contain repetitive cognitive tasks.

That's where AI agents could potentially create significant value.


But There Is a Major Difference Between "Can" and "Should"

This is where the conversation around AI agents becomes more serious.

Just because an AI can perform an action doesn't mean it should perform that action autonomously.

Imagine an agent that has access to:

  • Production databases
  • Financial systems
  • Customer records
  • Email
  • Cloud infrastructure
  • Internal applications

A mistake could have real consequences.

Therefore, enterprise agentic AI needs more than a powerful model.

It needs:

Permissions + Policies + Guardrails + Observability + Human Oversight

A useful architecture might look like:

                ┌───────────────────┐
                │   Human / Goal    │
                └─────────┬─────────┘
                          │
                          ▼
                ┌───────────────────┐
                │   AI Agent        │
                │ Reason + Plan     │
                └─────────┬─────────┘
                          │
              ┌───────────┼───────────┐
              ▼           ▼           ▼
          Multimodal    Memory       Tools
          Perception                 / APIs
              │           │           │
              └───────────┼───────────┘
                          ▼
                ┌───────────────────┐
                │ External Systems  │
                └─────────┬─────────┘
                          │
                          ▼
                ┌───────────────────┐
                │ Verify + Monitor  │
                └───────────────────┘

The important component isn't just the model.

It is the entire system around the model.


Human-in-the-Loop Will Still Matter

The future of agentic AI is not necessarily:

"AI does everything."

A more realistic enterprise model is:

AI handles routine decisions; humans handle exceptions, approvals and high-impact decisions.

For example:

Low risk

AI can automatically categorize an expense.

Medium risk

AI prepares a payment but requires approval.

High risk

AI analyzes the situation and recommends an action, but a human must approve it.

This creates a spectrum of autonomy rather than a simple "AI vs human" model.


The Rise of AI Workflow Designers

This shift also changes what AI skills will be valuable.

Prompt engineering was an important early skill.

But building useful agentic systems requires much more.

People increasingly need to understand:

  • Workflow design
  • APIs
  • Tool calling
  • Data retrieval
  • RAG
  • Memory
  • Authentication
  • Permissions
  • Observability
  • Evaluation
  • Error handling
  • Human approval workflows
  • Security

In other words:

The important skill is moving from writing better prompts to designing better AI systems.

A good AI workflow designer doesn't simply ask:

"What should the AI say?"

They ask:

"What should the AI observe, decide, do, verify and escalate?"


What Happens to Traditional Software?

AI agents won't necessarily replace traditional applications.

Instead, we may see a new software architecture where conventional applications provide reliable systems of record, while AI agents provide a natural-language reasoning and orchestration layer.

For example:

User
  ↓
AI Agent
  ↓
Reasoning / Planning
  ↓
 ┌──────────────┬───────────────┐
 ↓              ↓               ↓
Database       API           Enterprise App
 ↓              ↓               ↓
Result ─────────┴───────────────┘
              ↓
          AI verifies
              ↓
           User

The database still stores the data.

The ERP still manages transactions.

The CRM still manages customers.

The AI becomes the layer that understands the user's objective and coordinates the systems.


The Real Challenge: Trust

The biggest challenge may not be intelligence.

It may be trust.

Before organizations allow AI agents to operate production systems, they will need answers to questions such as:

  • Why did the agent make this decision?
  • Which data did it use?
  • Which tools did it access?
  • What actions did it take?
  • Who authorized those actions?
  • Can we reproduce the decision?
  • What happens when the model is wrong?
  • Can we stop the agent?
  • Can we roll back its actions?

This makes AI governance and observability just as important as model capability.


From AI Assistant to AI Coworker

Perhaps the simplest way to understand the transition is this:

Generation 1

AI as a search assistant

"Find this information."

Generation 2

AI as a content assistant

"Create this document."

Generation 3

AI as a reasoning assistant

"Analyze this situation."

Generation 4

AI as an agent

"Achieve this goal."

Emerging Generation 5

AI as a multimodal agent

"Understand everything relevant, decide what needs to happen, execute the workflow, and verify the outcome."

The boundaries between software, automation and AI are becoming increasingly blurred.


The Bigger Picture

The interesting question is no longer simply:

"How intelligent is the AI?"

A better question is:

"What can the AI perceive, what can it reason about, what can it access, and what can it safely do?"

Multimodal AI expands the AI's perception.

Reasoning models expand its problem-solving capability.

Agentic architectures expand its ability to act.

Tools and APIs expand its reach.

Memory expands its continuity.

Governance determines its boundaries.

Put these together and we get a new class of software:

AI systems that can see, think, act and learn from the results of their actions.

That could be one of the defining directions of enterprise AI over the next few years.

And perhaps the biggest shift is this:

We are moving from AI that generates answers to AI that participates in work.

Can AI Agents Run a Business Process Without Human Intervention?

 

Can AI Agents Run a Business Process Without Human Intervention?

Imagine starting your workday and discovering that several routine business processes have already been completed.

Customer emails were classified.
Inventory was checked.
Low-stock products were identified.
Invoices were matched.
Reports were generated.
Support tickets were categorized.
And nobody had to manually perform those steps.
This sounds like automation, but there is an important difference.
Traditional automation follows predefined rules.

AI agents can potentially understand a goal, make decisions, use tools and adapt their actions based on what they discover.

That raises an interesting question:

Can AI agents actually run a business process without human intervention?

The short answer is:

For some well-defined processes, yes. But full autonomy is not appropriate for every business process.

The real challenge is deciding where AI should act independently and where humans should remain involved.


What Makes an AI Agent Different From Automation?

Traditional automation generally follows a predetermined sequence.

For example:

If invoice received → extract invoice number → store invoice → send notification.

The rules are defined in advance.

An AI agent can potentially operate with more flexibility.

For example:

"Process incoming supplier invoices and identify anything that requires attention."

The agent might:

  • Read the invoice.
  • Identify the supplier.
  • Extract the relevant information.
  • Check purchase-order information.
  • Compare the invoice with receiving records.
  • Identify discrepancies.
  • Decide whether the invoice meets predefined business rules.
  • Route exceptions to the appropriate person.
  • Update the accounting workflow.

The agent is not simply executing one fixed script.

It is interpreting information and selecting actions within the permissions it has been given.


The Three Levels of AI Business Automation

Not every business process needs the same level of AI autonomy.

A useful way to think about AI automation is through three levels.

Level 1: AI Assistant

The AI provides information or recommendations.

Human → AI → Human Decision

Example:

"Analyze this month's sales and identify products with declining demand."

AI produces the analysis.

A person makes the decision.


Level 2: AI Copilot

The AI performs several parts of the workflow but requires approval for important actions.

Human → AI → Recommendation → Human Approval → Action

Example:

AI identifies products that should be reordered and prepares purchase orders.

The business owner reviews them.

After approval, the orders are created.


Level 3: Autonomous AI Agent

The AI can perform the complete workflow within predefined boundaries.

Goal → AI Agent → Tools → Decisions → Actions → Validation

For example:

"Monitor inventory and reorder approved products when stock falls below the defined threshold."

The agent can continuously monitor inventory and take action without asking for approval every time.

However, this only works safely when the process has clearly defined rules and appropriate controls.


What Business Processes Are Suitable for AI Agents?

Some processes are much easier to automate than others.

Good candidates typically have:

  • Clearly defined objectives
  • Repeatable workflows
  • Reliable data
  • Limited consequences when something goes wrong
  • Clear business rules
  • Well-defined permissions
  • Measurable outcomes

Examples include:

Customer Support Classification

AI receives incoming support tickets and categorizes them.

Document Processing

AI extracts information from invoices, forms and documents.

Internal Reporting

AI collects approved data and generates scheduled reports.

Inventory Monitoring

AI checks stock levels and identifies products requiring attention.

Data Quality Checks

AI identifies incomplete or inconsistent records.

IT Monitoring

AI analyzes alerts and prepares incident summaries.

Meeting Administration

AI can summarize meetings, extract action items and update task systems.

These are examples where AI can potentially operate with limited human involvement, depending on the organization's requirements.


What Processes Should Usually Keep Humans Involved?

Some decisions have consequences that make full autonomy inappropriate.

For example:

  • Large financial transactions
  • Employee termination decisions
  • Legal decisions
  • Sensitive customer disputes
  • Production infrastructure changes
  • Medical decisions
  • Security incident response
  • Access to highly sensitive information

In these situations, AI can still be extremely useful.

But instead of:

AI → Action

the workflow can be:

AI → Recommendation → Human Review → Action

This is commonly called human-in-the-loop design.


A Practical Example: E-Commerce Inventory

Consider an online business selling kitchen products.

The business has:

  • 500 SKUs
  • Multiple suppliers
  • Different reorder thresholds
  • Different supplier lead times
  • Seasonal demand

A simple inventory automation might say:

If stock < 20, send an alert.

An AI agent could potentially do more.

It might:

  • Monitor inventory.
  • Analyze recent sales.
  • Check historical demand.
  • Check supplier lead time.
  • Identify products approaching stock-out.
  • Calculate a suggested reorder quantity.
  • Check whether the supplier is approved.
  • Prepare the purchase order.
  • Execute the order if it falls within predefined limits.
  • Escalate unusual cases.

The business owner doesn't necessarily need to review every routine reorder.

But the workflow can be designed so that large or unusual purchases require approval.


Example: IT Operations

Now consider a database environment.

A monitoring platform reports:

"Database CPU utilization has exceeded 90%."

An AI agent could potentially:

Step 1

Analyze the alert.

Step 2

Check recent performance data.

Step 3

Identify expensive queries.

Step 4

Compare the current situation with previous incidents.

Step 5

Generate a diagnosis.

Step 6

Recommend corrective action.

For low-risk actions, the system might be allowed to perform predefined remediation.

For high-risk operations, it could request administrator approval.

For example:

Restart a non-production service

→ Potentially automated.

Change a production database configuration

→ Human approval.

Delete production data

→ Strong human authorization.

The important concept is not simply automation.

It is controlled autonomy.


The AI Agent Needs More Than an AI Model

A business-ready AI agent usually needs more than an LLM.

A simplified architecture might look like:

Business Goal

AI Agent / Orchestrator

Reasoning Model

RAG / Business Knowledge

Tools & APIs

Business Applications

Databases

Validation

Audit & Monitoring

Human Approval When Required

The language model is only one component.

The surrounding system determines what the agent can actually do.


Permissions Become Extremely Important

Imagine giving an AI agent access to your company database.

What should it be allowed to do?

Can it:

  • Read data?
  • Insert data?
  • Update records?
  • Delete records?
  • Execute stored procedures?
  • Access customer information?
  • Access financial information?

These permissions should not be left to the AI's judgment.

They should be defined by the system.

A useful principle is:

Give an AI agent the minimum permissions it needs to perform its job.

This is the same principle commonly used in secure software architecture.


What Happens When the AI Makes a Mistake?

This is perhaps the most important question.

Suppose an AI agent incorrectly identifies a supplier invoice as valid.

If the agent only prepares a recommendation, a human can catch the mistake.

If the agent automatically pays the invoice, the consequences can be much greater.

Therefore, AI workflows should consider:

What happens if the AI is wrong?

Possible controls include:

  • Validation rules
  • Approval thresholds
  • Transaction limits
  • Rollback mechanisms
  • Audit logs
  • Human escalation
  • Monitoring
  • Exception handling

The objective isn't to assume AI will never make mistakes.

The objective is to design the system so that mistakes are contained and recoverable.


The Importance of Guardrails

Guardrails define what an AI agent is allowed to do.

For example:

An inventory agent might be allowed to create purchase orders up to:

₹25,000

Anything above that requires approval.

A support agent might be allowed to issue refunds up to:

₹1,000

Anything larger requires human review.

An IT agent might restart approved development services but not production databases.

These rules turn unrestricted autonomy into bounded autonomy.


Can AI Agents Make Decisions?

Yes—but the type of decision matters.

Consider three categories.

Rule-Based Decisions

"If inventory < reorder level, create reorder recommendation."

These are relatively straightforward to automate.

Analytical Decisions

"Which products appear to have declining demand?"

AI can analyze multiple data points and provide an assessment.

High-Impact Decisions

"Should this employee be terminated?"

This involves significant human, legal and organizational considerations and should not simply be delegated to an AI agent.

The important question is not:

"Can AI make a decision?"

It is:

"What decisions should AI be authorized to make?"


AI Agents Need a Stop Button

An autonomous system should have mechanisms to stop or pause execution.

For example:

Unexpected behavior detected

Agent pauses

Human notified

Human investigates

Workflow resumes or is terminated

This is particularly important for long-running agents.

If an agent can continuously execute tasks, the organization needs visibility into what it is doing.


Monitoring AI Agents

Traditional software monitoring typically focuses on:

  • CPU
  • Memory
  • Errors
  • Response time
  • Availability

AI agents require additional monitoring.

Organizations may need to track:

  • Which tools the agent used
  • Which data it accessed
  • What decisions it made
  • Which actions it performed
  • How many model calls occurred
  • How much the workflow cost
  • How often humans intervened
  • How often tasks failed
  • Whether outputs met expected quality

This creates an AI audit trail.


The Cost Question

Autonomous AI is not automatically cheaper.

An agent may need to:

  • Call a model multiple times
  • Search documents
  • Query databases
  • Call APIs
  • Validate results
  • Retry failed operations

A simple task that previously required one API call could become a multi-step workflow.

Therefore, businesses should measure:

Cost per task

rather than simply asking:

"How much does the AI model cost?"

The real cost includes infrastructure, tools, monitoring, human review and failures.


AI Agents Need Evaluation

A normal software application can often be tested with predictable inputs and expected outputs.

AI agents can behave differently because their decisions may involve probabilistic models and dynamic tool use.

Therefore, businesses need to evaluate:

  • Accuracy
  • Reliability
  • Tool selection
  • Policy compliance
  • Security
  • Failure handling
  • Cost
  • Latency
  • Human escalation

The goal is to determine whether the entire workflow works reliably—not simply whether the AI's response sounds good.


Autonomous Does Not Mean Unsupervised

This is an important distinction.

An AI agent can operate autonomously while still being supervised at the system level.

For example:

Agent

Runs routine inventory checks automatically.

Monitoring system

Tracks its activities.

Policy engine

Enforces limits.

Human

Handles exceptions.

This model can be more practical than requiring a person to approve every single action.


A Useful Model: Human + AI

The future of business automation may not be:

Humans vs AI

It may be:

Humans + AI Agents

Humans can focus on:

  • Strategy
  • Relationships
  • Exceptions
  • Judgment
  • Creativity
  • High-impact decisions

AI agents can handle:

  • Repetitive analysis
  • Information retrieval
  • Routine workflows
  • Monitoring
  • Data processing
  • Report preparation

The objective is to assign each type of work to the system best suited to handle it.


A Practical AI Autonomy Framework

Before allowing an AI agent to act independently, businesses can ask five questions:

1. Is the process clearly defined?

If nobody can explain the process, automating it may be premature.

2. Is the data reliable?

AI cannot compensate for consistently poor source data.

3. What happens if the AI is wrong?

The potential impact should be understood.

4. Can the action be reversed?

Reversible actions are easier to automate safely.

5. Does the agent have appropriate permissions?

The agent should only access and modify what it actually needs.

If these questions have good answers, the process may be a reasonable candidate for increased AI autonomy.


The Future: Controlled Autonomy

The next stage of enterprise AI may not be about giving AI unlimited control.

Instead, businesses are likely to focus on controlled autonomy.

That means:

Clear goals

Limited permissions

Business rules

Monitoring

Validation

Human escalation

=

Controlled AI Automation

This approach allows organizations to benefit from AI agents without treating them as completely independent decision-makers.


Final Thoughts

AI agents can already be designed to perform multi-step tasks using models, tools, data sources and business applications.

The more interesting question is not whether an AI agent can perform a business process.

The question is:

Should it perform that process independently, and under what controls?

For routine, measurable and relatively low-risk workflows, greater automation may make sense.

For high-impact decisions, human review may remain essential.

The future of enterprise AI is therefore unlikely to be simply:

"AI does everything."

A more practical model is:

AI handles what can be safely automated.

Humans handle what requires judgment, accountability and oversight.

The real competitive advantage may come from designing the right boundary between the two.

And perhaps the most important AI skill for businesses in the coming years will not be building an AI that can do everything.

It will be knowing what to let AI do—and what not to let it do.

Beyond Prompt Engineering: The New Skill Is AI Workflow Design

 

Beyond Prompt Engineering: The New Skill Is AI Workflow Design

For the last few years, one of the most popular skills in Generative AI has been prompt engineering.

People learned how to write better instructions for AI models.

They experimented with different prompts.

They learned about roles, context, examples, structured outputs and chain-of-thought-style techniques.

And it worked.

A well-designed prompt can make a significant difference to the quality and usefulness of an AI response.

But AI is changing.

In 2026, the interesting question is no longer only:

"How do I write a better prompt?"

It is increasingly:

"How do I design a workflow where AI can actually complete useful work?"

This is where AI Workflow Design becomes important.


What Is AI Workflow Design?

AI workflow design is the process of designing a sequence of tasks where AI, software tools, business data and humans work together to accomplish a specific objective.

Instead of thinking about one prompt, you think about the entire process.

For example:

User Request

AI understands the objective

Retrieve relevant information

Analyze information

Use an external tool

Validate the result

Ask for human approval if required

Complete the action

This is very different from simply sending a prompt to an AI model.


Prompt Engineering vs AI Workflow Design

The difference can be explained with a simple example.

Suppose a business owner says:

"Find products that are running low in inventory."

Prompt Engineering Approach

You provide an AI model with an inventory file and ask:

"Analyze this inventory and identify products that need to be reordered."

The AI returns a list.

Useful.

But the process ends there.

AI Workflow Approach

The workflow could be:

  1. Retrieve current inventory.
  2. Retrieve recent sales.
  3. Check reorder thresholds.
  4. Identify products approaching stock-out.
  5. Compare supplier lead times.
  6. Calculate suggested reorder quantities.
  7. Generate a purchase recommendation.
  8. Ask the business owner for approval.
  9. Create the purchase order after approval.
  10. Record the activity in the system.

Now AI is not simply generating an answer.

It is participating in a business process.


Why Prompt Engineering Alone Is Not Enough

A prompt controls what the model should do with the information it receives.

But real-world applications often require much more.

They need:

  • Data access
  • APIs
  • Databases
  • Business rules
  • Authentication
  • Permissions
  • Validation
  • Error handling
  • Monitoring
  • Human approval
  • Audit trails

A great prompt cannot solve all of these problems.

This is why AI application development is increasingly becoming an architecture problem, not just a prompting problem.


The New AI Stack

A modern AI workflow can contain several layers.

1. User Interface

Where the user provides the request.

2. AI Model

The model interprets the request and generates reasoning or content.

3. Context / RAG

Relevant information is retrieved from documents, databases or knowledge bases.

4. Tools

The AI can interact with APIs, databases or applications.

5. Workflow Engine

Controls the sequence of operations.

6. Validation

Checks whether the result is correct.

7. Human Approval

Required for sensitive or high-impact actions.

8. Monitoring

Records what happened and identifies failures.

The prompt is still important.

But it is only one component of the system.


Think in Tasks, Not Prompts

One of the biggest changes in mindset is moving from:

"What prompt should I use?"

to:

"What tasks need to happen?"

Consider an employee onboarding process.

Instead of asking an AI:

"Write an onboarding email."

you could design a workflow:

New Employee Record Created

AI reads employee information

Generate personalized welcome email

Retrieve onboarding documentation

Create checklist

Notify HR

Send email after approval

Now the AI is integrated into a workflow.


AI Agents Make Workflow Design More Important

The rise of AI agents makes workflow design even more relevant.

An AI agent can potentially:

  • Plan tasks
  • Use tools
  • Retrieve information
  • Execute actions
  • Evaluate results
  • Continue working
  • Ask humans for assistance

OpenAI describes agents as systems that can independently accomplish tasks on a user's behalf, while Anthropic describes agents as systems that direct their own processes and tool use while working toward a task. (openai.com) (anthropic.com)

That means developers need to think about what the agent is allowed to do, not just what it is instructed to do.


A Simple Example: IT Incident Management

Imagine a database server generates an alert.

Old Approach

Monitoring system:

CPU usage exceeded 90%.

Engineer:

Investigates manually.

AI-Assisted Workflow

Monitoring system:

AI analyzes the alert.

Retrieves server metrics.

Checks recent database activity.

Searches previous incidents.

Identifies possible causes.

Creates an incident summary.

Recommends troubleshooting actions.

Engineer reviews.

Approved action is executed.

This workflow combines:

Monitoring + AI + Data + RAG + Tools + Human Decision-Making

The prompt is still present somewhere in the system.

But it is no longer the whole solution.


AI Workflow Design and RAG

RAG, or Retrieval-Augmented Generation, is another important component.

Suppose a company has:

  • 5,000 internal documents
  • Technical manuals
  • SOPs
  • HR policies
  • Product documentation
  • Support tickets

Instead of putting everything into one prompt, an AI workflow can retrieve the relevant information when it is needed.

For example:

User Question

Determine information required

Search knowledge base

Retrieve relevant documents

AI analyzes retrieved information

Generate response

This is more scalable than manually adding large amounts of context to every prompt.


Workflow Design With Multiple AI Agents

The workflow becomes even more interesting when several specialized agents are involved.

For example:

Research Agent

Finds information.

Analysis Agent

Analyzes the information.

Validation Agent

Checks the result.

Report Agent

Creates the final output.

This is the basic idea behind multi-agent AI.

The key skill is not simply knowing how to create an agent.

It is knowing where an agent actually adds value.


Human-in-the-Loop Design

A good AI workflow should not assume that AI must make every decision.

Some tasks should include human approval.

For example:

Low Risk

AI summarizes a document.

→ Automatic.

Medium Risk

AI prepares an email.

→ Human reviews.

High Risk

AI changes a production database.

→ Explicit approval required.

This creates a practical principle:

AI can automate the work without necessarily automating the final decision.

This distinction is extremely important in enterprise environments.


AI Workflow Design Requires Better Questions

Instead of asking only:

"What can AI do?"

workflow designers need to ask:

What is the objective?

What outcome are we trying to achieve?

What information does AI need?

Where does that information come from?

Which tools are required?

Does AI need a database, API, application or search system?

What decisions can AI make?

Which decisions require human approval?

What happens when AI is wrong?

Is there a validation or recovery mechanism?

How do we measure success?

Can we determine whether the workflow actually improved the process?

These questions are often more important than the wording of the initial prompt.


The Importance of AI Evaluation

A workflow can look impressive in a demonstration and still fail in production.

For example, an AI agent may correctly handle 9 out of 10 tasks.

But if the tenth task involves a critical business operation, that failure may be unacceptable.

This is why AI evaluation is becoming increasingly important.

Developers need to test:

  • Accuracy

  • Reliability

  • Tool usage

  • Failure recovery

  • Security

  • Latency

  • Cost

  • Consistency

Agentic systems are particularly challenging to evaluate because a task can involve multiple model calls and tool interactions rather than a single response. (anthropic.com)


The New Skill: Workflow Thinking

This may become one of the most valuable AI skills.

Instead of thinking:

"I know how to write prompts."

Think:

"I know how to turn a business problem into an AI-powered workflow."

For example:

Business Problem

Customer support takes too long.

Break the process down

Ticket → Classification → Knowledge Search → Response → Approval → Resolution

Identify AI opportunities

Classification + Retrieval + Drafting

Identify automation opportunities

Ticket routing + notifications

Identify human decisions

Complex complaints + refunds + sensitive cases

Build and evaluate the workflow

This is much closer to real-world AI implementation.


Who Needs AI Workflow Design Skills?

This skill isn't limited to AI researchers.

It can be useful for:

Software Developers

Building AI-powered applications.

Data Engineers

Connecting AI systems with business data.

IT Professionals

Automating operational workflows.

Business Analysts

Identifying AI automation opportunities.

Database Professionals

Building AI workflows around enterprise data.

Product Managers

Designing AI-enabled products.

Entrepreneurs

Automating repetitive business processes.

Students

Building practical AI projects rather than only chatbot demonstrations.


A Practical Learning Roadmap

Someone starting today can learn AI workflow design progressively.

Step 1 — Learn Generative AI

Understand:

  • LLMs

  • Tokens

  • Context

  • Prompting

  • Structured output

Step 2 — Learn APIs

Understand how applications communicate with AI models.

Step 3 — Learn RAG

Understand:

  • Embeddings

  • Vector search

  • Document retrieval

  • Knowledge bases

Step 4 — Learn Tool Calling

Learn how AI can interact with:

  • APIs

  • Databases

  • Files

  • Applications

Step 5 — Learn AI Agents

Understand:

  • Planning

  • Memory

  • Tools

  • State

  • Task execution

Step 6 — Learn Workflow Orchestration

Understand how multiple steps are connected.

Step 7 — Learn Evaluation and Security

Learn how to test and control the workflow.

This progression moves from:

Prompt → Application → Workflow → Agent


The Future of Prompt Engineering

Does this mean prompt engineering is disappearing?

Not at all.

Prompts will remain important.

An AI agent still needs instructions.

A RAG system still needs guidance.

A workflow still needs model behavior defined.

But prompt engineering may increasingly become one skill inside a much larger discipline.

The evolution could look like this:

Prompt Engineering

Context Engineering

AI Application Development

AI Workflow Design

Agentic AI Engineering

The important skill is becoming the ability to combine all of these components.


From Prompt Writer to AI Workflow Designer

The next generation of AI professionals may spend less time asking:

"What is the perfect prompt?"

and more time asking:

"What should happen before the AI runs?"

"What should happen after it responds?"

"What tools should it be allowed to use?"

"How do we validate the result?"

"When should a human intervene?"

"What happens when something goes wrong?"

These are workflow-design questions.

And they are becoming increasingly important as AI moves from simple conversation toward task execution.


Final Thoughts

Prompt engineering introduced people to the power of Generative AI.

But the next phase is bigger.

AI systems are becoming connected to databases, APIs, applications, documents and business processes.

AI agents can use tools.

Multiple agents can collaborate.

RAG can provide organizational knowledge.

Workflow engines can coordinate complex tasks.

Humans can remain part of the decision-making process.

The result is a new way of thinking about AI.

Don't just ask AI to generate an answer.

Design a system that helps AI complete useful work.

That may be the real skill beyond prompt engineering.

The future AI professional may not simply be someone who knows how to write a great prompt.

It may be someone who knows how to turn a real-world problem into a reliable, secure and measurable AI workflow.

What Happens When AI Agents Start Talking to Other AI Agents?

  What Happens When AI Agents Start Talking to Other AI Agents? Imagine asking one AI agent: "Prepare today's sales and inventory r...