When people talk about AI at work, the familiar picture is an employee opening a chat window and asking a model to draft an email, summarise a document or analyse some data. The task gets faster, but the person still gives instructions step by step and carries the result into the rest of the workflow.
AI agents move that boundary.
Instead of waiting for one prompt at a time, an agent can receive a goal, decide which steps are needed, use tools, inspect the result and continue through several actions within defined limits. That shifts the economic question from “How much faster can AI make one task?” to what happens when AI can manage a sequence of tasks across an entire workflow?
The OECD published Agentic AI in organisations in September 2026, drawing on interviews with organisations across regions and sectors. Its findings show that agentic systems are already being developed and deployed, while governance, measurement and operating practices remain at an early stage.
Our previous article Will AI Really Take Jobs – or Make Some Work More Valuable? examined why AI often changes the task before the whole job. Once several tasks are connected through an agent, the object being redesigned can become the workflow itself.
An AI agent is more than a smarter chatbot
A familiar chatbot follows a simple pattern:
user asks → AI answers → user decides what happens next
An agent can operate through a longer loop:
goal → plan → use tools → act → inspect results → adjust → continue or escalate
Anthropic describes agents as systems that can work across applications, write and execute code, manage files and complete tasks spanning multiple steps. This gives them more autonomy, meaning greater freedom to act within the tools, permissions and constraints provided by people.
Consider sales.
A chatbot might draft a follow-up email. An agent could receive the goal “follow up with high-potential leads,” check a CRM, review account status, draft outreach, choose a sending time, record the result, schedule a meeting and hand the case to a salesperson when human judgment is needed.
Most of those tasks already existed. What changes is the ability to connect them without requiring a person to mediate every handoff.
The economics change when AI moves from tasks to workflows
Early workplace AI has often focused on individual productivity: writing faster, searching faster or summarising meetings faster.
But businesses do not operate as isolated tasks. Work moves through people, teams, software systems and approval chains.
A procurement process might receive a request, check inventory, request quotes, compare suppliers, seek approval, create a purchase order, track delivery and pass information to accounting. Delay may come less from any single step than from waiting for information, rechecking data and coordinating between functions.
That is where an agent can have a different economic effect from a copilot used by one employee.
McKinsey’s 2026 Global Survey on AI found that 80% of respondents said AI had improved individual productivity, while only 37% attributed at least some EBIT impact to their organisation’s AI use. The gap does not prove that agents will solve the problem, but it shows why making one person faster does not automatically make the whole organisation more profitable.
If the workflow still contains duplicate approvals, file transfers and disconnected systems, the benefit can stop at the employee’s desk.
The money may be in the handoffs between steps
The business value of an agent is therefore not limited to labour hours saved.
It may come from reducing waiting time, copy-and-paste work, follow-up, data-transfer errors and the amount of human coordination required between systems.
An invoice process, for example, might receive a document, validate fields, match a purchase order, flag anomalies, prepare an accounting entry and route the case for approval. If an agent can handle the structured parts as one sequence, people may only need to intervene when the data does not match, a threshold is exceeded or something unusual occurs.
Agents are not zero-cost digital workers.
Organisations still pay for models, software integration, data preparation, permissions, monitoring, evaluation, security and human review of exceptions. Some workflows may generate large savings. Others may cost more to automate than they return.
That is why agent economics should be measured at the workflow level, not only by the number of employee hours apparently saved.
Human work may shift from doing every step to supervising the system
As agents take on more execution, the human position in a workflow can move.
Instead of:
do → hand off → check → continue → request approval
the pattern may become:
set goal → define constraints → review exceptions → decide → take responsibility
This connects directly to the labour-market issue in A10. If AI becomes better at some tasks, human value may shift toward judgment, context, communication and accountability. Agents intensify that question because they can take over sequences rather than only isolated tasks.
Microsoft’s 2026 Work Trend Index for Thailand reported that 32% of Thai workers in its sample qualified as “Frontier Professionals,” compared with a 16% global average, while Thai organisations still faced challenges moving from individual AI use into core organisational workflows.
This is Microsoft platform and survey evidence, not a nationally representative labour statistic. But it highlights an important distinction: being good at using AI individually is not the same as redesigning an organisation around it.
More autonomy does not mean removing people from the system
The attraction of an agent is that it can continue working with less step-by-step supervision. The same property also expands risk.
If a chatbot gives one bad answer, a person may notice before acting on it. If an agent misunderstands a goal and has access to several systems, the error can travel through multiple steps before anyone sees it.
Anthropic identifies risks including misreading user intent and prompt injection attacks that try to manipulate agents into actions they should not take.
This makes human-in-the-loop design more nuanced than requiring a person to approve every click.
An organisation can define:
- which systems the agent may access
- transaction or spending limits
- actions that require approval
- data that may not leave a boundary
- exceptions that must be escalated
- points where the process must stop for a human decision
The goal is not necessarily to remove people from the workflow. It is to choose where human judgment is most valuable.
Large companies and SMEs may not benefit at the same speed
Agent performance depends on more than the AI model. It also depends on data quality, system integration, permissions and how clearly the existing workflow is defined.
That gives a company with CRM, ERP, APIs, data governance and an IT team a different starting point from a small business still operating through spreadsheets, chat applications and documents spread across several systems.
The OECD’s 2026 D4SME Survey found that SME use of AI is rising and some firms are experimenting with AI agents, but strategic and secure integration into business operations remains uneven. Time constraints, maintenance costs and skills gaps are among the barriers.
This creates a competition question that is more interesting than who gets a chatbot first.
If agents genuinely reduce coordination costs, firms with better data and systems may lower the cost per transaction or process more work before smaller competitors can. If that advantage compounds, the productivity gap between firms could widen.
The opposite path is also possible. Cheap, easy-to-deploy packaged agents could reduce the technology disadvantage faced by SMEs. It is still too early to assume that large companies will be the only beneficiaries.
What if this goes much further toward AGI?
Today’s AI agents are not Artificial General Intelligence, or AGI, a concept used for AI with broad capabilities across many domains.
An OECD and European Commission AI literacy framework describes AGI as systems able to match or exceed the cognitive versatility and proficiency of a well-educated adult, and states that no such systems currently exist.
Movies often portray something much broader: a digital mind that can learn, plan, decide and act across almost any domain on its own. That image is far beyond the agents businesses are experimenting with today.
Current agents still operate with goals, tools, permissions and boundaries supplied by people, and they can still fail at seemingly ordinary tasks.
If systems eventually move from managing specialised workflows toward broad general capabilities of the kind associated with AGI, the implications for business and labour would be a different-order question. Before then, the more practical issue is where humans should sit in a workflow as agents take on more of its execution.
What to watch next
The first signal is how quickly agents move from pilots into production workflows. A successful demo is not the same as operating every day with real data, customers and money.
The second is whether companies begin measuring ROI at the workflow level. If reporting still focuses only on adoption rates or employee hours saved, the economic effect may remain unclear.
The third is how much permission organisations are willing to give agents – read-only access, drafting, transactions, or actions without approval. That reveals both confidence and accepted risk.
The fourth is exception rates and human escalation points. A system that looks autonomous on screen may still depend on substantial invisible human work.
For Thailand, another signal is whether SMEs can access reliable agent systems at an affordable cost. If agentic workflows require expensive data infrastructure and integration, the productivity gap between firms could widen. If packaged systems sharply lower those costs, smaller businesses could benefit sooner than expected.
The first wave of workplace AI made us ask, “Can AI help with this task?” AI agents push the question further: “Do we still need to design the workflow this way at all?” That may be where the business impact becomes deeper than simply having a faster writing assistant.
