Objective · tools · feedback
AI agents for business:
when autonomy is justified
An AI agent does not simply answer a question: it chooses the next step, reaches for permitted tools, checks the result and continues until the predefined stop condition. This flexibility is useful to no process by default — and demands measurable control.
Anton Konnov · 20 August 2026 · 10 minutes
Architecture choice
Chatbot, workflow or agent
Chatbot
Explains, searches and drafts text, but does not change the external environment.
Workflow
Executes a pre-defined sequence of steps and checks.
Agent
Chooses its own steps and tools depending on intermediate results.
Multi-agent system
Separates roles only where independent contexts deliver measurable benefit.
Criterion
When an agent is genuinely needed
An agent approach makes sense when a task has a clear objective but an unknown number of steps in advance; it requires collecting data from several sources, selecting tools and adjusting the plan based on actual results.
If the route is stable, rules are formalised and exceptions are rare, conventional automation or workflow is usually simpler, faster and more predictable. Agent complexity must be justified by result quality, not by technology novelty.
Example
Why does an order look delayed: workflow or agent
A hypothetical example, not a real client story or deployment. The task is an internal summary of why an order appears to be delayed. The summary does not promise to fix the order and does not make commercial decisions. The permitted sources are the CRM order record, the warehouse event log and the shipment status. Before comparing them, confirm that the records refer to the same order and note when each was last updated.
A workflow is enough. If the answer comes from the same fields in known systems and predictable checks — the order never reached the warehouse; it was packed but not shipped; the carrier has not updated the status for a long time — a fixed route with rules will do. Several sources and branches do not by themselves call for an agent: workflows can branch too.
An agent is a candidate. When the next useful step depends on what has been found and cannot reasonably be prescribed in advance as a short route. For example, the warehouse log contains a reference to a returned item, and what to check next — a replacement, a split shipment or a manual correction — only becomes clear from that return's history in a permitted system or from a specific clarification through the agreed tools. If such variants are few and recurring, that is again a workflow branch. Even then, the agent is compared against the simplest workable option, without assuming it will win.
The output. A compact internal summary: checked observations, each tied to its source and update time; separately, hypotheses, open questions, an owner and the next check. A status mismatch is not a proven cause of delay; incomplete, stale or conflicting data cannot establish the cause with confidence. A finished summary does not mean the problem is resolved.
Boundaries of this scenario. Read and draft only; contacting the customer, changing the order and issuing a refund are decided by an authorised person. These are the limits of this example, not a general rule for every agent. If data is missing or conflicting, or the agreed time, cost or iteration limit is exceeded, the work stops and hands over what has been gathered, what remains unchecked and who takes the next step.
Acceptance. Test routine cases, conflicting data and an unavailable source. Compare against the current or simple route: the share of usable summaries correctly grounded in sources, corrections and false completions, total cycle time, human review effort, and compute and tool cost. A more accurate summary does not by itself mean less manual work.
For the technical side, see How AI agents work.
Six steps
How to run a controlled pilot
Check for a simpler solution first
Compare the agent against search, a single model call, RAG and a fixed workflow. Document what process variability cannot reasonably be described by rules.
Define the outcome and stop condition
Set the input, expected artefact or action, quality criteria, time and cost limits, maximum iteration count and situations for human hand-off.
Connect context and tools
Provide only the necessary sources and operations. For each tool, describe its purpose, parameters, typical errors and a verifiable environment response. Separate memory from the task's temporary context.
Separate permissions
Read, draft, modify data, publish and pay — these are different access levels. Irreversible, external and sensitive actions stay behind human confirmation or a separately and strictly limited control loop.
Test trajectories, not just answers
Verify which sources and tools are selected, how the agent responds to system unavailability, conflicting data and incorrect intermediate results. Normal, edge and failure scenarios are needed.
Launch with observation
Preserve the action sequence, tool results, stops, human intervention, cost and final acceptance. Expand autonomy only after sustained test passes.
Control
Four boundaries of autonomy
Data
Which sources the agent may read and what is permitted to store in memory.
Actions
Which tools are available and which operations require confirmation.
Resources
Time, iteration, model, compute and external call limits.
Escalation
When the agent must stop, explain its state and escalate to a human.
Acceptance
What to measure in the pilot
Proportion of tasks completed against a verifiable quality criterion.
Number of corrections, manual interventions and false completions.
Compliance with access permissions, confirmations and stop conditions.
Cycle time and cost per successful task relative to the current process.
Repeatability of results after changing inputs and tool availability.
Mistakes
What most often complicates the system
Avoid starting with multiple roles, shared memory and dozens of tools. Each new component multiplies the number of trajectories and makes diagnosis harder. Start with one useful scenario, an observable trace and clear acceptance criteria.
It is also dangerous to judge an agent by the persuasiveness of its final text. A correct answer reached through wrong data or a prohibited action is not a successful result.
Sources
Technical references
First step
Pick one variable task
We will compare workflow and agent, define tools, access boundaries and pilot criteria.