Resource · Executive guide
An agent is software that uses an AI model to work toward a goal. It chooses actions, uses tools, checks what happened, and continues until it finishes or asks for help.
Have you already tried any of these?
A real brief
“Prepare and send the invoice for order 1043”
Invoicing agent
Chooses the next step based on the result
CRM and orders
Looks up authorized data
Invoicing and email
Runs specific actions
The agent cycle
Click any step in the diagram to see what happens at that point in the example.
The user defines the expected result. A good brief includes context, limits, and the condition that shows when the work is done.
Order invoicing example
Prepare and send the invoice for order 1043.
Goal identified
Systems in the example
The model interprets the current state and the orchestrator decides which capability it can use. It adapts the path to each case as it goes.
Order invoicing example
First it needs to locate the order and validate the customer.
Next action: query CRM
Systems in the example
The software runs an authorized operation: query a CRM, call an API, create a document, or request an approval.
Order invoicing example
It looks up order 1043 and retrieves line items, amount, and tax data.
Tool executed
Systems in the example
The result comes back to the agent. It may detect incomplete data, an error, or a need for human validation before continuing.
Order invoicing example
The order exists and is paid, but the customer's tax ID is missing.
Issue detected
Systems in the example
With the new information it chooses another step, asks for help, or finishes. Time, cost, and permission limits keep the cycle from running unchecked.
Order invoicing example
It requests the missing data. After validation, it creates the invoice and sends it.
Result completed and logged
Systems in the example
Repeat the cycle
Returns to step 02 with the new data.
The agent delivers the result and leaves a record of what it did: actions taken, tools used, and data handled. That trail makes the process auditable and repeatable.
Order invoicing example
Invoice sent to the customer and order updated in Holded.
Result logged and auditable
This is the condensed version of the cycle. Open this page on a computer to see the full interactive diagram.
Autonomy always has limits. Permissions, budget, and traceability define how far it can go on its own.
Chatbot, assistant, or agent
These words describe how a system is used. The same application can answer like a chatbot in one task and act like an agent in another.
Its main job is to interpret messages and generate a response. It can be very useful for drafting, summarizing, or reasoning with you.
Example
You paste in an email and it proposes a reply.
Examples you may already know
It combines conversation with controlled access to documents or internal data. It often uses RAG (search over your own documents) to find sources before answering.
This is what many people simply call an «assistant» or «a chat with documents».
Example
It answers an operational question and cites the applicable internal procedure.
Examples you may already know
It chooses actions based on what it finds. It can query systems, transform data, create records, and request validations.
Example
It reviews an order, generates the invoice, asks for approval if there is an exception, and logs the result.
Examples you may already know
Conversation
Its main job is to interpret messages and generate a response. It can be very useful for drafting, summarizing, or reasoning with you.
Example
You paste in an email and it proposes a reply.
What it adds
Examples you may already know
Lookup
It combines conversation with controlled access to documents or internal data. It often uses RAG (search over your own documents) to find sources before answering.
This is what many people simply call an «assistant» or «a chat with documents».
Example
It answers an operational question and cites the applicable internal procedure.
What it adds
Examples you may already know
Execution
It chooses actions based on what it finds. It can query systems, transform data, create records, and request validations.
Example
It reviews an order, generates the invoice, asks for approval if there is an exception, and logs the result.
What it adds
Examples you may already know
Quick rule
If it waits for your next message to continue, it is a chatbot.
If it only searches and cites your own information, it is a knowledge assistant.
If it runs steps and chooses the next one based on what it finds, it is an AI agent.
Personal assistants
Personal assistants are general-purpose agents: they live in your channels, have their own environment to act, and remember what they did. Very useful for one person, but they are not the same as an agent designed around a business process.
Personal assistant
Who it works for
For one person and any task: today it books a trip, tomorrow it summarizes your email.
When it acts
Always on: it works in the background and pings you when it needs something.
Where it works
In its own environment (a cloud computer or your machine) with your personal accounts and apps.
Memory and data
It remembers preferences, projects, and conversations, stored in personal accounts.
Channels and permissions
It reaches you on WhatsApp, Telegram, or email, with your personal account permissions.
Company agent
Who it works for
For a specific business process, with a metric to improve.
When it acts
When the process calls for it: an order, a ticket, or an invoice comes in.
Where it works
Connected to the ERP, the CRM, and company systems through controlled integrations.
Memory and data
Data handled under company policy and GDPR, with a defined retention period.
Channels and permissions
It lives in the team's tools, with role-based permissions, a log of every action, and approvals.
OpenClaw Open source · Your machine or server
It installs on your own machine or on a server and reaches you on the channels you already use: WhatsApp, Telegram, Discord, or email. State, memory, and credentials stay on your machines.
Hermes Nous Research · Your machine or server
An open-source agent that learns from its own work: it creates skills, improves them with use, and keeps them. It reaches you through multiple messaging platforms.
Grok Bot xAI · Cloud
Each bot has its own cloud computer. It signs into your tools, works in the background, and comes back when it needs approval.
Muse Meta · Cloud
Connected to Facebook, Instagram, and third-party apps. It drafts and sends emails, books travel, or makes purchases, from its app or from WhatsApp.
Dots OpenAI · Cloud
Always-on agents with their own cloud computer and browser. They start from ChatGPT memory and keep working even when your machine is off.
A personal assistant that works with client or company data from personal accounts leaves that data outside IT's control: no action log, no role-based permissions, and no retention policy. Your team may already be using them. It is worth deciding where to set the limits.
Anatomy of an agent
The end experience depends as much on the model as on the software around it, the connected tools, and the operational controls.
It interprets language, analyses the situation, and proposes the next action. Its capability matters a lot, but a model on its own only produces an output.
Design question
What level of quality, speed, and cost does this task need?
Names you may have seen
It defines the work, the applicable policies, and the information available at each step. It can include documents retrieved with RAG.
Design question
What information does it need, and which business rules must it follow?
Names you may have seen
Sometimes called a harness. It keeps state, presents tools to the model, executes its requests, handles errors, and decides when to stop.
Design question
How are steps, retries, costs, and exceptions controlled?
Names you may have seen
These are concrete operations the orchestrator can run: look up an order, update the CRM, create a file, or send an approval request.
Design question
Which systems does it access, and what actions can it run in each one?
Names you may have seen
Identity, permissions, human validation, logs, evaluations, and alerts turn a demo into a system that can operate safely.
Design question
What gets logged, who approves, and how is the system stopped?
Names you may have seen
Layer 01
It interprets language, analyses the situation, and proposes the next action. Its capability matters a lot, but a model on its own only produces an output.
Design question
What level of quality, speed, and cost does this task need?
Names you may have seen
Layer 02
It defines the work, the applicable policies, and the information available at each step. It can include documents retrieved with RAG.
Design question
What information does it need, and which business rules must it follow?
Names you may have seen
Layer 03
Sometimes called a harness. It keeps state, presents tools to the model, executes its requests, handles errors, and decides when to stop.
Design question
How are steps, retries, costs, and exceptions controlled?
Names you may have seen
Layer 04
These are concrete operations the orchestrator can run: look up an order, update the CRM, create a file, or send an approval request.
Design question
Which systems does it access, and what actions can it run in each one?
Names you may have seen
Layer 05
Identity, permissions, human validation, logs, evaluations, and alerts turn a demo into a system that can operate safely.
Design question
What gets logged, who approves, and how is the system stopped?
Names you may have seen
What “custom” means
Configuring instructions can be enough for a simple need. A custom agent makes sense when it has to integrate, apply its own controls, and respond to a specific process.
ChatGPT, Gemini, or Claude with their standard capabilities. Useful for individual tasks and exploration.
Examples
Choose this level if… A good instruction solves the task, with no internal data or actions in your systems.
A GPT, Gem, or equivalent configuration with instructions, files, and tools available on the platform.
Examples
Choose this level if… It needs your documents, tone, and format, but the platform tools are enough.
Orchestration, CRM or invoicing integrations, permissions, interface, and evaluation built around a business process. Even when a system has no modern integrations, a custom connector can be built.
Examples
Choose this level if… It must operate inside your systems (ERP, CRM, or tools without standard integrations) and meet your own business requirements: permissions, approvals, and logs defined by the company.
Moving up to a custom agent is within reach for most companies. If you want to systematize operations, integrate your tools, and stop repeating manual work, this is your level. Levels 01 and 02 are enough to start today.
Going deeper
Explained simply for people who want to know more.
An API is a set of rules that lets two programs talk to each other without a person in the middle: one asks for information or requests an action, and the other replies in a predictable way.
It is everyday infrastructure: when your website shows live stock from Holded, or your CRM records a sale made in the online store, there is an API in between. It has been around for decades, long before AI.
Not every tool has an API. In those cases you need a custom integration that can extract and write data into the tool automatically.
The model does not “load” tools on its own. The orchestrator shows it a description of the allowed operations. When the model requests one, the software validates the request, runs it, and returns the result.
The connection can be coded directly against an API or exposed through MCP (Model Context Protocol). MCP offers a standard way to publish tools and resources so compatible clients can discover and use them. It is the same mechanism Claude, ChatGPT, or Gemini use when you connect external tools.
MCP simplifies integrations, but it does not turn an application into an agent. It also does not replace authentication, permissions, validations, or business limits.
Popular MCP servers
A skill is a package of instructions, procedures, and resources that expands what an agent knows how to do. Instead of explaining how to do a task in every conversation, the knowledge is documented and the agent loads it when the case requires it.
Difference from tools: a tool is an operation the software runs (look up an order, send an email). A skill is knowing how to do something: the steps, the criteria, and the formats. An agent combines both: it decides with judgment and executes with its tools, following the skill's procedure.
For a company, skills are how you capture the way your team works: the rules, exceptions, and formats that today live in a few people's heads.
It is an agent specialized in working with code. It can read a repository, edit files, run tests, observe errors, and correct its approach. Its loop is visible: it changes something, checks the result, and continues.
The same pattern can be applied to operations, finance, or customer support. The tools, permissions, and success criteria change.
Coding agents you may have seen
“Open” usually refers to models published with their weights, which you can install on your own infrastructure. “Closed” describes models offered as a service by their vendor. Neither option by itself determines where your data is processed.
Privacy and deployment are another axis: public cloud, the vendor's enterprise environment, private cloud, or your own infrastructure. The choice depends on sensitivity, performance, cost, and operational capacity. For personal data, GDPR is the starting point: process only the data you need and document where it is processed.
Closed (proprietary) models
Open (open-source) models
Memory keeps relevant information across steps or sessions. RAG retrieves fragments from authorized sources when they are needed. They are different mechanisms, and both need access and expiry policies.
Observability logs actions, results, costs, and errors. It lets you evaluate the system, investigate an incident, and improve the process with evidence.
RPA (robotic process automation) is software that operates other applications through the interface: it opens windows, fills forms, and copies and pastes data. It has been automating repetitive tasks for years in systems that do not offer an API.
An RPA flow follows fixed steps. If the screen moves or an unexpected case appears, it stops. It runs what was programmed and does not adapt its behavior to the situation.
With agents, RPA has a useful role again: the agent decides what to do in each case and can rely on an RPA bot to interact with legacy systems that have no API. The decision belongs to the agent; the screen manipulation belongs to RPA. It is a realistic way to automate processes that run on legacy software.
You do not have to wait for your vendor to ship AI features. An MCP server or an API integration can be built specifically for your systems: a legacy ERP, custom-built software, shared spreadsheets, or applications without a modern interface.
The relevant question is which data and actions the agent needs, and whether there is a safe way to access them. When the data lives in a specific system, it is almost always possible to build that access with the right permissions and controls.
Application in an SME
An agent is useful when the task mixes variable information, several systems, and decisions that do not fit a fixed sequence.
Handle an order issue, check the delivery note, coordinate the next action, and notify the customer without reviewing dashboards one by one.
Popular tools
Ecommerce · logistics · tickets
Review invoices, match them against delivery notes, speed up bank reconciliation, and chase missing documents before month-end close.
Popular tools
Accounting · ERP · email · documents
Resolve questions with internal sources, classify and route tickets, and open a case when needed.
Popular tools
RAG · helpdesk · identity
Research an account, prepare the context, and update opportunities. Reply to each lead in minutes, not hours: first-response speed often decides the sale.
Popular tools
CRM · web · calendar
Draft content in the brand voice and prepare a reading of each campaign by crossing data from every platform.
Popular tools
Ads · analytics · content
Answer employee questions by citing internal policy and prepare the document onboarding for each new hire.
Popular tools
Internal policies · ATS · documents
Compare supplier quotes, follow open orders, and spot invoices that do not match what was received.
Popular tools
Suppliers · orders · invoices
Review contracts clause by clause against your own templates and flag deviations before signature.
Popular tools
Contracts · templates · records
Cases by industry
The examples above are grouped by department, but each industry has its own repetitive tasks with variable information. These cases show concrete processes where an agent is a good fit; open the ones that feel familiar.
Online store · carriers · messaging
CRM · documentation · messaging
Case files · templates · deadlines
Insurers · calendar · messaging
Transport · invoicing · ERP
Budgets · subcontractors · reports
Suppliers · POS · invoicing
Issues · suppliers · reports
Ads · analytics · projects
ATS · CVs · reports
Support · product usage · CRM
Quality · production · ERP
Carriers · policies · commissions
Portfolios · regulation · CRM
Catalog · LMS · messaging
Volunteering · grants · reporting
CRM · valuation · workshop
Bookings · fees · loyalty
Calendar · contributors · payments
WhatsApp/email · history · calendar
Proposals · tracking · reporting
Leads · quotes · photos
These cases are also solved with custom development: the solution is designed around each company's specific process.
Diagnosis · custom design · own integration
How to decide
When a use case fits, two decisions remain: the type of system and the scope. These are the signals for answering each one.
Design decision
The practical difference is who decides what to do next. In conventional automation we define the steps and possible paths in advance, even if some of them use AI. An agent receives a goal and decides which tool to use and what to do next based on the results.
Signals that point to an agent
Signals that point to conventional automation
In practice, most companies need both. The choice depends on the specific task: automation for what is predictable, and an agent for what changes in each case. Combining them in the same process is the usual approach.
Scope decision
Start with the simplest level that solves the task well, and move up only when the process requires it.
A general-purpose application is enough
ChatGPT, Gemini, or Claude as they are, with good instructions.
A configured agent is enough
A GPT, Gem, or equivalent on the platform you already use.
A custom agent is the better fit
Built around your process, systems, and controls.
From demo to production
The controls you need depend on the agent's autonomy. Choose a level to see what each condition requires: looking up data, acting in your systems, or paying and communicating.
Agent autonomy
The agent only looks up information. Low risk: these are the controls that apply.
Least privilege
EssentialEven if it only looks things up, access is limited to the data it needs. Read access without personal credentials.
Logging and evaluation
RecommendedIt logs what it looked up and when. That is the basis for trusting the next levels of autonomy.
Approval where it matters
RecommendedWith read-only access there is almost nothing to approve. Still define who responds when something fails.
A pilot with one metric
RecommendedValidate answer quality with real cases before expanding usage.
Validation
An eval is a test with a case, an expected result, and an acceptance criterion. It runs automatically whenever something changes: the instructions, the tools, or the model itself. Without evals, you cannot confirm the agent's effectiveness or the impact of changes to it.
Evaluation suite
Agent that invoices in Holded
3 / 3 CASES PASSED
Order with a matching delivery note
Invoice created in Holded and sent to the customer
Order without a matching delivery note
Process paused and the responsible team notified
Amount above the approval limit
Escalated to human validation before invoicing
When a change breaks a case, the eval flags it immediately. The failure is fixed before it reaches production, not after a customer notices.
Define a suite of real cases and their expected result. That will be the agent's quality criterion. If it passes them, it is doing what you expect.
Measure the success rate against those cases (for example, 48 of 50 passed, 96%) and adjust instructions, tools, and limits until the result is stable.
Run the suite with every model or process change and sample real cases to catch drift before your customers do.
In short
An agent uses an AI model to choose actions, use tools, and check results until it finishes the task or asks for help.
It moves forward without waiting for your next message: it chooses the next step based on what it finds at each one.
An agent's reliability depends on the full system: instructions, connected tools, and operational controls.
It adds value when the task mixes several systems, unpredictable inputs, and decisions that do not fit a fixed sequence.
Where to start
A tightly scoped pilot, least privilege, human approval where the impact requires it, and a metric to compare. Sometimes the better answer is still a simpler automation.
Details you need in order to evaluate adopting agents.
Next step
If you have identified a task with variable information, several systems, and repetitive decisions, you can explore how we build agents for businesses.
No-commitment first conversation. We will also tell you if a simpler automation is a better fit.