An AI agent is software that uses a model to interpret a goal, select approved tools, take bounded actions, check the result, and hand work to a person when confidence or authority runs out. A chatbot answers. Fixed automation follows a known path. MASS builds agents as described on the Agentic AI service page.
Agents versus chatbots and automation
| Chatbot | Fixed automation | Agent | |
|---|---|---|---|
| Job | Return a message | Run a known if-then path | Advance a goal across steps |
| Tools | Usually none, or search | Deterministic connectors | Approved tools, chosen per case |
| When to use | Q&A on a corpus | The path never branches | The path branches and a person still needs a prepared next step |
If every case follows the same steps, a rule or a form is cheaper and easier to audit. MASS will say so. See also when an agent is the wrong tool.
What is an AI agent in business?
A useful business agent usually has three parts: a reasoning model, tools that read or change business systems, and instructions that define the job and its limits.
A useful business agent usually has three parts: a reasoning model, tools that read or change business systems, and instructions that define the job and its limits. OpenAI’s current agent guidance describes the same foundation: model, tools, and instructions. If the model is not controlling workflow execution, the system may be AI-powered, but it is not an agent.
One complete workflow: inbound support triage
This is a design walkthrough, not a published MASS production result.
- Intake. A ticket arrives with the customer message, account id, and channel.
- Read. The agent may call
get_ticketandget_account_summarywith read-only keys. It cannot see other customers. - Retrieve. It searches the approved help corpus for this product and plan. If retrieval confidence is low, it stops and escalates.
- Draft. It writes a reply and a one-line rationale citing the passages used. Nothing is sent yet.
- Decide. If the intent is a password reset how-to, a specialist can accept the draft. If the intent is a refund, cancellation, or legal complaint, the agent only prepares a brief for a person.
- Act (optional, later). Write tools—
post_reply,update_ticket—stay off until evaluation on a sample of drafts is acceptable. - Log. Tool calls, sources, and the human accept/edit/reject decision are stored for review.
Permissions
Read, draft, and write are separate. Give the agent the minimum tool set for that workflow. Adjacent accounts, payroll, and unrelated files stay out of reach. High-impact actions require a named person. This matches how MASS describes pilots on the Agentic AI page: specialists accept, edit, or reject before anything reaches a customer.
What AI agents handle well
Agents are strongest where work is repetitive but not perfectly predictable. Traditional automation needs a fixed path; an agent can interpret an email, extract intent from a document, choose from several approved actions, and adapt when information is missing.
1. Triage and routing
An agent can classify enquiries, identify urgency, gather missing details, and route work to the correct team. The result is faster response without pretending every request can be resolved automatically.
2. Research and synthesis
Agents can search approved documents, compare records, summarize long material, and produce a structured brief. They are especially useful when information is scattered across PDFs, tickets, CRM records, and internal knowledge.
3. Tool-based operations
With narrow permissions, an agent can update CRM fields, create a draft report, prepare an invoice packet, schedule a follow-up, or open a ticket. This is where agents differ from answer-only chatbots.
4. Long-running coordination
An agent can monitor a workflow over time: wait for a document, check whether an approval arrived, remind the owner, then continue from the saved state.
Ask: “Can we describe the goal, approved data, allowed tools, completion criteria, and escalation rule?” If not, the workflow is not ready for an agent.
What an AI agent should not do alone
Agents are not independent employees, universal experts, or a substitute for accountability. They can misread context, use a tool incorrectly, or produce a plausible answer that is not supported by evidence.
- High-stakes decisions: hiring, medical, legal, credit, safety, and financial decisions need qualified human responsibility.
- Irreversible actions: large refunds, payments, deletions, contract acceptance, and public communications should require confirmation.
- Unclear policy: automation magnifies ambiguity. If employees disagree about the process, the agent will not repair the underlying operating model.
- Unbounded access: an agent should not receive broad system permissions “just in case.” Give it the minimum required access.
NIST’s AI Risk Management Framework emphasizes defined human roles, continuous risk management, and lifecycle monitoring. In plain terms: someone must own the agent, know what it is allowed to do, and review how it performs after launch.
How to choose your first AI agent workflow
- Start with volume. Pick work that happens often enough for time savings to matter.
- Prefer reversible actions. Drafting and recommending are safer starting points than sending and approving.
- Measure a baseline. Record current handling time, error rate, backlog, and escalation rate.
- Define success. Specify what “done” means and what evidence the agent must preserve.
- Launch narrowly. Begin with one team, one workflow, and a visible review queue.
Evaluation and human oversight
Before write permission, score a labeled set of cases on outcome, evidence, tool correctness, and whether the agent escalated when it should have. After launch, sample traces weekly. A person owns the agent: they know the allowed tools, they can turn write off, and they review misses. NIST’s AI Risk Management Framework is a useful structure for that ownership—not a certificate.
A safe agent architecture
| Layer | Purpose | Example control |
|---|---|---|
| Instructions | Define the job and boundaries | Never issue a refund above $100 |
| Knowledge | Ground decisions in approved sources | Use the current policy library only |
| Tools | Read or act in business systems | CRM update with scoped permissions |
| Guardrails | Detect unsafe or out-of-scope behavior | PII filtering and action risk tiers |
| Human handoff | Transfer uncertain or sensitive work | Escalate after two failed attempts |
| Evaluation | Measure quality over time | Weekly review of sampled traces |
Frequently asked questions
What is the difference between an AI agent and a chatbot?
A chatbot primarily returns messages. An AI agent manages steps toward a goal and can use approved tools to retrieve data or take action. Some chat interfaces contain agents, but the interface alone does not make the system agentic.
Can an AI agent replace employees?
It is more accurate to think in terms of tasks and workflows. Agents can remove repetitive coordination and preparation work, while people retain judgment, relationships, accountability, and exception handling.
How long does an AI agent take to build?
A narrow proof of value may take weeks; a dependable production workflow takes longer because integrations, permissions, evaluation, exception handling, and monitoring matter more than the initial demo.
Build this kind of system with MASS: Agentic AI services. Related: support-agent design and automation ROI.