The distinction that matters
An AI agent is software that uses a language model to work towards a goal and can use tools — look something up, check a system, create a record, send a message — rather than only producing text.
Set against the neighbouring terms:
- A chatbot converses. It answers questions and collects information. Nothing changes in your systems.
- An assistant converses with access to context — your documents, your data — but still mainly produces answers.
- An agent converses and acts. It can retrieve a real answer from a live system, create an enquiry record, check availability, or trigger a workflow.
The addition of action is what makes agents valuable and also what makes them risky. Once software can change something, the interesting engineering question stops being "can it answer well" and becomes "what is it allowed to do, and what happens when it is wrong".
What agents are genuinely good at
Narrow, repetitive jobs where the information exists and the action is well defined.
- Qualifying enquiries. Asking the questions your sales process needs — requirement, quantity, location, timeline, rough budget — and creating a structured record before a person joins.
- Answering from your own material. Retrieving the relevant passage from your price lists, policies or specifications and answering from it.
- Routing. Classifying a request and putting it in front of the right person with the details already gathered.
- Status lookups. Checking an order, booking or ticket in a connected system and reporting back.
- Internal knowledge. Answering staff questions from documentation that exists but nobody can find quickly.
- Draft preparation. Assembling a quotation or a summary for a person to check and approve.
Notice that several of these end with a human. That is not a limitation of current technology so much as a sensible design choice about where commitment happens.
What agents are bad at
Equally important, and less often stated by people selling them.
- Open-ended goals. An agent told to "grow the business" will produce activity, not results. Scope is everything.
- Long chains of dependent steps. Reliability compounds downward. A step that is 95 per cent reliable is fine; ten of them in sequence is not.
- Judgement calls with real consequences. Pricing negotiation, credit decisions, medical or legal matters, anything irreversible.
- Situations not covered by their material. A well-built agent says so. A badly built one invents something plausible.
- Emotional situations. An upset customer wants acknowledgement from a person. Automation here does more damage than the time it saves.
How answers stay accurate
The concern people raise first is invention — the agent confidently stating something untrue about your business. It is a real risk with a general-purpose model, and the mitigation is architectural rather than a matter of instructing the model to be careful.
Grounding. The agent retrieves relevant passages from your approved material and answers from them, rather than generating from what the model absorbed during training.
Honest failure. When retrieval finds nothing relevant, the correct behaviour is to say so and offer a person. This has to be configured and tested deliberately; it is not a default.
Scope limits. Requests outside the agent's defined job get redirected rather than attempted.
Testing with awkward cases. Before launch, we test with questions we know are not covered, to confirm it declines rather than improvises.
Log review. Most of an agent's eventual quality comes from reading the first few weeks of real conversations and fixing weak answers at source, usually by improving the underlying material.
Guardrails worth insisting on
If you commission an agent, these should be in the specification rather than discussed after something goes wrong.
- An explicit job description. What it handles, and what it must never attempt.
- Escalation topics agreed in writing. Complaints, refunds, negotiation, medical, legal, financial, and anything involving sensitive personal data.
- Scoped action permissions. The agent can only do what you have granted, and high-impact actions require human approval.
- Complete logging of conversations and actions, reviewable by your team.
- Disclosure that customers are dealing with an automated assistant, with a human route always visible.
- An off switch you can operate without contacting your provider.
A provider reluctant to commit to any of these is telling you something useful.
A worked example
Consider a company supplying industrial components, receiving perhaps thirty enquiries a week across WhatsApp, email and its website.
Without an agent: enquiries land with whoever sees them. Some are viable, some are for products the company does not stock, some are from buyers wanting quantities below the minimum order. A salesperson spends time on all of them equally, and the useful ones sometimes wait a day because the person was on a call.
With a well-scoped agent: every enquiry gets an immediate acknowledgement. The agent asks what item, what quantity, where it is going and by when. It checks the requested item against the product list and tells enquirers plainly when something is outside range — helpfully, and without occupying a person. Viable enquiries are recorded with all details and routed to the right salesperson with a notification.
What the agent does not do: quote a final price, negotiate, promise a delivery date, or handle a complaint about a previous order. Those go to a person.
The gain is not that the agent replaces the salesperson. It is that the salesperson starts every conversation informed, and no viable enquiry waits a day because someone was busy.
Do you need one?
Work through this honestly.
- Is your enquiry volume high enough? Below a certain level, a person handles it fine and an agent is an expensive solution to a small problem.
- Are the questions repetitive? If every enquiry is genuinely different, an agent has little to work with.
- Is your information documented? If not, that is the first project. An agent cannot answer from material that does not exist.
- Would plain automation be enough? Often yes. If the workflow follows fixed rules, you do not need a language model. See what AI automation involves.
- Would a chatbot be enough? Frequently, and it is cheaper and lower risk. Start there and add actions once conversation quality is proven. See what an AI chatbot is.
Our general advice is to start with the simplest thing that could work. A narrow agent doing three things reliably is worth considerably more to a business than a broad one attempting anything.
Thinking about an AI agent?
PresenticAI builds narrowly scoped agents with grounded answers, defined permissions, logged conversations and human escalation — and will tell you when simpler automation would serve you better.