An AI agent is a software system that takes a goal, decides the steps to reach it, and uses real tools — a CRM, a calendar, a database, an API — to carry them out. Unlike a chatbot, it acts rather than answers, and runs inside limits you define, with a logged trail and a path back to a human.
Most companies that tell us AI did not work for them tried the same thing: a chat window bolted onto a website, trained on a help centre, answering questions nobody had. It deflected some tickets. It changed nothing about how the business ran.
The gap is not model quality. It is that a chatbot ends where the work begins. Someone still has to open the CRM, check whether the lead is real, book the slot, update the record and tell the rep. An agent is the thing that does that part.
What an AI agent actually is
Strip away the marketing and an agent is four parts wired together. Take any one away and you are back to a chatbot.
- A goal it is responsible for — qualify this lead, resolve this ticket, keep this pipeline accurate.
- Context it can rely on — your documents, your product data, the customer's history, retrieved at the moment it is needed rather than crammed into a prompt.
- Tools it can call — the CRM write, the calendar booking, the database query, the internal API, the refund endpoint.
- A loop with a stopping rule — it acts, checks what came back, decides whether it is done, and knows when to stop and fetch a human.
That fourth part is where most in-house attempts come apart. A model asked to keep going until the job is done will happily keep going forever, or declare victory on a half-finished task. Production agents earn their reliability from the boundaries around the loop, not from the prompt inside it.
Agent vs chatbot vs workflow automation
These three get used interchangeably in sales decks and they are not the same tool. Picking the wrong one is the most expensive mistake in an AI project, because you only find out after the build.
| Chatbot | Workflow automation | AI agent | |
|---|---|---|---|
| Handles | Questions with known answers | Fixed, predictable sequences | Goals with variable paths |
| Decides | Nothing — it retrieves | Nothing — you wrote the branches | Which step comes next, at runtime |
| Touches your systems | Rarely | Yes, on rails | Yes, through scoped tools |
| Breaks when | The question is new | The input is shaped unusually | The goal is ambiguous or the tools fail silently |
| Right for | FAQ deflection | Invoice routing, form-to-CRM | Qualification, triage, live calls, research |
A rule worth keeping: if you can draw the whole process as a flowchart without a single diamond that says it depends, you want workflow automation, and it will be cheaper, faster and more reliable than an agent. Agents earn their cost where the path genuinely varies with the input.
Where agents pay for themselves
Across the systems we have shipped, the work that repays an agent has a recognisable shape: it happens constantly, a person makes a small judgement call each time, and the judgement follows patterns that already exist in your data.
- Inbound lead qualification — reading a form fill or an email, checking fit against real criteria, enriching it, and either booking a call or closing it out. Covered in depth in AI sales agents.
- Live phone conversations — answering, understanding, looking things up mid-call and acting on what it finds. See how a real-time voice agent works.
- Tier-one support — resolving the request rather than deflecting it, which means refunds, address changes, order lookups and returns, not just article links. More in AI customer support agents.
- CRM hygiene and handoffs — the unglamorous work of keeping records accurate, described in CRM and workflow automation.
- Internal research and reporting — pulling from several systems, reconciling them, and writing the summary a person would otherwise assemble by hand every Monday.
Notice what is not on that list: anything where being wrong is expensive and hard to detect. Pricing approvals, legal drafting, medical or financial advice, irreversible financial movements. Agents can prepare that work. They should not sign it.
What it takes to keep one reliable
A demo agent works on the happy path. A production agent works on the eleventh unusual case at 2am, and the difference is entirely engineering.
- Scoped tools. Each tool gets the narrowest permission that lets it do its job. An agent that can read every customer record because that was easier is a data incident with a delay on it.
- Grounded answers. Anything factual comes from retrieval over your own content, with the source attached — see RAG and knowledge systems.
- Confidence thresholds. Below a bar, the agent stops and escalates rather than guessing confidently, which is the failure mode users forgive least.
- Human approval on irreversible actions. Sending money, deleting data, and committing to a customer are approval gates, not autonomy.
- Traces on everything. Every step, tool call, input and output logged, so a bad outcome is reproducible.
- Evaluation sets. A fixed collection of real cases the agent is re-run against before any prompt, model or tool change ships.
That last one is the discipline most teams skip and most regret. Without an evaluation set, every change to an agent is a guess, and the only detector you have is a customer complaint.
How a first agent should be scoped
The instinct is to start with the biggest problem. The better move is to start with the problem whose success is easiest to measure, because the first project's real job is to teach the organisation what these systems can and cannot do.
A good first workflow
- Happens dozens of times a week, so results arrive in days, not quarters.
- Has an obvious metric — response time, qualification rate, tickets resolved without a human.
- Touches one or two systems, not seven.
- Has a safe failure mode: the worst case is a person doing what they already do.
A bad first workflow
- Spans four departments and needs everyone's agreement to launch.
- Depends on data that lives in a spreadsheet somebody maintains by hand.
- Has no agreed definition of a correct outcome.
- Fails expensively — money moves, or a customer commitment is made.
One well-chosen workflow in production beats a strategy deck covering nine. It also gives you the thing no vendor can hand over: an internal, evidence-backed sense of where the boundary sits.
The honest limitations
Agents are not a replacement for a broken process. If the underlying workflow is undefined — nobody agrees what a qualified lead is, or when a refund is allowed — an agent will not resolve that ambiguity. It will encode whichever version it was given and apply it at scale, which is worse.
They are also not free after launch. Models change, your product changes, your customers ask new things. Budget for the agent as a system that is maintained, the same way you budget for any other production service.
And they do not remove the need for people. What they remove is the part of the work where a person is acting as a router between two screens. In practice, teams running good agents do not get smaller — they stop spending their day on the queue and start spending it on the exceptions the agent escalates.
Where to start
Pick the workflow your team complains about most, and write down what a correct outcome looks like in one sentence. If you can write that sentence, an agent can probably be held to it. If you cannot, that is the actual first project — and it is worth doing regardless of whether AI ever touches it.
How to take an AI agent from idea to production
The sequence we use on a first engagement, from picking the workflow through to a supervised rollout.
Find the workflow
Map where repetitive judgement work happens and how often. Choose one workflow with high volume, a clear definition of success, and a safe failure mode.
Define the correct outcome
Write down, in one sentence, what a correct result looks like — and collect twenty real past cases with their right answers. This becomes the evaluation set.
Design the agent
Decide what the agent is responsible for, what tools it may call, what knowledge it needs retrieved, and exactly where it must stop and ask a person.
Build the integrations
Wire the tools into the real systems under scoped credentials, with the write paths behind approval gates until the agent has earned them.
Run supervised
Put the agent in front of live traffic with a person reviewing every action. Fix what the traces reveal, and re-run the evaluation set after each change.
Release the gates and monitor
Remove approval on the actions the agent has proven, keep it on the ones that are irreversible, and monitor resolution rate, escalation rate and cost per run.