The short answer: An AI agent is software that is given a goal and works through the steps to reach it: it reads, looks things up, uses other systems, checks what it has produced and hands back a result. Unlike a chatbot, it does a piece of work rather than just answering a message. Done well, it works within clear limits, and a person stays in charge of the decisions that matter.
Read the transcript
An AI agent is software that is given a goal, and works through the steps to reach it.
It reads, looks things up, uses other systems, checks its work, and hands back a result.
People still own the decisions.
A chatbot answers the message in front of it. Automation follows fixed rules.
An agent sits between the two: it reads messy material, and carries out the steps.
Ours build a training library from real calls, coach the coaches, and research a company before the call.
PlanningAssist, in beta, writes its answer only from the passages it finds, with a numbered source for every point.
If nothing matches well enough, it says so rather than guessing.
We decide in advance where a person looks at the work.
Start small, with a process your team does often, where most of the effort is reading, sorting, drafting or checking.
Read the article at freestart.ai/articles.
There are plenty of explainers about agents. This one is written by people who build them: we build agents for our own sales teams and planning work at Freestart, and we use them ourselves. So I'll explain the idea plainly, then show you what ours actually do, including where we keep people in the loop.
What an agent is
An agent is software that:
- Takes a goal. Not a single question, but a job: "build a reference library from this material", or "find out about this company before we ring it".
- Works through the steps. It breaks the job down and does the parts in order, rather than producing one reply and stopping.
- Uses tools and data. It can search documents, look up a public register, read a website or pull records from another system.
- Checks its work. A well-built agent tests its own output against the sources, and says so when it can't find what it needs instead of guessing.
- Hands back a result, for a person to use, review or approve.
The last point matters most. An agent does the reading, sorting and drafting. People still own the decisions.
How an agent differs from a chatbot and from ordinary automation
A chatbot answers the message in front of it. It's useful, but each reply is one step, and it's up to you to decide what to do next.
Ordinary automation, such as rules, scripts and workflows, follows fixed instructions: "if this field says X, send it to Y". It's dependable and cheap, but it can only handle the cases someone wrote a rule for. Give it a messy document and it has nothing to say.
An agent sits between the two. It can read unstructured material like a chatbot, and it can carry out a sequence of steps like automation, but it decides how to work through each case within the limits you set.
| Chatbot | Rules and automation | Agent | |
|---|---|---|---|
| What you give it | A question | A trigger and fixed rules | A goal |
| What it does | Replies once | Follows the same steps every time | Plans and carries out several steps |
| Handles messy, unstructured material | Yes, in conversation | Poorly | Yes |
| Uses other systems and data | Sometimes | Yes, as programmed | Yes, as the task needs |
| Checks its own output | Rarely | Only what the rules test | It should, by design |
| Best for | Quick answers | Predictable, high-volume steps | Reading, sorting, drafting and checking |
None of these replaces the others; good systems often use all three.
What ours do
Everything below is in use at Freestart unless I say otherwise.
Building a training library from real calls (TrainingHub)
TrainingHub is the training system for our sales teams. Part of it is a reference library: the set of resources we use to train staff. An agent builds that library from our training material and CallHub call transcripts. Reps study it as short chapters, with the original handbook one click away.
It's typical agent work: long, unstructured source material, read and organised into something people can study.
Coaching the coaches (CallHub)
CallHub records and transcribes our sales calls. One of its features, manager coaching, looks at a different kind of conversation: the coaching sessions between managers and their reps. It reviews the transcripts of those sessions and gives each manager feedback on how to coach better.
The agent does the close reading of long transcripts. What the manager does with the feedback is still up to them.
Researching a company before the call (ProspectHub)
ProspectHub builds a sales list from a sector and an area, then researches each company when asked: what it does, who runs it, what its customers say and whether it fits the campaign. The research covers services, size, target market, certifications and reviews, with a link to each source. Owners and directors are matched to Companies House and checked against the trading address.
New prospects are classified when they're imported, and companies that don't fit the campaign are flagged. A person decides whether to reassign or exclude them. ProspectHub doesn't decide who to ring; our staff still check and add to the research themselves. It's the starting point, not the last word.
Finding the right policy and showing the source (PlanningAssist)
PlanningAssist answers planning questions, and it's in beta. You ask in plain English and add your postcode.
In plain words, this is how it finds the answer. We've broken local plans, national policy and official guidance into short passages, each tagged with where it came from, and filed them by meaning as well as by their exact words. Your postcode tells PlanningAssist which planning authority, and which nation's policy, applies. It then searches for the passages that match your question, by meaning and by keyword, and writes the answer only from what it finds. Every point cites a numbered source you can open and read yourself.
It also checks itself. A draft whose citations don't point to a passage it actually found is written again, and if nothing matches well enough, it says so rather than guessing. That approach, looking things up before answering, is usually called retrieval.
Where agents fit, and where they don't
Agents fit well where the work is:
- Repetitive reading: transcripts, documents, websites, policy.
- Sorting: putting things into categories, flagging what doesn't fit.
- Drafting: a first version of a summary, a note or an answer.
- Checking: comparing a draft against its sources.
They fit badly where the work is a judgement call with real consequences: whether to hire someone, whether to take on a customer, what advice to give, whether something is right to publish. An agent can prepare the ground for those decisions. It shouldn't make them.
Keeping people in charge
Three principles guide how we design agents:
- Review points. We decide in advance where a person looks at the work. In ProspectHub, people decide what happens to a company that doesn't fit. In PlanningContent, the system our editors use to write and refresh Planning Handbook guidance, an editor sees each claim's verdict and citations, and nothing is exported until they approve it.
- Sources shown. Where an agent gives an answer, it shows what the answer rests on, so a person can check it rather than take it on trust.
- Nothing sent without a person where it matters. If a mistake would reach a customer, a client or the public, we design the agent so that a person signs the work off first.
How to pick a first process to hand to an agent
Start small. A good first candidate usually ticks most of these:
- Your team does it often, and it follows a familiar pattern.
- Most of the effort is reading, sorting, drafting or checking.
- The source material already exists: documents, transcripts, records.
- A person can easily tell a good result from a bad one.
- A mistake is caught at a review point before it does harm.
- Someone on your team owns the process and will review the output.
If a process needs a judgement call at every step, or there's no material for the agent to work from, it's probably not the right place to start.