Guide · AI agents
What is an AI agent?
How an agent differs from a chatbot and from automation, where it earns its keep, and how a person stays in charge.
- 1ReadTakes in what arrives: an invoice, an email, a chat message, a photo.
- 2DecideWorks out what it is and what should happen next.
- 3DraftPrepares the action: a quote, a reply, a check, a journal.
- 4ApproveA person approves anything that matters — the exact lines.
- 5ActCarries it out once, and logs who approved what.
In short
An AI agent is software that reads what comes in — documents, emails, chat messages — works out the next step and prepares the action: a quote, a reply, a check or an approval request. Unlike a chatbot it does work, not just conversation; unlike fixed automation it copes with messy, varied inputs. Good agents work inside your company's rules, see only what the person using them may see, and ask a person before anything important happens.
01Not a chatbot
AI agent, chatbot or automation: what's the difference?
All three use software to save time. They differ in what they take in, what they give back, and what they do when something looks wrong.
| Chatbot | Rule-based automation | AI agent | |
|---|---|---|---|
| What it does | Answers questions in a chat | Runs the same fixed steps every time | Reads, decides the next step and prepares the work |
| What it takes in | Questions typed into a chat | Data in an exact, expected format | Documents, photos, emails and chats as they come |
| What it knows | General knowledge | Only what it was programmed with | Your documents, price lists and policies — within each person's access |
| When something's odd | May answer confidently anyway | Stops, or carries on wrongly | Flags it and asks a person |
| What you get | Text | A finished step | A draft or an action, waiting for approval when it matters |
02Where it helps
Where AI agents earn their keep
Agents pay off wherever people read, compare and re-type all day. Some examples, by department:
Compare supplier quotes, check invoices against POs and contracts, flag price and GST differences, check expense claims, turn timesheet photos into pay.
Build quotes from the price list, check discounts against the rules, and route anything outside them to a manager.
Answer warranty and servicing questions from your own policies, and draft replies for a person to send.
Answer policy questions with the handbook section named, and sort CVs into shortlist, maybe and not a fit — with the evidence.
File what arrives and find any document by what's in it — showing each person only what they may open.
One snapshot of what needs attention today, across departments.
03In practice
What it looks like in practice
Here's an agent comparing supplier quotes. It reads five quotes in three formats, lines up the same items, finds the best buy — and notices that one supplier's sums don't add up.
- Mei Ling · Purchasing
Which of these supplier quotes is the best buy?
5 quotes · PDF, Excel, Word - Collie
Best buy for the full list: Harbour Office Supply. One more thing — Lion City Trading's quote adds up wrong on line 7.
Quote comparison
- Quotes read
- 5 (PDF, Excel, Word)
- Best buy
- Harbour Office Supply
- Total
- S$18,072.60
- Line 7, Lion City Trading
- S$180.00 too high
- Mei Ling · Purchasing
Thanks — I'll ask Lion City to correct it.
Logged to the audit trail
Nothing here needs a prompt engineer. The person asks the way they'd ask a colleague, and checks the result.
04Guardrails
How a person stays in charge
The real question isn't whether an agent can act — it's what it may do alone. Good agents are fenced in four ways.
Anything that moves money waits for a person, who approves the exact lines — once. Change anything and it's a new request. Nobody approves their own.
An agent sees only what the person using it may open. HR's records stay with HR.
Every request, approval and action is logged with who and when, in a log that shows if anyone changes it.
ID, card and bank numbers are masked in what the agent shows and logs.
05Start small
How to start with AI agents
Start with one chore, not a transformation programme.
- 1Pick one choreFrequent, rule-bound and dull: invoice checks, quote comparisons, claim checks.
- 2Write down the rulesWho may approve what, which limits apply, which documents are the source of truth.
- 3Start with draftsLet the agent prepare and a person approve everything at first.
- 4MeasureCount hours saved and errors caught on real work, for a month.
- 5Widen with careLet it do more alone only where the rules are clear — and keep approvals wherever money moves.
06Choosing
Questions to ask before you choose an agent
Ask these of any AI agent, ours included.
0 of 8 checked
Sources
- PDPC — Model AI Governance Framework
- PDPC — Advisory Guidelines on the use of personal data in AI recommendation and decision systems (PDF)
Checked 12 October 2026. Rules and prices change — confirm what applies to your business with the official source.
FAQ
Common questions.
What is an AI agent in simple terms?
Software that reads what comes in — documents, emails, chat messages — works out the next step and prepares the action, such as a quote, a reply or a check, inside your company's rules. It asks a person before anything important happens.
What is the difference between an AI agent and a chatbot?
A chatbot answers questions in a conversation. An AI agent does work: it reads documents and messages, decides the next step and prepares the action — a quote, a check, a reply — and asks a person to approve what matters.
Are AI agents safe for business data?
They can be, if they see only what each user may see, ask before acting on anything that matters, log every action, and keep data where you decide. Ask where the AI runs and whether your data is used for training.
Will AI agents replace staff?
Agents take over the reading, comparing and re-typing; people keep the decisions, the approvals and the relationships. Most teams spend the hours saved on work they never had time for.
What is Collie?
Collie is HeySheep's private AI platform. It gives each department — finance, sales, customer service and HR — its own AI agent, running on a box inside your office, with no cloud AI by default and no usage bills. Agents draft and check; a person approves what matters.
Keep reading
