Sales agents
Answer product questions from approved information, qualify buying intent, collect details and hand warm opportunities to your salespeople.
AI Agents
AI agents that actually do useful work.
Not another chatbot. We build AI agents for businesses in South Africa that answer real questions, qualify leads, research information and trigger workflows — working only from information you have approved, and handing over to a person the moment judgement is needed.
If the work follows fixed rules once mapped, an agent is the wrong tool — Automation & Integrations is, and we will say so.
One question, six steps, bounded the whole way
- Question arrivesCustomer · Team member
- Agent interpretsWhat is being asked
- Approved knowledgeYour information only
- Business rulesWhat it may do
- Acts in your systemsCRM · Email · WhatsApp
- A person takes overJudgement · Approval
What an agent actually is
Automation follows rules. An agent interprets.
Deterministic automation runs predefined steps: when this arrives, do that. It is the right tool for most repetitive work, and it is what our Automation & Integrations service builds.
An agent earns its keep where the input has to be interpreted before the next step is knowable — reading an enquiry, working out what is actually being asked, choosing the next action from a bounded set. It works from information you have approved, inside rules you have set, and it stops and hands over the moment the next step needs a person.
A concrete example
What a properly built sales agent does with one enquiry
A customer asks which of your products suits a specific operating requirement. Here is the whole job — and, just as deliberately, where it ends.
Understands the request
Reads the enquiry as a question about a real requirement, not a keyword to match.
Asks clarifying questions
Operating conditions, volumes, constraints — the questions your salespeople would ask first.
Compares approved product information
Works only from the specifications and materials you have signed off. Nothing is improvised.
Answers the product questions
Plainly, citing what it knows, and saying so when the answer is not in the approved material.
Collects contact details
Name, business, the right channel to continue on — captured in a structured, reusable form.
Qualifies buying intent
Timeline, budget signals and fit, against the questions you have agreed matter.
Creates or updates the CRM lead
The conversation lands in your system as usable data, not a transcript somebody has to mine.
Hands the opportunity to a salesperson
With the context attached. The relationship, the negotiation and the commitment belong to a person.
What it never does: invent a specification, quote a price you have not approved, or promise a delivery date it cannot know. Where the answer is not in the approved information, the honest move is built in — it says so, and hands over.
Where agents fit
Six kinds of agent we build
Each is the same idea — interpretation inside boundaries — pointed at a different repetitive job, and chosen by the shape of the repetition rather than enthusiasm for the technology.
Customer support agents
Handle the questions customers ask the same way every week, check order and request status, and escalate anything unusual to a person.
Internal assistants
Answer staff questions from company documents and policies, so people stop interrupting each other for information that is written down.
Research agents
Gather and summarise repetitive research — suppliers, prices, coverage, submissions — with sources attached, for a person to judge.
Document agents
Pull structured detail out of quotes, applications, invoices and forms, and put it where it belongs instead of leaving it in a PDF.
Workflow agents
Take one event — an order, a booking, a cancellation — and carry out the several bounded updates that should follow from it.
More than one of these? Normal. Projects start with the single job that repeats most and carries the least risk — and extend only once that one has earned it.
What a useful agent is made of
Knowledge, tools, rules — and a person in charge
An agent is not a model with a logo. It is a small system, and every part of it is decided with you before anything goes live.
Knowledge
Approved company information only — nothing scraped, nothing guessed.
Tools
CRM, email, WhatsApp and business systems, with the minimum access needed.
Rules
What it may read, may do, must never do — and when it must stop and ask.
Human control
Approval points, escalation and handoff, with a named person accountable.
The line we hold
We don’t automate decisions that shouldn’t be automated.
AI is used here to reduce unnecessary work — the reading, sorting, drafting and fetching that sits around a decision. The decision itself stays with the person responsible for it. Approvals, customer commitments, pricing, anything with legal, financial or safety weight: defined human responsibility, by design.
Where a customer is talking to an agent, they can tell. Where the agent is unsure, it says so and hands over. And everything it does is logged, so you can always see what happened and why.
How it runs
Find the job first. Build last.
The first two phases exist to protect you from an expensive agent nobody needed. Sometimes their honest conclusion is a simpler automation — and then that is what we build.
Find the job
Identify the repetitive work that genuinely needs interpretation, and what it costs today.
Define the boundaries
Agree what the agent may know, may do, must never do, and when it must hand over.
Connect the systems
Wire in the approved knowledge and the tools it needs, with the minimum access that works.
Build & test
Run it against real past cases — the normal ones and the awkward ones — before it meets a customer.
Deploy gradually
Start narrow, with a person reviewing. Widen its scope only as the evidence supports it.
Improve
Review what it handled and what it escalated, and tighten the boundaries and knowledge from real use.
Scope and price
AI agent projects are scoped individually
There is no fixed agent package, because the honest price depends on the workflow’s complexity, the condition of the data and knowledge involved, the integrations required, the approval points you need, and how the agent has to operate day to day.
Every project gets a written scope and price before work begins, with third-party and AI usage costs identified separately. Where the agent needs a portal, dashboard or interface around it, that is scoped with Custom Apps & Business Systems.
Where to start
If enquiries are part of the job, start with the review
Most agent conversations begin with enquiries: too many arriving, or too many waiting. The free Website Review covers that journey — how enquiries arrive, what happens to them, and where they stall — which is exactly the evidence an agent project starts from.
If the job has nothing to do with the website, skip it and describe the work instead. That is the faster route.
Free Website Review
Get a Free Website Review
Free, with no obligation, and normally back within two business days.
If the steps never change once they are mapped
Automation & IntegrationsIf the agent needs a portal, dashboard or system around it
Custom Apps & Business SystemsQuestions about putting an agent to work
Will it make things up?
This is the right first question. The agent works only from information you have approved, and where the answer is not in that information it says so and hands over — that behaviour is built and tested before anything goes live, on your real past cases. No AI output is inherently accurate, which is exactly why the boundaries, the approved knowledge and the handoff exist.
How is this different from a chatbot?
A typical chatbot matches keywords to canned replies and traps people in menus. An agent interprets the request, works with your actual business information and systems, takes bounded actions — updating the CRM, booking the follow-up — and hands to a person with context attached. The difference shows up in what happens after the conversation, not in the chat window.
Which systems can it connect to?
CRM, email, WhatsApp, databases and other business systems — where those platforms expose a supported connection, which depends on the provider and your account types. That is confirmed during scoping rather than assumed. Anyone promising a specific integration before checking your accounts is guessing.
What does an agent cost?
Scoped individually — complexity, data, integrations, approvals and operating requirements genuinely change the answer, so a fixed price on this page would be a made-up number. You get a written scope and price before work begins, and third-party and AI usage costs are identified separately. If a simpler automation covers the job, we quote that instead; it is usually cheaper.
Do we actually need one?
Possibly not. If the work follows fixed rules once mapped, deterministic automation is simpler, cheaper and easier to trust — and it is what we recommend. The agent is for the part that genuinely needs interpretation. Working out which is which is the first phase of every project.
Next step
Find the job an agent should be doing
Tell us the questions your team answers over and over, or the enquiries that wait too long for a person. We will tell you whether an agent is the right tool — and say so plainly when it is not.
Does the repetition start at your website’s enquiries? A free Website Review covers that journey.
Or email lance@wedevelop.co.za