Custom AI Agents
An agent that does one job properly.
Most AI spending buys a tool nobody uses. We build the other thing: a single agent with a named job, plugged into the systems you already run, that does the work end to end and hands off to a named person at a point we agree before anything is built.
What You Get
The parts that decide whether it works.
One Named Job
An agent is scoped to a single task with a defined start and finish. Anything broader is a demonstration, and a demonstration is not a thing that runs on Tuesday.
Connected To Your Systems
It reads and writes where the work already lives — your mailbox, your accounting package, your spreadsheets — rather than asking anybody to work somewhere new.
A Person Signs Off
Every agent has a checkpoint where a human approves, edits or rejects the output. We tell you where that checkpoint is and why it is there.
Escalation Written Down
What the agent must never decide alone is agreed in advance and written into the build: refunds, discounts, anything legal, and anything it is not confident about.
Where Your Data Goes
Before anything is built you are told which model processes what, where it runs, and what is kept. Information that cannot leave South Africa does not have to.
Measured Against The Old Way
We record how long the job took before and how long it takes after. An agent nobody can show a number for is one nobody can defend.
The jobs an agent can actually be given.
An agent built to do one named job end to end — draft the quote, reconcile the delivery notes, answer the first line of support.
An agent works where the job is repetitive, the inputs arrive in a form a machine can read, and somebody can tell when the output is wrong. That last one matters most: a task nobody checks is a task nobody notices going wrong. The jobs below have all three. The audit is what tells you which of them you have.
Jobs worth giving an agent
- Reading the inbox and drafting the quote
- Reconciling delivery notes against purchase orders
- First-line support, with escalation to a person
- Pulling figures into the monthly report
- Chasing overdue invoices on a schedule
- Turning meeting notes into tasks and dates
- Checking applications against a known set of rules
What lands in your hands
Not "AI for your business". One job, named, with a start and a finish: read the enquiries in the sales inbox and draft a quote against the price list; match delivery notes to purchase orders and flag what does not agree; answer the questions support gets every week and pass the rest to a person. A person still owns the outcome, and we tell you where that person sits.
What's included
- A working agent doing one named job
- Written scope naming what it will not do
- A named human checkpoint in every run
- Escalation rules for what it cannot answer
- Connections to the systems you already run
- Before and after timings for the job
- A written record of what it did
- Handover session and a plain-English run guide
How An Engagement Runs
Audit first, then build what earns its keep.
Four stages, in order. Nothing is built before the audit says it is worth building, and nothing goes live before you have watched it run beside the manual process.
01. Audit
We sit with the people doing the work and record where the hours actually go, task by task. You get the findings in writing whether or not we build anything.
02. Pilot
One task, built and run beside the manual process rather than instead of it, so you can compare the output before anything is trusted with real work.
03. Deploy
The agent goes into your existing systems, with the handover, the access and the written record of what it does and where a person has to approve it.
04. Operate
We watch what it gets right and what it gets wrong, fix the second, and tell you when a task has changed enough that the agent should be retired.
Questions
Before you commit.
Does this replace our staff?
It replaces tasks, and we will tell you which ones before anything is built. An agent that drafts quotes takes the typing, not the relationship: the person still checks the figure, still handles the customer who wants something unusual, and still owns whether it goes out. Most of the time the part that needed a person becomes the job. If you are buying this to cut a role, say so on the first call — it changes what we would build.
Where does our data go, and does it leave South Africa?
It depends on the model and where it runs, and you are told which before anything is built. A hosted model generally means your inputs are processed abroad, commonly in the European Union or the United States, which section 72 of POPIA permits where the provider is contractually bound to standards substantially similar to ours. Where information cannot leave the country, a model running on our South African servers or on your own hardware is the alternative, and it is usually slower and more expensive.
What if it gets something wrong?
It will, and the build assumes it. Every agent has a checkpoint where a person approves the output before it reaches a customer or a ledger, and the things it is not allowed to decide alone are written down before we start: refunds, discounts, anything legal, anything it is not confident about. It also keeps a record of what it did and why, so a wrong answer can be traced rather than argued about. An agent with no human checkpoint anywhere in it is not something we will build.
Can you do this without an audit first?
Sometimes. If you already know the job, can name the person who owns the outcome, and can show us real examples of the work — including the awkward ones, because the exceptions decide whether an agent can do it — there is nothing to audit and we scope it and build it. How much we need to see is settled on the call rather than set as a threshold here. If you are describing a feeling that things take too long, the audit is cheaper than building the wrong agent.
Which AI model do you use?
Whichever fits the job, and it changes. Providers release new versions often and the sensible choice for a task today is often not last year's, so the agent is built with the model behind it swappable rather than welded in. What we will not do is name a provider on a marketing page and quietly use a different one: the model, where it runs and what it costs to run are in the scope document you sign.
Ready to bring your idea to life?
Let's create something amazing together.
Firm quote against a scoped roadmap before any work begins.