04 / 06
Discipline
Agentic Process Automation
Software that carries out a multi-step task inside your business, using your systems and your data, and hands off to a person at the points where a person should decide.
Not a chatbot on your website, and not a subscription to somebody else’s platform. An agent that does a real piece of work you currently pay a person to do by hand, wired into the systems that work already lives in, with rules about what it may do on its own and what it must escalate.
This is worth doing for some processes and is a waste of money for most, which is why the first piece of work is finding out which.
Approach
The method, in full
How we approach it.
The method, published in full, because it is the part that differentiates us and it is not a secret.
1. Audit where automation pays
We work through your processes and find the ones with volume, with rules that can be written down, and with a real cost attached to doing them by hand. Most candidates fail this test. Automating a task that runs four times a month is a hobby.
2. One real process, not a platform
We build for a specific process end to end rather than delivering a general capability and hoping a use is found for it. General automation platforms get bought with enthusiasm and abandoned quietly, because nobody owns the work of pointing them at anything.
3. Integrate with what you have
The agent works inside your existing architecture, against your real data, with the same engineering standards as everything else we ship. Where it needs an integration that does not exist, that is engineering work and we will say so.
4. Tune against your criteria
Model selection and, where it earns its keep, tuning, measured against your data and your definition of a correct outcome. A public benchmark score tells you nothing about whether it gets your invoices right.
5. Guardrails and a human in the loop
What it may do unsupervised, what needs approval, what it must never do, and a full record of what it did and why. We took this from building compliance systems where every action needs an evidence trail, which turns out to be the right instinct for systems that act on their own.
6. Operate it
Like everything else we ship. An agent nobody monitors is worse than no agent, because it fails silently and confidently.
What AI actually changes, and what it does not
AI is genuinely useful. It drafts, it explains, it writes the boring code, it gets a first version onto the page. We use it and we should. The specific danger is that it makes output easier without making output valuable. Somebody who does not understand your problem can now generate a convincing proposal for the wrong solution in minutes.
A developer who does not understand architecture can generate code that works on Friday and hurts for a year. So it sits in the hands of people who stay responsible for the result. You cannot blame a tool for your judgement.
The position most vendors will not take
Most processes should not be automated with an agent. Some are better handled by ordinary code, which is cheaper, testable, and produces the same answer every time. Some should be simplified out of existence. Some are cheap enough by hand that the automation would never pay back its own maintenance.
Reaching for an agent first is how organisations end up with an expensive system producing plausible answers nobody can check. We would rather establish that in a paid assessment than discover it together during a build.
Proof
Where this is already running.
An AI-assisted feature we built runs in production inside a mining operations application, with the infrastructure, tests, and hosting around it that any other part of that system has. It ships as part of software people use to do their jobs.
That is one production deployment rather than a practice with a decade behind it, and you may as well know the size of it. What it means is that we have already done the unglamorous work of putting a model-driven feature into a system that has to keep running, which is where most demonstrations stop.
Engagements
Typical shapes this takes.
An automation assessment
Fixed fee, fixed duration. We map the candidate processes, score them, and write up which are worth automating, which should be ordinary software, and which should be left alone.
One process, built
Only after the above, and only where the numbers support it. Delivered as Scoped Delivery.
Operating it
Monitoring, guardrail review, and adjustment as the process changes.
Deliverables
What you get from the assessment.
A map of the processes examined
What each one does now, and what it costs.
A score for each
Against volume, rule clarity, current cost, and risk.
A recommendation per process
Agent, ordinary software, simplify, or leave alone.
Costed options
For the ones that pass.
The reasoning
In language you can take to a board.
The report is yours
Whether or not we do the work.
Related
Next step
Start with the assessment.
We do not sell agent builds off a website. If you find someone who does, ask what they are proposing to automate before they have looked at your business.
Claer & Volker
453 Winifred Yell St, Garsfontein
Pretoria, South Africa
info@claervolker.com
© 2026 Claer & Volker
·
·