01
It holds permissions
Credentials that reach real systems, which means the blast radius of a mistake is now a question someone has to be able to answer.
Technologies
Agentic AI and workflow automation
An agent that only talks is a demo. An agent that acts is a system — with permissions, failure modes and an audit requirement, and almost none of it decided by prompt quality.
The Filter
Four questions
Ships With
An evaluation harness
Irreversible Steps
Behind approval
Provider Stance
No reseller agreements
The Transition
The moment an agent stops answering and starts acting, it becomes production software.
Calling tools, writing to your ERP, sending messages, moving a ticket — the moment a model does any of that it acquires permissions, failure modes and an audit requirement. That transition is where most agent projects either become valuable or become a liability, and the difference has almost nothing to do with prompt quality.
02
It has failure modes
Multi-step processes fail halfway. Loops retry. A wrong answer no longer stops at the screen — it lands in a system of record.
03
It carries an audit requirement
Somebody will eventually ask which tools were called, with what arguments, on whose authority, and what came back.
Finding the Work
Finding the work worth automating
The most reliable way to waste six months is a use case workshop that produces a prioritisation matrix. The opportunities that pay are found by sitting with the people doing the work and watching where the time actually goes, which is usually somewhere nobody would have put on a slide. Our practical filter is four questions.
01
Does this task happen often enough that automating it matters?
02
Is the input available in a machine-readable form, or does somebody have to find it first?
03
Is a wrong answer recoverable, and how quickly would anyone notice?
04
Does a specific person own the outcome, or is it spread across three departments in a way that means nobody will adopt it?
Where Projects Actually Fail
Tasks that fail the last question fail after deployment, which is the most expensive place to fail.
Engineering
What separates a working agent from a fragile one
None of this is exotic. It is the ordinary discipline of production software, applied to a component that happens to be non-deterministic.
01
Scoped permissions
An agent should hold the narrowest credentials that let it do its job, and the blast radius of a mistake should be something you can state in one sentence.
02
Approval gates on anything irreversible
Reading is cheap to get wrong. Sending, paying, deleting and committing are not, and those steps belong behind a human confirmation until the error rate is measured and accepted.
03
An evaluation set that exists before it ships
Agent quality drifts with every model update, prompt change and tool change. Without a fixed set of cases and an automated way to run them, you cannot tell an improvement from a regression — and you will hear about the regression from a user.
04
Idempotency and rollback
Multi-step processes fail halfway. The system has to know what it already did, and be able to undo it.
05
Logging a reviewer can follow
Which tools were called, with what arguments, on whose authority, and what came back. This is what makes an incident investigable, and it is also most of what an auditor asks for.
06
Cost and latency budgets
Agent loops that retry are the standard way a pilot’s economics quietly stop working at production volume.
Governance
Governance, without theatre.
For European organisations the obligations follow from where the system sits in the risk classification, and a good deal of what the EU AI Act asks for is documentation, human oversight and traceability that a properly built agent produces as a by-product.
Our role is to make sure the engineering produces that evidence during the build, rather than having it reconstructed afterwards by people who were not there. We work alongside your legal and compliance function rather than replacing their judgement.
Produced During the Build
Documentation Human oversight Traceability
What the Engineering Leaves Behind
A record of which tools were called, and with what
The authority each action was taken under
The approval points and who cleared them
A fixed evaluation set and its results over time
The rollback path for every irreversible step
Build, Buy, or Both
Buy the orchestration layer. Build the domain logic.
The frameworks, model providers and agent platforms are moving quickly, and they are not where a manufacturer or an energy company builds an advantage.
Usually buy
The orchestration layer, the model providers and the agent platform. This is a fast-moving commodity layer, and running your own is rarely where the advantage sits.
Usually build
Your data, your process knowledge and the tool integrations nobody else can replicate. This is the part that is actually yours.
What moves the line
Data residency requirements, how far your workflow deviates from what the platforms assume, and what you can realistically staff.
How we compare providers
Against your actual requirements rather than a feature matrix. We hold no reseller agreements with any of them.
Where This Applies
Where the bottleneck is finding and reconciling information, not writing it
Technical document and drawing digitisation
Structured data pulled out of PDFs, scans and CAD exports — and then checked, which is the half that usually gets skipped.
Procurement and negotiation preparation
Assembling a fact base from systems that do not talk to each other, ahead of a conversation where that base decides the outcome.
Engineering and quality processes
Workflows carrying a heavy documentation load, where the evidence trail is as much of the work as the decision.
Internal knowledge work
Where the bottleneck is finding and reconciling information across systems rather than producing the final document.
A pilot needs a defined kill criterion, so that stopping is a normal outcome rather than an admission of failure.
From the real workflow to a scoped pilot
How a first engagement looks
We spend time in the real workflow with the people who run it, then come back with what survives the filter.
1
Sit in the real workflow
Time with the people who run the process, watching where it actually goes rather than collecting nominations for it.
2
Two or three candidates
The tasks that survive the four questions — frequency, machine-readable input, recoverability, and a named owner.
3
Build-or-buy for the platform layer
A recommendation for the orchestration layer against your data residency, deviation and staffing constraints.
4
Architecture, then a scoped pilot
An architecture that includes the evaluation harness and the approval model from the start, and where it makes sense a scoped pilot with the kill criterion agreed before it begins.
Related
The decisions and capabilities either side of an agent programme
Technologies
Generative AI
The model layer underneath an agent — where it creates measurable value, and how to deploy it with governance and control.
Technologies
Synthetic Data & Simulation
Where an agent’s evaluation set comes from when you cannot collect or label enough real cases.
Services
AI Management Consulting
Vendor-neutral advisory on the decisions around an agent programme, from sourcing to delivery leadership.
Services
AI Make or Buy Decision Support
Where the line between platform and domain logic falls in your case, decided against requirements rather than vendor slides.
Technologies
Edge & On-Premise AI
What changes when the agent has to run inside your perimeter — device class, power budget, and an update path you control.
Blog
Stop Chasing AI Quick Wins
What Uber’s Agentic Pods get right: the most valuable workflows are discovered inside the work, not nominated in a workshop.
Book a Strategy & Architecture Review
Bring us the workflow you are not sure you should automate.
We will tell you whether it survives the filter, what it would take to run it safely, and where stopping would be the right call.
Book a Strategy & Architecture Review nAIxt Technologies GmbH
Am Forst 2
82166 Gräfelfing, Germany
+49 89 54196515
info@naixt-technologies.de