Agentic AI:
From experiments to work that runs itself

Most companies have stopped asking whether AI is useful. The question now is why so little of it ever reaches daily operations. A team runs a pilot, builds a demo that impresses everyone in the room and then nothing happens. The demo sits there. Six months later someone asks what came of it and the honest answer is that it never made the jump from “look what this can do” to “we use this every morning.”

That jump is what agentic AI is really about. A chatbot answers your question and leaves you to do the actual work. An agent does the work. It reads the request, checks your systems, prepares the result and only pulls in a person when something genuinely needs one. Answering versus acting. That is the whole difference and it is bigger than it sounds.

We spend most of our time at Acumen helping companies make that jump in a way that survives contact with a real business, not just a slide deck. Here is what agentic AI actually does, where it pays off and how we go about building it.

AUTHOR – Niels

What is agentic AI?

Plenty of companies already use AI to answer questions or draft text. Agentic AI is a different animal. It plans work that takes several steps, uses your own tools and systems to get through them and carries the task to the finish. A person stays involved, but only where a person is actually needed.

An example makes it concrete. Ask a chatbot to draft a reply to a supplier and it will, once you ask. An agent notices the supplier email the moment it lands, looks up stock in your ERP, prices the order against your own rules, drafts the quote and sends it for approval. The technology underneath is much the same. What it does to your working day is not.

So what can an agent handle on its own? It can read a document and pull out what matters, whatever format it arrives in. It can update records straight in your systems rather than making someone copy and paste. It can decide within rules you set, connect several systems in a single flow and pick up a new order or ticket the second it appears and act on it. And throughout all of that, it keeps a person in the loop and in control. That last point is not a caveat. It is the design. The agent takes the repetitive nine-tenths of a job so people can spend their time on the tenth that needs a human.

Why this matters now

The problem in most organisations is not a shortage of AI experiments. It is a shortage of AI that actually runs, on its own, day after day. Around nine in ten companies use AI somewhere in the business (McKinsey). Fewer than one in ten see AI deliver real, measured. Analysts expect that to shift fast: Gartner projects that 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from less than 5% a year earlier. Whoever closes the gap between testing and using will feel it first.

What can it be used for?

Agents work best in the corners of a business where the same thing happens on repeat, the rules are recognisable and the hours add up. A handful of patterns show up in almost every company we talk to.

Document intake is usually the obvious place to begin. Invoices, contracts, safety data sheets, supplier forms: processing them by hand is slow and mistakes creep in. An agent reads them, sorts them, routes them and leaves only the odd exception for a person to look at.

Quoting is another. A slow or inconsistent quote loses business. An agent gathers the data, applies your pricing rules, handles the approval routing and turns a two-day job into a two-minute one.

Then there is the support inbox, where requests stack up before anyone with the right expertise even sees them. An agent reads each one, answers the simple ones and sends the rest to the right team straight away, at three in the afternoon or three in the morning.

And there are the decisions that stall for no good reason other than the data living in five different places. An agent stitches those systems together and puts the relevant picture in front of whoever has to make the call.

Added value for your business

None of this stays abstract for long. The payoff tends to show up in a few ways you notice almost immediately.

Things move faster, because a job that used to eat half an afternoon is done while the team gets on with something else. The output is steadier, because an agent does not get tired or cut corners when a deadline looms. It does not keep office hours either, so work that arrives overnight is handled by morning. And the people freed from the repetitive part get their attention back for the decisions that genuinely need them.

The reassuring part is that you do not need a company-wide programme to get there. You start with one process, prove it works and grow from there. A run of small, solid wins beats a single enormous rollout that everyone is quietly afraid of.

Collaboration and way of working

Getting an agent to hold up in real operations is as much about how you build it as what you build. We treat each project as a partnership. The solution takes shape gradually, with regular feedback, rather than us vanishing for a few months and coming back with a finished box. The people who know the process inside out help check our assumptions, kick the tyres on early results and draw the line on what the agent should and should not do without a human.

That closeness is not a formality. Agentic AI reaches into real decisions, so trust and openness matter more than usual. Talking often and staying aligned is what keeps a solution honest about the messy constraints of the actual job instead of a tidier version of it.

The way we work runs in three steps. We start by finding the right use case, which means sitting down to brainstorm where an agent genuinely fits and then checking whether the data behind it holds up. Most AI projects come unstuck here, on a fuzzy problem or shaky data, long before the model is ever the issue. Next we prove it. We build a scoped proof of concept that runs one clearly bounded slice of the process end to end. Not something that only behaves on stage, but a working solution you can watch handle real cases, with a person stepping in wherever a decision carries weight. Only once it has earned its keep do we scale it, wiring it into your systems and turning to the next use case.

Technical implementation

An agentic solution has to earn its place in daily use, which means it has to be dependable, controllable and able to grow with you. We build on whatever platform you already rely on rather than pushing one stack on everyone. Sometimes that is LangChain and LangGraph for flexible open-source orchestration. Sometimes it is Databricks when the data and compute get heavy, IBM watsonx Orchestrate for enterprise-grade deployment, or Microsoft Foundry on Azure.

The foundation changes; the principles do not. A person stays in the loop wherever a decision carries real risk. The agent does the repetitive work and raises its hand the moment something falls outside its rules. And the whole thing stays transparent enough to change later, so it grows with new requirements instead of hardening into a box nobody wants to open.

“Turn repetitive, rules-based work into a process that runs itself, without giving up control of the decisions that matter.”

Product setup and current status

Agentic AI is not a someday idea for us. We have it running in production right now, in places that look nothing alike.

One is Creyten, an AI research assistant for Belgian tax law built on Fisconetplus, which has grown into a product in its own right. Tax lawyers and advisors get precise, sourced answers in seconds instead of losing an afternoon to legal databases. For a manufacturing client we built a pipeline that reads incoming technical input and matches it against the right options in a large catalogue, learning from feedback as it goes. What used to be manual now handles itself. And for a global materials group we showed how a multi-agent buying assistant, wired into SAP Ariba and ServiceNow, can steer employees to the right purchasing channel and cut down on policy slips and rework.

Three sectors, three very different problems. The common thread is that each one took a process that used to rest entirely on people and handed the repetitive part to an agent, without pushing people out of the decisions that count.

Looking ahead

What these projects have in common is not clever technology on its own. It is clever technology meeting people who know the domain and a working relationship close enough to be honest. The companies getting results are not the ones chasing whatever model launched last week. They are the ones that picked a real problem, proved something small and built out from there.

Agentic AI is early and it will keep getting better. Our hunch is that the edge will not go to whoever has the newest model. It will go to the companies that already know how to put one to work.