Back to Blog
What AI agents are actually worth: 80% less manual documentation work at CCOM
Customer Stories7 min read

What AI agents are actually worth: 80% less manual documentation work at CCOM

CCOM AS cut manual documentation work by 80% and improved quality at the same time. The return did not come from a clever model. It came from doing it in a structured way, one workflow at a time, on a platform built to run agents in production.

Claus Fasseland

There is a kind of work that never shows up in a business case, because nobody ever decided to do it. Documents arrive. Someone reads them. Someone works out what matters. Someone keeps a register up to date. Someone chases a supplier for a missing piece of paper, and then chases again two weeks later.

It is not hard work, and that is exactly why it is expensive. It absorbs good people for hours every week, it never finishes, and it gets worse as the business grows.

What one customer got out of it

CCOM AS is a Norwegian maritime services company. They have been running a team of Symbi agents across that kind of process since the spring of 2026. Their own assessment:

Symbi-built agents have reduced manual documentation work by 80%, and strengthened quality at the same time. The platform makes it straightforward to operate the agents and keep oversight of them.

CCOM AS

Read the second half again, because it is the part that decides whether the first half lasts. Cutting the effort is easy to demo. Keeping quality while you cut it, and being able to show afterwards what was done and by whom, is what makes it something you can actually run a business on.

The return comes from the structure, not the model

Plenty of companies have tried AI on this kind of work and got very little back. Usually it is the same story. Someone pastes documents into a chat tool, gets good answers, and nothing changes, because the answers live in a chat window and the register still needs updating by hand. Or a developer writes a script, it works for a month, then the format changes and nobody owns it.

The difference is not the model. It is whether the work is set up as an operation.

That means a few unglamorous things. Each job runs on a schedule or on a trigger, without anyone remembering to start it. Its output goes into the systems the business already uses, not into a new place people have to check. Somebody can see what it did last week. And when it is wrong, there is a way to correct it that does not involve rewriting anything.

Start with one job, and let it prove itself

We do not begin with a platform rollout, and we do not recommend anyone else does either.

Pick the single worst job in the process. Automate that one. Let it run in production for a few weeks while people watch it. Then look again, because the next bottleneck is usually somewhere nobody predicted, and it is far cheaper to find that out by running than by planning.

The financial logic here is simple. If the first job does not pay for itself, you have spent a few weeks, not a year, and you stop. If it does pay, you have a working piece of the process and a much better idea of what the next one is worth. Every step after that is a smaller decision made with better information.

CCOM got to a full team of agents this way, one at a time, over about five months. At no point was there a big bet to approve.

Why several agents instead of one

The most common question we get is why not build one capable agent that does everything.

The reason is cost and control, and they pull in the same direction. Different steps deserve different treatment:

  • Judging whether something is relevant is a decision. It should come with a written reason and a confidence score so a person can check the thinking, not just the answer.
  • Sending an email to a customer or a supplier cannot be undone. A person approves it before it goes out, every time.
  • Removing duplicates from a register is mechanical. It should simply run, with nobody watching.
  • Writing the audit trail involves no judgement at all. It makes no AI call, so it costs nothing per run and can go as often as you like.

Put all of that in one agent and you have to choose one setting for the lot. Either you make a person approve the mechanical work, which throws away most of the saving, or you let the irreversible work run unsupervised, which is how organisations end up not trusting the system at all.

Where that line sits is a setting, not a rule we impose. You decide which kinds of action need a person and which should simply run, and you move the line as trust builds. Plenty of steps are better off fully automatic from day one, and paying someone to approve them would be the expensive mistake.

Splitting the work means each step is only as automatic as it deserves to be. It also means that when something goes wrong, and it will, you know which job to look at.

What the Symbi app actually does for you

An agent that works in a demo and an agent that is still running in month six are different products. Almost all of that difference is operational, and it is what the Symbi app is built around.

The agents work inside your systems. They read and write your spreadsheets, your SharePoint, your helpdesk. Your data stays where it already is. There is no shared pool and nothing is pooled across customers. If you stopped using Symbi tomorrow, your register would still be sitting where it always was, up to date.

You choose what waits for a person. By default anything irreversible does: outbound messages are drafted by an agent and appear in a queue for approval, with the text there to edit before it goes, and whoever approved it is recorded. Where a step does not warrant that, you set it to run on its own. Most teams start cautious and loosen it as they see the work.

Uncertainty is visible. When an agent is not sure, it marks the item as needing human review instead of guessing. That queue is a work list, not a hidden failure.

You can change how an agent works without a developer. Instructions are edited in the app and published as a new version. Publishing runs a check first and refuses the change if something is wrong, it makes you write down what you changed and why, and it records who published it.

Every run is logged, with its cost. You can see what each agent did, why, and what it spent. That matters more than people expect at approval time, because "it saved us time" is an argument and "here is what it did, and it cost this" is evidence.

That last point is why CCOM's assessment mentions oversight in the same breath as the 80%. The saving is what you buy. The oversight is what lets you keep it.

This is not really about maritime

The nouns change from business to business. Invoices instead of order lines, contracts instead of declarations, applications instead of components. The shape is the same everywhere: important, repetitive, document-heavy work that quietly takes up a day or more of somebody's week.

If that describes something in your back office, three things carry over from a project like CCOM's:

  1. Start with one workflow, not a platform. Take the worst job first.
  2. Split the work by how much control each step needs, not by how much one model could technically handle.
  3. Decide early where the evidence lives. The register is what you build. The proof is what you will be asked for, and it is the easiest thing to leave until it is too late to reconstruct.

If your back office has a queue that never empties, or a register that is only as current as the last thing someone remembered to type in, it is worth a conversation. Book a 30 minute Agent Audit and we will find the one workflow most worth handing over first. No obligation, and you stay in control the whole way.