AI
Which processes can be automated?
A process can be automated when it runs often, when the result can be checked, when an error can be corrected before the customer notices, when someone owns the rules and when the person approving has time for the volume. Miss one of the five and the work moves somewhere else.
A process can be automated when it runs often, when a correct result can be checked afterwards, when an error can be corrected before it reaches the customer, when someone owns the rules and when the person approving has time for the volume. The short version: a process can be automated when an error is cheaper to correct than the work is to do.
Apps AS automates processes in its own products and for customers. This article draws on the process that writes the articles on apps.no, on how Threll.ai and Munin are set up, and on a count made in Apps AS' own CMS on 5 October 2026.
What does it mean for a process to tolerate automation?
A process tolerates automation when a step of the work can be removed without the result getting worse, and without the risk growing larger than the gain. Tolerating it means the process still produces the right result when nobody looks at each individual case.
The bar is higher than whether something can be automated at all. Almost any task can be run by an agent. The question is what happens on the day the agent gets it wrong, and who notices the error. A process that only works as long as somebody reads through every result is digitized. The work has moved to a screen and is still there.
Which two questions matter most?
Two questions matter most: how often the process runs, and what an error costs to correct. Frequency decides whether building is worth it. The cost of an error decides where the person has to stand.
| The error is cheap to correct | The error is costly to correct | |
|---|---|---|
| The process runs often | Automate fully. A wrong article draft on apps.no is thrown away before any customer sees it. | Automate, with an approval in front of the action that cannot be undone. |
| The process runs rarely | Digitize with a form and a template. Building and running an agent costs more than the work it removes. | Leave the job to a person. There are too few cases to learn from, and each error costs too much. |
The top right of the table is where most agentic projects end up, and where the approval has to be placed correctly. Threll.ai, the voice agents Apps AS develops, splits an outbound call into two operations: the agent prepares the call, and the call is only placed once a confirmation is given. A phone call cannot be taken back, so the confirmation sits in front of the action.
Which five criteria do we check before automating?
Apps AS checks five criteria in a fixed order before a process is automated. If the process fails on one criterion, we fix the process before anything is built.
- Criterion 1
Frequency: does the process run often enough? Set the cost of building, running and the review that remains against the work removed in one year. The process fails when the sum comes out negative.
- Criterion 2
Ground truth: can a correct result be checked afterwards? The check has to run against something outside the system, such as a paid invoice, a closed case or an answer the customer accepted. The process fails when the only thing able to judge the result is the system that produced it.
- Criterion 3
Reversibility: can an error be corrected before the customer notices? A draft, a proposal and an internal record can be corrected. A sent email, a payment and a phone call cannot. The process fails when the action has left the building the moment it is taken, and no approval stands in front of it.
- Criterion 4
Owner: is there a named person who can change the rules? The process fails when the rules exist only in the head of whoever did the job before.
- Criterion 5
Capacity: does the person approving have time for the volume? Automation raises the volume arriving at the step that still needs a person. The process fails when the approval step becomes the queue.
The first four criteria are familiar from process improvement. The fifth criterion is the one most often missing from the plan, because capacity in the approval step only becomes a problem once the agent works.
Gartner predicted in June 2025 that over 40 percent of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls. The costs correspond to the frequency criterion, the value to the ground-truth and owner criteria, and the risk controls to the reversibility criterion. Capacity in the approval step is not on Gartner's list.
Why does the approval step become the queue?
The approval step becomes the queue because automation removes everything that used to take time before the approval. The approval step was small as long as the rest of the process was slow. Once the rest runs by itself, the approval is the only step left.
The process that writes the articles on apps.no shows the pattern. An agent picks a topic from the weekly plan, reads the archive, writes the Norwegian text, translates it, makes the image and files everything in the CMS as a draft. One step is not automated: a person reads it and presses publish. A published text can be withdrawn, but by then it may already have been read and indexed.
The figure 2 of 7 says nothing about the quality of the drafts. The figure says where a faster process has to find its speed: in the approval step. A faster agent only makes the queue longer. Why the human review is a fixed cost is covered in Manual review of AI agents is not a cost you can remove. How to measure whether a step of the work has actually been removed is covered in How do you measure whether automation worked?.
What do you do with a process that cannot be automated yet?
A process that cannot be automated yet is fixed on the criterion it fails. The agent gets built afterwards.
- If the process fails on ground truth, write down what a correct result is, with ten examples two people agree on.
- If the process fails on reversibility, split the action into a proposal the agent makes and an execution a person confirms.
- If the process fails on ownership, give the process a name and a date for its next review before anything else happens.
- If the process fails on capacity, split the approval by risk. In Munin, Apps AS' own customer platform, you decide which workflows run unattended, which escalate and which wait for approval. The split between the three decides how much a person has to read.
Why an automated process needs a named owner after launch is covered in the anchor article Personalization is an operations task, not a campaign. Why agents earn most in the handovers between people is covered in Where AI agents actually earn their keep.
Place one process in the table
Pick one process you are considering automating. Write down how many times a month it runs, and what it cost the last time an error in the process had to be corrected. Send the two numbers to hello@apps.no. We will reply with which of the five criteria the process fails on first, and what we would fix before anything is built.
Sources
- Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, press release 25 June 2025, read 5 October 2026. gartner.com
- Apps AS, the Norwegian article archive on apps.no, count of article drafts created 28 September–2 October 2026, counted 5 October 2026. apps.no
- Apps AS, Munin, case page read 5 October 2026. apps.no
- Apps AS, Threll.ai, case page read 5 October 2026. apps.no
About this article. Written by Espen Hareide, co-founder and partner at Apps. A satellite to the pillar From digitization to automation. The table and the five criteria are Apps AS' own. The count of article drafts was made in Apps AS' own CMS on 5 October 2026; it does not separate out who created each draft, and does not say why the five unpublished drafts are where they are. The description of outbound calls in Threll.ai is based on Apps AS' own product setup as documented on 5 October 2026. Gartner's figure is a forecast from June 2025, not a measured result. This article uses no customer figures. Translated from the Norwegian original.
FAQ
- How many times a month does a process need to run before automating it pays off?
- There is no fixed threshold. Work out what it costs to build and run the agent for a year, add the time a person still spends on review, and compare that with the hours that actually disappear. Processes that run only a few times a month almost always lose that calculation.
- Can we automate a process that has a lot of exceptions?
- Count the exceptions first. If they are a minority, the agent can handle the main path and pass the exceptions on to a person. If the exceptions are the majority, you have many small processes rather than one, and you should pick the largest of them and start there.
- Who should approve what the agent does?
- The person who owns the process, or someone the owner appoints. The developer who built the agent is poorly placed, because the developer judges whether the agent works and not whether the result is right for the customer. Write down one name and how many cases that person can get through in a day.
- What do we do if only one employee knows what a correct result looks like?
- Set aside time with that employee and write down ten or twenty examples with the correct answer before anything is built. The examples become the ground truth the agent is tested against. Without them the process stops on the day that employee is off sick or leaves, with or without an agent.
- Should we start with the process that takes the most time today?
- Not necessarily. The processes that take the most time often have costly errors and many exceptions. Start instead with a frequent process where errors are cheap to correct, and learn how operations and approval work there before you take on the heavy process.
