A customer reads an AI-written service message and still does not know whether the appointment is booked. The company saved time drafting the email. The customer now has to call, and someone must resolve the uncertainty. Faster writing has moved work rather than removed it.
For Norwegian businesses using AI in customer communication, the approval question should be concrete: Can the intended reader work out what happens next, what they must do and what the company has actually promised? Grammar and a professional tone are necessary checks. They do not answer that question.
New research gives a reason to look beyond sentence length
An English-language preprint by Thomas Hikaru Clark, Matias B. Martin and Kyle Mahowald, submitted on 21 October 2026, found that leading models favoured several language patterns associated with reading difficulty. Its text comparisons and separate training experiment suggest that pressure to compress information deserves scrutiny. This was not a human comprehension trial, a Norwegian-language study or proof of how commercial models were trained. Results also varied by model and task. It does not establish that AI writing is always harder to understand.
The management implication is to assess the text in the situation where it will be used. A short sentence may leave the reader to reconstruct who is responsible, which condition applies and whether a deadline is theirs or the company's. A longer explanation may make the task easier.
This matters beyond a few experimental users. Statistics Norway's September 2026 figures show that 48% of enterprises with at least ten employees use one or more AI technologies, across the surveyed industries excluding finance. That is an adoption figure, not a measure of customer writing or its quality.
The missing information may be a business decision
Consider a hypothetical Norwegian service company. An appointment has been cancelled. The customer must reply with two suitable alternatives by Thursday at noon. The company will confirm a new appointment by Friday at 16:00. Until that confirmation arrives, the new appointment is not booked.
A deliberately constructed poor message reads:
“Rescheduling is subject to capacity clarification and confirmation of customer availability.”
There are no spelling errors. But who should contact whom? Is the customer waiting for a call? Has the company reserved a time? The sentence conceals the work behind abstract nouns and omits essential instructions.
A clearer message, using the facts in the hypothetical brief, would be:
“Your service appointment has been cancelled. Please reply to this email with two suitable times by Thursday at noon. We will confirm a new appointment by Friday at 16:00. Your new appointment is not booked until you receive our confirmation.”
The improvement required information from the person responsible for scheduling. Asking a model to “make it clearer” would not establish those deadlines. If the business has not decided when it will respond, the writer needs that decision. The model must not invent a delivery promise to make the paragraph flow.
In a real message, use calendar dates where weekday names would become ambiguous, particularly in reusable templates. Check that the reply channel works and that someone owns the incoming requests. Clear wording cannot repair a process that has no capacity to fulfil the promise.
These messages are invented examples, not outputs from the study or records of a customer incident.
Ask readers to use the message
Språkrådet's guidance on testing texts recommends involving intended users and checking whether they can find, understand and use the information. It describes lightweight methods as well as larger studies. Its public-sector setting does not prevent a private business from applying that practical approach.
For the service message, give a relevant reader the text and a realistic situation. Ask four questions without explaining the answer first:
- What would you do now?
- When do you need to do it?
- What will the company do, and by when?
- What has already been confirmed, and what is still undecided?
Ask the reader to point to the words supporting each answer. If they reach the right conclusion only after an employee explains it aloud, the message has not done the job on its own.
Test the actual format. An email on a phone, an SMS and a letter can expose different problems. An instruction hidden below a long introduction may be missed even when every sentence is understandable.
A few conversations will uncover problems, but they will not establish an error rate for the whole customer base. Include readers with different familiarity with the service and different Norwegian-language proficiency. Revise the message and try it with fresh readers. Having the same model impersonate customers is not a substitute for observing people use the text, as discussed in the article on synthetic customer research.
Give AI a useful brief and a clear limit
AI is useful for producing alternatives, adapting an approved message to another channel and finding phrases that deserve explanation. There is no need to discard those benefits. Give the tool the information that makes a good draft possible:
Reader and task: who receives the message, what they already know and what they need to do.
Approved facts: the actual dates, prices, conditions, contact route and commitments. Identify what remains undecided.
Writing instruction: use ordinary language, name who does what and put the next action where the reader will find it. Do not add facts or promises. Flag missing information for the responsible employee.
Do not impose an arbitrary word limit that forces essential conditions out of the message. A short summary can help, provided the reader can reach the complete terms and distinguish the summary from them. Nor should every technical term disappear. Explain terms the reader needs to make the decision, rather than replacing them with something inaccurate.
For a Norwegian business publishing in both English and Bokmål, review the two versions against the same factual brief. Check ordinary Norwegian usage, names of services and the strength of commitments. “We expect to reply” and “we will reply” are different promises. An English draft that reads well is not evidence that its Norwegian translation works.
Use approved tools and permitted information. Testing a fictional message usually does not require uploading customer records. The employee AI policy guide covers the separate decision about allowed uses and information.
Manage the whole customer outcome
Start with one frequently used message that already causes questions. Give a customer-service or operations owner responsibility for its facts, testing and revision. Record misunderstandings, corrections and repeat contacts alongside drafting and review time. Falling contact volume alone is not proof of success if customers have given up or cannot reach you.
Scale the checking to the consequence. A routine appointment reminder should not need a committee. Instructions involving safety, substantial payments or contractual rights need the relevant specialist as well as language review. When a template, service condition or model changes materially, check the affected messages again.
Management owns this process. If AI-generated communication becomes important to the company's service delivery or risk, the board should ask what evidence shows that customers understand material messages and who corrects failures. Directors do not need to approve individual emails.
Choose one message this week. Put it in front of someone who would actually receive it. Their account of what they would do next is more useful than another assurance that the text sounds professional.
Sources checked on 27 October 2026. The service company, messages and proposed acceptance questions are illustrative. The working method is the author's recommendation, not a claimed measured improvement or a legal standard for all private companies. The illustration is AI-generated.