Skip to main content
AI

When not to use AI in operations: four cases

When not to use AI in operations: exact results, costs that grow with volume, decisions you have to explain, and processes that already work.

Borja Martínez Azcarraga · Consultant

May 4, 20266 min readUpdated on October 21, 2026

A language model earns its place when the task is about interpreting free text and a mistake is cheap to spot and fix. It can summarise a call, sort an ambiguous email or draft a reply that someone reviews. When the task needs something else (an exact result, a decision you have to justify, or a high volume of cases a rule already handles), a written rule, a formula or an automation without AI usually does the job better. These are the four cases where it pays to stop before adding it.

1. When the result has to be exact and repeatable

Working out the VAT on an invoice, checking a tax ID or applying a price list each have one valid answer. Take the check letter of a Spanish tax ID (NIF): you divide the number by 23 and look up the remainder in a table published by Spain's Ministry of the Interior. Given the same number, the formula returns the same letter.

A language model doesn't work that way. OpenAI's API documentation says its outputs are non-deterministic by default and that, even when you fix the parameters to make them repeat, repetition isn't guaranteed. For a draft, that's fine. For an amount that goes into your books, it isn't.

Decision: if a formula or a rule gives the right answer, use it. If the data arrives as free text, AI can extract it, but the rule does the calculation and the check.

2. When the cost grows with volume and a rule already does the job

Every call to a model has a cost that depends on the provider, the model and the length of the text. A rule based on sender, subject line or specific words costs nothing per run. With a handful of cases a month you'll barely notice the difference. With thousands, you'll see it on the bill and in processing time.

Decision: before choosing, work out the monthly cost with your provider's current prices: cases per month times the average cost per call. Compare it with the cost of maintaining the rule. If the rule covers most cases, keep AI for the ones the rule can't sort.

3. When you have to explain every decision

In hiring, credit approval or a disciplinary decision, "the model decided" isn't an explanation. In the EU, Article 22 of the GDPR gives people the right not to be subject to a decision based solely on automated processing when it produces legal effects or similarly significant effects, with exceptions and with safeguards such as human intervention. The EU AI Act also classes systems used to recruit people or to decide on promotions and terminations as high-risk (Annex III, point 4). Following the 2026 AI Omnibus, those obligations apply from 2 December 2027.

Decision: in these processes, the rule and the reason for each decision are written down and a person decides. AI can organise the paperwork or prepare a summary, as long as the person deciding can check the original. If you're unsure whether your case falls under these rules, or which rules apply outside the EU, get legal advice before you roll it out.

4. When the process already works

Adding AI to a workflow that works, without a specific problem to solve, adds pieces to maintain: one more provider, instructions to adjust when the model changes, edge cases to review. The workflow doesn't get better and the maintenance grows.

Decision: before you touch it, write down in one sentence what problem AI solves (time, errors, delays), how you'll measure it and who will maintain the change. If you can't answer all three, leave it as it is.

An example: the finance inbox

Picture the finance inbox at a services company. Supplier invoices, customer questions about payments, bank notices and marketing emails all land there. Today someone goes through it by hand every morning and passes invoices on to accounting.

This is how it looks with AI only where it helps:

  1. Data that comes in: an email from a regular supplier with a PDF invoice.
  2. Owner: the finance team member who reviews the inbox today.
  3. Rules first: known senders are sorted by address and subject line, without AI. Marketing emails are filtered the same way.
  4. AI where there's free text: supplier, tax ID, net amount, VAT and total are extracted from the invoice. Emails no rule can sort get a suggested category.
  5. Rule-based check: the tax ID is validated with its formula, and net plus VAT has to match the total. The supplier has to exist in the accounting master data.
  6. Exception: if any check fails or the sender is new, the email goes to a review queue with the reason. No amount reaches accounting without passing the checks.

Observable result: the finance team member no longer reviews the whole inbox, only the exceptions queue, and each exception says which check failed. If the model extracts an amount wrongly, the sum doesn't add up and the error shows before anything is booked.

Where it does fit

Outside those four cases, AI tends to help with free-text tasks that tolerate errors: summarising meetings or calls, suggesting a category that someone confirms, searching internal documentation or drafting replies that a person reviews before they go out. In all of them, someone knows the result is a suggestion and has an easy way to correct it.

Four questions for this week

Pick a process where you're thinking of adding AI and answer in writing:

  1. Is there a single correct answer that a formula or a rule can give?
  2. Do many cases a month go through it that a rule could handle? Work out what each call would cost at current prices.
  3. Will someone have to explain to a person or an auditor why each decision was made?
  4. What specific problem does the process have today, how will you measure it and who will maintain the change?

If the first three answers are "no" and the fourth names a specific problem with an owner, AI has a place. If any of the first three is "yes", start with the rule and keep AI for what the rule doesn't cover.

If you want to run this assessment in-house, there's AI training for companies and an extension on evaluating AI use cases.

Next step

If, after the four questions, you have a process where AI fits but it needs to be integrated with your tools and with checks like the ones in the example, see how we approach AI strategy or book a call. Come with the process in mind and the check that should stop an error before it reaches a customer or your books.

Sources

  1. OpenAI, API documentation: Advanced usage, "Reproducible outputs": outputs are non-deterministic by default and the seed parameter offers best-effort reproducibility. Vendor documentation; other providers may behave differently.
  2. Spain's Ministry of the Interior, NIF/NIE check digit calculation (in Spanish).
  3. Regulation (EU) 2016/679 (GDPR), Article 22.
  4. Regulation (EU) 2024/1689 (AI Act), Annex III, and European Commission, AI Omnibus enters into force (27 July 2026), which sets 2 December 2027 for Annex III high-risk rules. Legal texts; not a substitute for legal advice. Sources checked on 21 October 2026.

Let's talk.

Book a free intro session so we can understand where you stand and how we can help. No strings attached.