Ask a team what AI should automate and you'll get a list of forty things, most of them the tasks people dislike rather than the ones that cost the most. The useful question is narrower: which one or two processes, automated properly, would give you back the most time with the least risk? Answering it well takes a method and, more often than not, a visit. This is what we look for when we spend one to three days in a client's office, and how you can apply the same method yourself.

Why we look in person

Every organisation runs two versions of each process: the one in the handbook and the one at the desk. The handbook says orders come in through the portal. The desk says half of them arrive by email as PDFs, someone retypes them into the ERP, checks the price against a spreadsheet and emails the customer back from their own inbox. The expensive part of the work lives in that second version, and it rarely comes up in a meeting because nobody thinks of it as a process. It's just what Tuesday looks like.

That's the idea behind forward deployed engineering, the model OpenAI, Anthropic, AWS and Microsoft all built new units around in 2026: the people who build the system go and see the work before they decide what to build.

How the visit works

The visit takes one to three days, depending on how many teams are involved. The shape rarely changes:

  1. Before we arrive. One short call to agree which systems you use, which teams we'll visit and whether personal or patient data is involved.
  2. Start with the goal. An hour with whoever owns the budget: what would better look like, in hours, errors, response times or revenue?
  3. Sit with the people doing the work. Not a workshop. We watch real tasks being done, ask why each step exists and note every place where information is copied, checked or chased.
  4. Count. How often each task happens, how long it takes and how often it goes wrong. Rough numbers are fine; no numbers are not.
  5. Check the plumbing. Which systems have usable APIs or exports, who controls access, and where personal data flows.
  6. Read back the shortlist. Before we leave, a plain conversation about what we found, what we'd automate first and what we'd leave alone.

Afterwards you get a written scope with a fixed price and a date for the first system. Not a slide deck, and not an invitation to a longer discovery phase.

The six questions we score every candidate on

Each candidate process gets a score from one to three on six questions. It's deliberately crude. The point is to make the trade-offs visible, not to produce false precision.

QuestionScores high whenScores low when
VolumeIt happens daily, dozens or hundreds of timesIt happens a few times a month
Time per instanceEach one takes minutes of skilled attentionEach one takes seconds
Rule clarityA new colleague could learn it in a dayIt depends on years of judgement
Input formatEmails, forms or documents with a recognisable shapeOpen-ended conversations on any subject
Data accessThe systems involved have APIs or reliable exportsThe data lives in someone's head or a locked legacy system
Cost of a mistakeErrors are cheap and easy to catch with a human checkOne error harms a patient, a customer or your licence to operate

An illustrative example: imagine a wholesaler whose sales team receives 60 order emails a day, each retyped into the ERP in about six minutes. That's six hours of skilled time every day. The rules are clear, the inputs are recognisable, the ERP has an API, and a person can approve each order before it's booked. It scores high on almost everything. The same company's annual pricing review scores low: it happens once a year and runs on judgement.

The best first candidate is rarely the one with the single highest score. It's the one with no low scores: frequent, clear, reachable and safe to check. A process that scores three on volume but one on cost of a mistake can still be automated, but it needs a very different design, with a person reviewing every output.

The processes that usually make the list

Every business is different, but the same patterns turn up again and again:

Notice how few of these are glamorous. That's the point. MIT's 2025 study of generative AI in business found that more than half of generative AI budgets went to sales and marketing tools, while the biggest returns came from back-office automation (Fortune). Dutch businesses show a similar tilt: among companies using AI, 35% use it for marketing or sales and 32% for administration or management tasks (CBS). For more ideas by department, see our list of practical AI automation ideas.

The processes we usually leave alone

Quick wins first, foundations second

Order matters as much as selection. A good first system is narrow enough to ship in weeks, visible enough that people notice, and judged against a baseline taken before the build. That baseline is how you'll calculate the return on the automation, and it's what earns the second project. Foundations such as a shared data layer or an integration hub come next, once there's a result to justify them. If you're weighing an agent that works across several systems, our guide to reliable AI agent workflows covers the design questions.

How to prepare your team for the visit

Most companies don't need an AI strategy to start. They need one well-chosen process, automated properly and measured honestly. Neurova AI builds custom AI applications and integrations that way, from Eindhoven, for clients across Europe. If you'd like us to find yours, ask for an on-site visit.

Frequently asked questions

Which business processes are best to automate with AI? Processes that are frequent, time-consuming, clear enough to check and fed by data you can access. Typical examples are email and document intake, invoice and order processing, quote preparation, scheduling, first-line customer questions, internal knowledge lookup and routine reporting. Rare, judgement-heavy decisions are usually poor first candidates.

How do you identify automation opportunities in a business? Watch the work rather than only asking about it. Sit with the people who do it, count how often each task happens and how long it takes, and note where information is retyped between systems. Then score each candidate on volume, time, rule clarity, input format, data access and the cost of a mistake.

What should not be automated with AI? Low-volume work that depends on expert judgement, decisions about people such as hiring or credit, made without human review, and processes that are broken or about to change. Automating a bad process only makes it fail faster, so fix the process first.

Written by Andreas Chitos, founder of Neurova AI. He builds AI systems and medical software from Eindhoven, the Netherlands. Get in touch.