Artificial intelligence applied to business processes is not “put a chatbot on the website”. An assistant that answers public FAQs can help, but it rarely moves margins. The cases that pay are those where AI reads, classifies or proposes on a flow that is slow, repetitive or error-prone — and a person stays accountable for the outcome.

If you start from the tool (“we want gen AI”) you will get a demo. If you start from the process (“every quote takes three hours of copying from PDFs and the ERP”) you can measure whether a model cuts that time without inventing data.

Five uses we see in SMEs, without science fiction

1. Extraction from documents. Delivery notes, invoices, specs, order emails: structured fields from messy text. AI proposes; the ERP or custom software validates. The saving is data-entry hours, not the model’s “creativity”.

2. Classification and routing. Tickets, certified email, form requests: category, urgency, queue. It does not replace support. It stops everything landing in one inbox and sales opening the same attachment three times.

3. Search over internal knowledge. Manuals, closed jobs, price lists, procedures. An engine that answers with citations beats a chatbot that improvises. Quality depends on permissions and a tidy document base, not on the fashionable model.

4. Operational forecasts, not oracles. Reorder points, likelihood a deal closes, anomalous consumption: often classical models, fitter than chat. You need clean history. If data lives in five spreadsheets, integrate first, then forecast.

5. Quality and compliance checks. Compare a quote to standard terms, required fields, values out of range. AI is a second pair of eyes, not the final signature.

In all of these the software around the model matters more than the model: where data lives, who approves, how an error is corrected, how you record what the system did.

What is a bad first project

Replacing salespeople. Letting the model decide discounts or invoices alone. “An AI that understands the company” with no list of documents and systems. Training a private model when an existing one plus process rules would do. Sending personal data to a public service with no agreement and no filter.

AI amplifies the process you have. If the process is chaotic, it amplifies chaos — with a confident tone that misleads.

Four questions to judge a use case

  1. Which repetitive activity eats hours every week and has a checkable output?
  2. Is the input already digital (files, mail, database) or still on paper?
  3. Can a person fix an AI mistake in minutes, or can it cause immediate harm (money, legal, safety)?
  4. Who owns the final yes?

If question three is “immediate harm” and nobody signs, it is not a first case. If question one is clear and three is “we correct it in the UI”, it deserves a short prototype.

From trial to process

A pilot lasts weeks, not a quarter of slides: one flow, one document type, one user group, one metric (time, errors, items processed). Only if the numbers hold do you wire CRM, ERP or custom software and add audit and fallback.

DPH embeds models and automation in processes; we do not sell a disconnected chatbot. If you have a concrete flow to shorten, describe it in a quote request: we will say whether AI belongs there or whether you need an integration or a custom application first.