Short answer
AI reliably helps where a task is high-volume, language- or document-heavy, and tolerant of review: extracting data from supplier documents, drafting replies, classifying and routing enquiries, summarising long threads, and answering internal questions from your own documentation. It performs badly as the final authority on numbers, prices, stock, or compliance decisions, where a deterministic rule in the system is both cheaper and correct.
The test we apply before adding AI
Three questions. Does the task happen often enough that saving minutes matters? Is a near-correct answer useful, or does it have to be exactly right? And is there a human or a rule that can catch a bad output before it reaches a customer or a ledger?
If the answers are yes, near-correct is useful, and yes, AI is a good fit. If the task must be exactly right with no check, the answer is code and constraints, not a model.
Use cases that hold up
- Document intake — pulling line items, quantities, and references out of supplier PDFs and emails into a structured record for review
- Enquiry triage — classifying inbound messages by intent, urgency, and product, then routing with a suggested reply
- Drafting — first-pass quotes, follow-ups, and descriptions that a person edits rather than writes
- Internal answers — a search layer over your own procedures, specs, and past orders so staff stop asking the one person who knows
- Summarising — turning long threads or call notes into a record attached to the customer or order
Use cases that create work
Autonomous customer replies with no review, on anything commercially sensitive. Pricing or discount decisions generated rather than calculated. Stock or availability answers inferred instead of read from the system. Forecasts presented with confidence on thin data. Chat interfaces added to a process that was fine as a form.
The failure mode is not dramatic. It is a slow accumulation of outputs somebody has to check, which is the same manual work you were trying to remove.
AI belongs inside the system, not beside it
Most disappointing AI projects are standalone tools bolted onto a business that still runs on spreadsheets. The model has no reliable data to work with and its output has nowhere to go. Fix the operational layer first: once orders, stock, and customers live in one system, AI has structured context and its output can be written back where it counts.
Written by the Atomberg team, based on systems we build and run for European operators. See client work.