Services · AI automation
AI automation, where it solves a real task
AI handles a narrow set of tasks well: classification, extracting information from unstructured text, summarisation, retrieval, content-based routing. The rest of a business flow is deterministic logic, which is cheaper, more predictable and easier to debug. Our job is to say which is which.
Our position
In a market where every vendor says "AI", the most useful thing we can offer is an honest boundary.
AI fits when the input is unstructured and the rule cannot be written: an email from which the client’s intent must be understood, an invoice in an unknown format, a description from which an attribute must be pulled. AI does not fit when the rule exists: if the value exceeds a threshold, if the supplier is on list X, if the date has passed. There a simple condition is better on every measure.
Tasks AI does well
- Classification: which category a request, email or document belongs to.
- Extraction: which amount, date or company name appears in free text.
- Summarisation: what a long conversation thread essentially says.
- Retrieval: which document or answer is relevant to a question.
- Content-based routing: who it should reach, judging by what it says.
Tasks we do not recommend AI for
- Calculations. A formula is exact; a model estimates.
- Written business rules. If it can be expressed as a condition, write it as a condition.
- Irreversible decisions without human review.
- Any step where an error cannot be detected afterwards.
What a correctly built AI flow looks like
Select a step for detail.
Unstructured input
An email, a scanned document, a description written by a client. If the input is already structured, the AI step is pointless and gets removed.
A strictly bounded task
The model gets a narrow, verifiable task: "extract the amount and the due date", not "process the invoice". A broad task produces output you cannot validate.
Output with a confidence score
Every result carries a confidence measure. Without it you cannot build a responsible flow — you have no way of knowing what deserves review.
Human review threshold
Above the threshold the flow continues automatically. Below it, a person sees it. The threshold is calibrated on your real data, not on a default.
Mechanical validation after
The result is checked with deterministic logic: is the amount within a plausible range? Is the date in the future? Does the company exist in the system? AI proposes, validation confirms.
Corrections are recorded
When a person corrects a result, the correction is logged. That is how you see, after a month, where the model is systematically wrong — and intervene there.
Cost ceiling
Every AI step has a variable cost. The flow has a cap: per request, per day, per month. A flow without a cap is an open financial risk.
Reference architecture — not a client result.
What affects the cost
The volume of inputs processed, because models are billed by usage. A flow handling thousands of documents a month has a real recurring cost that must be budgeted up front, not discovered later.
How narrow the task is. A well-bounded task needs a smaller, cheaper model.
The accuracy required. Going from 90% to 98% often costs more than the entire rest of the flow, and sometimes reviewing the remaining 10% by hand is cheaper than the difference.
Frequently asked
Do we need AI?
Probably not for most steps. At the audit we tell you which step justifies it and which does not. If none does, we say so — automation works perfectly well without it.
Does our data end up training a model?
That depends on the provider and plan, and it is settled explicitly before implementation rather than assumed. Where the requirement is that data must not leave a given perimeter, we either choose an architecture that respects it or drop the AI step.
What happens when it is wrong?
That is what the confidence threshold and the validation step are for. An AI flow without review is not automation, it is guessing at scale.
What does it cost monthly?
It has a recurring cost that depends on volume, and we estimate it before implementation with the assumptions visible. We do not build a flow whose monthly cost you cannot anticipate.
Other solutions
Want to find out whether your process is worth automating?
Free initial audit, 60–120 minutes. No obligation to implement. If automation does not make enough economic or operational sense, we will tell you.