Applied AI that reaches production

What applied AI means in practice

Applied AI is a model wired into a process that already exists, with a defined input, a defined output and someone accountable for both. It is not a chatbot bolted onto a website, and it is not a research project.

The distinction shows up in one question: what happens when the model is wrong. A demo has no answer. A production system has a confidence threshold, a fallback path and a log that lets somebody find out why.

How an AI system gets to production

Four steps, in this order. Skipping the third is the most common reason a pilot never ships.

1. 01

   Name the decision

   Pick the specific decision or task the model takes over, and write down what a correct answer looks like. If that cannot be written down, the problem is not ready for a model yet.

2. 02

   Ground it in your data

   Retrieval over the documents, records and systems the operation already has, so answers come from your reality instead of the model’s general knowledge. This is where most of the engineering time goes, and it is the part that decides quality.

3. 03

   Evaluate before shipping

   A test set built from real cases, scored against the definition from step one. Without it there is no way to tell an improvement from a regression, and every prompt change becomes a guess.

4. 04

   Monitor quality and cost

   Token cost per call, latency, failure rate and the share of answers a human overrode. Model pricing and behaviour change under you; a system without monitoring degrades silently.

Where applied AI does not pay for itself

A model is the wrong tool when the task is fully deterministic. If the rule can be written as a rule, write the rule: it is cheaper, faster and it never drifts. Invoice matching by exact identifier does not need a language model.

It is also the wrong tool when volume is low. A process that runs eleven times a month rarely repays the evaluation and monitoring that keep an AI system honest, and the honest recommendation there is to automate the surrounding steps and leave the judgement to a person.

The third case is missing data. If the answers the model would need are in nobody’s system — only in someone’s head — the first project is capturing them, not generating them.

Questions we get asked

What is the difference between an AI pilot and a production AI system?

A pilot proves a model can do the task on sample data. A production system handles live data with gaps and duplicates, is scored against a test set on every change, has a fallback path when it fails, tracks cost per call, and enforces access boundaries. The model is the same; the work around it is what differs.

What is RAG and when do you need it?

Retrieval-augmented generation grounds a model’s answers in your own documents and records instead of its general training. You need it whenever a correct answer depends on facts specific to your organisation, which in business operations is almost always.

How do you know if an AI system is working?

With a test set built from real cases and scored against a written definition of a correct answer, plus monitoring of cost per call, latency, failure rate and how often a human overrode the output. Without those, an improvement and a regression look the same.

When should a company not use AI for a process?

When the rule can be written as a rule, when volume is too low to repay evaluation and monitoring, and when the facts the model would need are not in any system yet. In the third case the first project is capturing the data, not generating answers.

Does Verttical build AI agents?

Yes, as part of a system rather than as a standalone product: an agent that takes an action needs the same grounding, evaluation and monitoring as any other AI component, plus a clear boundary on what it is allowed to do without a human.

What does an AI implementation engagement start with?

A discovery stage of two to eight weeks that identifies which decisions in the operation a model can take over and what a correct answer looks like for each. Some of those turn out to be better solved without a model, and that answer is part of the output.

Start with the discovery stage

Tell us which decision in your operation takes the most human time. We reply within one business day.

Related services

• Custom software development

  Systems built for how one operation works, not configured from a template — including the cases where buying beats building.

• Business process automation

  We measure what a process costs to run before automating it, and say so when the answer is to leave it alone.