AI implementation: how should your company prepare?

Considering an AI implementation in your company, but unsure how to prepare for the first conversation with a software house? It is easy to start with a model or a tool. Without naming the process AI should support, it is hard to assess scope, available data, risks and how to evaluate the result.

Before choosing technology, describe the task, its users and the decision the solution should support. This makes it easier to decide whether AI is appropriate or whether rules-based automation or a process change would be enough.

This article sets out questions to prepare before an AI implementation: goals, data, security, costs, the technology partner and a pilot.

Why the business goal comes before the technology

It is easy to begin with a model, chatbot or tool. Start by describing the problem you want to solve. Identify the process, the people it affects and the work that currently takes time, requires manual checking or makes decisions harder. Only then can you assess whether AI is a suitable approach.

Possible areas to investigate include handling repetitive enquiries, organising documents or analysing sales information. These are examples to assess, not a promise that AI will be the best solution. In some cases, a clearer process or conventional automation may be enough.

How to describe the goal of an AI implementation

Write down what should change and how you will evaluate it. Replace a broad goal such as “improve customer service” with one specific task. For example, test whether AI can draft answers to repetitive questions and assign someone to review them. Measure the current workflow before a pilot so you can compare the result with a baseline rather than rely on impressions.

Prepare a short description of the process, its exceptions and the outcome you expect. It does not need to be a complete specification. It should give the delivery team enough context to assess feasibility.

Set realistic expectations for AI

AI cannot decide for the company which problem is worth solving or who owns the decision. Its output may need human review, especially when an error could affect a customer, finances or safety. During a pilot, test quality on examples from daily work. Decide what should happen when the system cannot give a reliable answer.

Data quality, availability and provenance

Before speaking with a delivery partner, check what information the selected use case requires, where it is stored and who can authorise its use. Consider completeness, freshness, format and whether sample data reflects situations the team encounters in practice.

Data may be spread across a CRM, databases, spreadsheets and documents. Decide which source is current, who controls access and whether the information includes personal or confidential data. Some data may need to be organised, limited or anonymised before testing.

Assessing technical infrastructure

Map the systems used by the selected process, the interfaces they provide and the constraints that need to be considered. An AI solution may require connections to existing applications, access controls, activity logs or a test environment.

An older system does not always need to be replaced. First establish whether the required data can be accessed safely and the output passed to the right step in the process. The integration scope should follow the specific use case.

Identify the data and risks involved

If the solution will process personal or confidential information, or influence decisions about people, define access, where the data is processed, how long it is retained and how incorrect outputs will be handled. Before sending data to an external service, review the provider’s terms and discuss requirements with your data protection lead or legal adviser.

Obligations depend on the system’s use and the company’s role. The European Commission publishes an official overview of EU AI rules. Use it to identify requirements that may apply to your use case.

Consider risk throughout the system lifecycle

Security does not end at launch. Decide who monitors outputs, how errors are reported and when a person must approve a result. The NIST AI Risk Management Framework organises risk work around Govern, Map, Measure and Manage. It can provide questions to consider, rather than a ready-made implementation specification.

What affects the cost of an AI project

There is no single price for an AI implementation. Cost depends on process scope, data availability, integrations, the level of control required, testing and the chosen infrastructure.

List one-off and recurring costs: problem discovery, development or configuration, integrations, model and infrastructure usage, monitoring, maintenance, team training and handling exceptions.

Match the commercial model to the scope

When scope and acceptance criteria are clear, it is easier to agree a price for a defined stage. If data, output quality or integration feasibility still need to be tested, split the work into stages with a decision point after each one.

Before work starts, agree the stage budget, deliverable, assumptions, exclusions and what must be true before the next phase begins.

Include the cost of running the solution

Allow for recurring model or API usage, data storage and processing, monitoring, support, updates and the time people spend reviewing outputs. The scale depends on usage and the operating model. Decide who will own the solution after launch.

Check how the partner investigates the problem

A portfolio can show the type of work a company delivers, but also consider whether the team asks about the process, data, users and success criteria. Ask how output quality will be tested and what happens when the system is uncertain. Also ask how it connects to existing tools and who maintains it.

Questions for the first conversation

Prepare questions such as: What data is required? What will the pilot involve? Who approves the output? Which cases will the system not handle? How will external service costs be tracked? What will the company receive at the end of each stage?

A suitable partner should also ask about goals, constraints, exceptions and the people who use the process. If a proposal arrives before these points are understood, ask for its assumptions and risks.

Agree on communication

Before starting, decide how progress will be reported, who makes decisions and how scope changes are approved. Regular reviews help the team decide whether to proceed based on test results.

Prepare the team for a change in how work is done

An AI implementation can change how tasks are divided. Employees should know which tasks the system handles, which remain with them and when an output needs review or correction. Involve people who perform the process in mapping exceptions and testing. They know situations that may not appear in formal documentation.

Provide training and a clear escalation path

Users should know how to work with the system, recognise an uncertain answer and report a problem. Prepare a straightforward fallback to manual handling if the solution stops behaving as expected.

Not every employee needs to understand the model technically. Some people should understand the solution’s purpose, its limitations and how it is monitored.

Before launching a pilot, record the current baseline and evaluation criteria. Choose measures that fit the process. These might include task time, corrections, output quality or cases handed to a person.

Choose the examples to test, decide who will review the result and define the quality threshold for continuing. Measure the same indicators before implementation and during the pilot. If the system does not meet the criteria, revise the scope or pause until the cause is understood.

Preparing an AI implementation in a company starts with a process and a specific task, not with choosing a model. Identify the available data, how it will be protected, who will review outputs and how you will decide whether a pilot meets its goals.

If you are unsure whether AI is the right approach, describe the problem and the current way of working. In a free technology consultation, we can discuss a suitable first step: clarifying the process, automation, integration or further project discovery.