AI implementation guide

AI integration starts with one useful workflow.

AI integration connects a model to your existing software, approved information and business process. The work includes data access, application logic, evaluation and ongoing operation—not just adding a chat window.

Start with a task you can describe and assess: finding an answer in approved documents, extracting fields for review, or drafting a response for a team member. Establish what a successful result looks like before choosing a model or requesting a build estimate.

By Elevéa Cloud · Published 6 October 2026

Which business task is a good starting point?

Choose a repeated task with accessible inputs, a clear reviewer and an observable outcome. Measure the current process first: how many requests arrive, how much work each takes and what makes a result acceptable. A task with no reliable source information or no way to assess quality is a poor first pilot.

For example, an internal knowledge assistant could find the relevant section of an approved operating manual and provide a linked answer. A document workflow could suggest invoice fields for a person to verify before an existing system accepts them. These are illustrative scopes, not claims about completed Elevéa client projects.

Aim for a bounded first release. Giving an assistant permission to read a record is a different requirement from allowing it to change that record, send a customer message or approve a payment. Scope and review those actions separately.

Automation, RAG or fine-tuning?

The approaches can be combined, but they solve different problems. Use the simplest approach that meets the workflow and quality requirements, then compare more complex options against that baseline.

Match the approach to the problem
Approach	When to consider it	What still needs engineering
Rules-based automation	The steps and decisions can be stated reliably as rules.	Validate inputs, handle exceptions and monitor failed jobs. A language model may add unnecessary uncertainty.
AI with existing tools	The task involves interpreting language or drafting a structured response.	Connect the model to approved APIs. Validate its output and keep business permissions in application code.
Retrieval-augmented generation (RAG)	Answers need information from documents or records that can change.	Retrieve relevant, permitted sources at request time. Check source quality, freshness and whether the answer is supported.
Fine-tuning	A repeatable behaviour or output pattern remains difficult after improving instructions and examples.	Establish a baseline and a representative test set first. Fine-tuning is not a live knowledge database or a permissions system.

A retrieval system can supply current information, but it can retrieve the wrong passage. A fine-tuned model can follow an output pattern, but it can still produce an incorrect answer. OWASP also notes that neither RAG nor fine-tuning fully mitigates prompt injection. Evaluate the complete application, including tools and permissions.

Prepare sources and access together.

Identify who owns each source, whether it is current and which users may read it. Resolve duplicated or conflicting documents before relying on them for answers. Preserve identifiers so an answer can point back to its source.

Access checks belong in the application and retrieval layer. An instruction such as “only show authorised information” is not a substitute for checking the signed-in user’s permissions before providing records to a model.

Agree the data-handling requirements.

For an international project, confirm the countries of operation, required languages, permitted data locations, retention periods and vendor terms that apply to the actual deployment. Test language and locale requirements with representative examples.

Use redacted samples during scoping. Decide what may appear in operational logs, who can access them and how long they are retained. Do not assume every model, plan or hosting region has identical data practices.

How do you know an AI integration works?

Create a test set before tuning the system. Include ordinary requests, exceptions and requests it should not fulfil. Keep some examples separate from the examples used to improve the system so the final check measures more than familiarity with the demonstrations.

Starting scenarios for a knowledge or document workflow
Scenario	Test input	What to check
Normal request	A representative task with a known, reviewed answer.	Correctness, usefulness, response time and the cost of the complete task.
Missing information	A question that the approved sources cannot answer.	The system explains the gap or escalates instead of inventing a supported answer.
Access boundary	Two users ask for the same record but have different permissions.	Restricted content is not retrieved, quoted or exposed through a tool response.
Untrusted content	A document includes instructions to ignore the task or reveal unrelated information.	The application preserves its access and action boundaries. Record failures; one passing test is not proof of security.
Failed dependency	An API times out or returns a malformed result.	The user receives a useful status, retries are bounded and side effects are not duplicated.
Changed source	A policy or record changes after the first successful answer.	The next answer uses the permitted current source, with stale caches handled explicitly.

Set thresholds with the business owner based on the consequence of an error. Measure answer quality, unsupported statements, failed tasks, latency and cost separately. An average quality score can hide an unacceptable access-control failure. A model-generated grade may help scale review, but should be checked against human judgement on representative samples.

Anthropic’s evaluation guidance recommends specific, measurable, achievable and relevant success criteria. NIST’s voluntary AI Risk Management Framework also treats evaluation and risk management as part of the system lifecycle. Neither a checklist nor use of a particular model certifies the finished product.

How much does an AI integration cost?

A useful estimate separates the implementation from ongoing operation. Implementation includes workflow discovery, data preparation, interfaces, integrations, access controls, evaluation and rollout. Operation includes model use, retrieval or search, hosting, monitoring, human review and maintenance.

Compare the cost of an accepted task.

Cost per accepted task = total operating cost for the period ÷ tasks completed to the agreed standard in that period.

Include failed attempts and retries in total cost. If no tasks meet the standard, this metric cannot be calculated. Compare it with the existing process and review the quality of both.

A low price per model request can be misleading if one task needs several requests, substantial document processing or frequent human correction. Use expected volumes and a tested workflow for an estimate, then check actual usage during a limited pilot. No universal price or return-on-investment figure fits every integration.

What should happen before a wider launch?

Run a limited pilot. Start with a defined group and a clear owner for reviewing failures.
Control actions. Require review for consequential actions and validate tool inputs independently of the model.
Plan failure handling. Set usage limits, timeouts, retry rules and a way to disable the feature without disabling the core product.
Document the handover. Record accounts, permissions, source refresh, monitoring, incident contacts and responsibilities for ongoing changes.
Re-test changes. Model, prompt, retrieval and integration updates can change behaviour. Compare them against the same acceptance set before release.

Security testing, legal review or a more extensive evaluation may be needed for the particular application. These launch checks are a starting point for scoping, not an assurance that any deployment is safe or compliant.

Prepare an AI project brief.

Write down the task, users, source systems, permitted actions, acceptance criteria, expected volume and the person responsible for the result. This makes proposals easier to compare and keeps a demonstration from being mistaken for a finished integration.

Download the project brief

Elevéa Cloud offers AI development and integration services for businesses worldwide. For a customer-facing portal or an internal workflow, the work may also involve web application development or an existing CRM integration.

Sources and scope.

This guide explains buying and scoping decisions. Technical examples are illustrative; delivery claims and client results are limited to the evidence on our company and project pages.

Sources reviewed 6 October 2026.