Your team already uses a CRM, a customer portal or an internal application. You want to add AI, but replacing the software would disrupt work that runs well today. The first decision is which task to improve within that existing setup.
AI integration into existing software involves more than connecting a model. The feature needs access to the right records and a defined place in the workflow. You also need to decide who checks its output and what happens when it cannot complete a request.
Choose a first project you can evaluate with real examples. Assign someone who understands the work to review the results before you extend it to other teams.
Choose a workflow with a measurable problem
Start with the task that takes too much effort today. Describe who performs it, what information they use and what happens after they finish.
For example, a request to “use AI in customer service” leaves most of the scope undecided. A more useful brief would ask for draft answers to delivery questions, based on order records and the approved shipping policy. A support employee would check each answer before sending it.
Record a baseline before implementation. Depending on the task, this might include handling time, correction rate, unanswered requests or the amount of work completed without escalation. Include review time when comparing the new process with the old one.
Document processing
AI can help extract information from incoming documents and prepare it for an existing system. A useful starting point could be reading supplier forms and proposing structured records for review.
The integration still needs rules for missing fields, contradictory values, duplicate documents and unreadable files. The system should show the source behind important extracted values so reviewers can verify them quickly.
Internal knowledge search
Employees may need answers spread across manuals, policies and project documentation. An assistant can retrieve relevant passages and draft an answer with source references.
Access permissions must apply before restricted material reaches the model. An employee who cannot open a document should not discover its contents through a generated summary. The feature also needs a process for changed, removed and conflicting documents.
Customer support assistance
Begin with suggestions to support staff when the cost of a wrong answer is meaningful. The assistant might classify a request, retrieve relevant account details and draft a response.
Distinguish preparing a reply from taking an action. Issuing refunds, changing account details or making commitments to a customer requires explicit business rules and appropriate approval.
Decide whether the task needs AI
Some work is better handled by a standard integration. Copying an approved invoice total into accounting software usually needs reliable field mapping and error handling. Interpreting inconsistent descriptions in incoming documents may benefit from a model.
Use predictable rules for calculations, permissions and state changes. Use AI where interpreting language or varied inputs is valuable. Combining the two keeps the workflow easier to inspect.
Anthropic recommends starting with the simplest approach that works in its engineering guide to effective agents. For your project, compare a model-based feature against a rules-based alternative on the same examples before adding more complexity.
How AI connects to an existing application
A typical integration retrieves the data needed for a request and passes it to a model. The application checks the response before displaying it. If the result will change a record, an additional step verifies that the action is allowed and the proposed values are valid.
For an order assistant, this means the software retrieves the correct order, supplies relevant shipping information and checks which changes the current user may make. The model should not decide the user’s permissions.
The connection method depends on your software. A documented API is usually easier to maintain than automation that operates a changing user interface. Missing APIs, poor data quality and unclear ownership can become larger parts of the project than the AI component itself.
OWASP’s authorization guidance remains relevant here: application access rules should continue to be enforced when an AI feature requests data or performs an action.
What drives the cost of AI integration
Separate implementation cost from ongoing operating cost. A model subscription or API bill covers only part of the project.
Initial development
An estimate should cover the work needed to understand the process and prepare the data. It should also include connections to existing systems, interface changes and testing. Ask how deployment and ongoing maintenance will be handled. Features that change records or serve several user roles need more detailed permissions and acceptance tests.
A draft generator for one internal team has a different scope from an assistant that updates customer accounts across several systems. Ask vendors to explain that difference in their estimates.
Recurring operation
Alongside model usage, budget for hosting and any search infrastructure the feature needs. Someone will also need to monitor failures and check quality after updates. Include retries and human review when calculating cost per completed task.
Use representative examples to estimate normal and peak usage. Check provider prices and data handling terms at the time of procurement because these can change.
An illustrative break-even calculation
Suppose a team processes 2,000 requests per month and the new workflow saves a net three minutes per request, after review and corrections. That releases 100 working hours. At an assumed internal cost of $35 per hour, the capacity has a planning value of $3,500 per month.
If ongoing operation costs $700 per month, that leaves $2,800 in monthly capacity value before recovering implementation cost. These figures are assumptions for the example. They do not represent provider prices or a Five Bricks client result. Time saved becomes cash savings only if spending falls; otherwise, the benefit depends on how the team uses the available hours.
Repeat the calculation with fewer users and smaller time savings. If the business case no longer holds, use the pilot to test those assumptions before committing to a larger build.
Run a pilot with clear acceptance criteria
Select a narrow set of users and representative examples, including difficult and incomplete inputs. Keep an evaluation set separate from the examples used to tune the feature.
Decide what counts as an acceptable result. For document extraction, measure critical field accuracy and reviewer effort. For support drafting, measure whether the answer is supported by the available information and how much editing it needs. For actions that change records, verify both permission checks and the final system state.
Record failures and make it easy for users to complete the task manually. Expand the pilot after the quality, operating cost and workflow fit meet the agreed criteria. Improvements to the model or prompts should be checked against the same examples before release.
What to prepare for a development estimate
Bring a description of one workflow, the systems involved, example inputs and the expected output. Include approximate monthly volume, user roles, review requirements and the consequences of an incorrect result.
If an existing tool already meets those needs, compare its configuration and operating costs with custom work. Our guide to custom applications versus off-the-shelf software can help frame that decision.
Discuss your AI integration with Five Bricks
Contact Five Bricks with the task you want to improve and the software your team uses. An example with confidential information removed will help us understand the input, the expected result and the integration work involved.
Frequently asked questions
Often, a feature can be added through existing APIs or a separate service. The appropriate approach depends on access to data, the application’s architecture and the actions the feature must perform.
Not necessarily. Start by testing whether an existing model with appropriate context can meet the requirement. Consider additional training only when evaluation identifies a specific need and suitable data is available.
Some bounded tasks can. Decide based on demonstrated reliability and the consequences of mistakes. Drafting an internal summary and changing a customer’s payment details should have different approval rules.