For the product you sell

How to add AI features to an existing product

Choose a task your users need help with, check what your data and permissions allow, and plan how to evaluate the result. This guide covers feature options, integration decisions, and the first release.

Choose a task to improve

Look at how people use the product today. Identify a task, the information it requires, and what makes it difficult. Record how the task is done so you can compare a proposed feature with the current process.

These questions can help identify candidates. The examples are illustrative; the answer depends on your users and workflow:

  • Where do users produce a first draft?

    For example, a support representative prepares a reply from plan details. Check whether a suggested draft would help, what sources it needs, and how the person would review it.

  • Where do they wait?

    A request may wait for classification, missing information, or approval. Find out which step causes the delay and whether it needs AI, a simpler rule, or a process change.

  • What information is hard to find?

    Consider a user looking for a clause across contracts. Check whether the documents contain the answer, who can access them, and how the user would verify the result.

For each candidate, agree what a useful result looks like. That may include accepted drafts, task completion time, correction rates, and feedback from the people doing the work.

Four ways to use AI in your product

These patterns help describe what a feature does. The examples are illustrative, and you can combine patterns when the task requires it.

  • Assist

    Prepare a draft or suggestion for someone to review. For example, a support representative checks a proposed reply before sending it. Evaluate the content, the changes reviewers make, and the time the full task takes.

  • Retrieve

    Find answers in documents or records the user is allowed to access, with sources they can inspect. An example is finding a contract’s renewal terms. Our Document Intelligence & Retrieval service covers retrieval and its integration with your tools.

  • Extract and classify

    Prepare structured fields from documents, messages, or images. For example, extract an invoice’s supplier and amount for review before creating a record. Test missing fields, ambiguous values, and how corrections enter the workflow.

  • Automate agreed actions

    Let an agent act within defined permissions. For example, it may route a request under agreed rules and send exceptions to a person. Decide approval requirements for each action using evaluation results and the consequences of an error.

Choose the first feature based on the task, available data, integration effort, and risks. An agent can be a starting point when its scope and controls are clear; a draft also needs evaluation and appropriate review.

Check the product and its constraints

Review the interface, permission rules, data model, and integrations the feature will use. Agree any changes these need before estimating the implementation.

  • Permissions

    Define which records each user and organization can access. Enforce those rules in retrieval and actions, and test requests that must be denied.

  • Latency

    Measure response time with representative requests and workload. Agree how long users can wait for this task and how the interface handles delays, cancellations, and failures.

  • Operating cost

    Estimate the expected workload, request sizes, retries, infrastructure, and review effort. Compare options under that load and define spending limits and monitoring.

  • Evaluation

    Build a test set from representative tasks, including missing data, incorrect inputs, and permission boundaries. Agree acceptance criteria and combine test results with user feedback and production monitoring.

Data readiness

Inspect the data for the proposed task: its quality, structure, access, and gaps. Include any preparation in the plan. Our Data Readiness for AI service helps define that work.

Compare models on your task

Evaluate candidates against the same cases and requirements. Compare quality, response time, total operating cost, data handling, and the work your team would need to maintain them.

Hosted models via API
Check quality on your cases, service limits, data processing locations, retention and training terms, and cost at the expected workload.
Open-weight models
Assess the license, hosting requirements, performance, and operational work. Compare the full cost and data flow with the alternatives for your workload.
Conventional machine learning
Include it among the candidates for prediction or classification. Test accuracy, response time, and the explanations your users or reviewers need.
Fine-tuning
Identify the gap it would address and whether suitable training and evaluation examples are available. Compare the result and maintenance effort with other ways of addressing that gap.

Keep the model behind an interface your product owns. Changing it can still require updates to prompts, tool calls, or data handling. Evaluate the replacement on your use cases before releasing it.

Plan the first release and its review

Agree who will use the first release, what they can do with it, and how problems reach your team. Define the conditions for expanding access, changing permissions, or stopping the feature.

  1. Choose an initial group

    Select users and tasks that let you assess the feature with manageable review and support. Decide the size and rollout criteria for your context.

  2. Show the source, allow the edit

    Make sources available for checking where answers depend on retrieved information. Let users review and edit drafts, and provide a way to report incorrect results.

  3. Agree approval rules by action

    Define which actions need approval and who reviews them. Changes depend on evaluation results, the consequences of an error, and your team’s decision. Fewer corrections alone do not justify removing review. Our AI Agents & Assistants service includes defining these controls.

  4. Review usefulness and quality

    Measure task completion, accepted drafts, corrections, and continued use where relevant. Read those results alongside user feedback and the consequences of mistakes.

  5. Plan monitoring and incident review

    Define the records needed to investigate outputs and actions, who receives alerts, and when to pause the feature. Agree access and retention limits for those records.

  6. Describe the task clearly

    Use a label that explains the action and what the user should review. Illustrative labels include “Prepare a reply” and “Find answers in these documents.”

How long, and what it costs

We estimate the first feature after reviewing the workflow, data, integrations, and evaluation needs. Include approval controls, monitoring, and operating responsibilities in that scope. With a defined use case, you can start directly with implementation.

If the scope still needs work, a two-week AI Discovery Sprint has a fixed price and produces prioritized opportunities, a feasibility assessment, and a roadmap with the scope and acceptance criteria for a first delivery. Your team or another provider can use the roadmap. The sprint fee is credited 100% if you contract implementation with us within 90 days.

Check these gaps before release

  • Unclear task.

    Name the user, the task, and the result the feature is meant to help them achieve.

  • Narrow test data.

    Include representative inputs, exceptions, and failures when evaluating the feature.

  • Untested permissions.

    Check permitted and denied requests for each relevant user or organization.

  • No comparison with the current process.

    Record how the task works today and agree what would count as an improvement.

  • Unclear review responsibilities.

    Identify who checks results, handles exceptions, and can stop an action.

  • Unplanned operation.

    Agree monitoring, support, spending limits, and how changes will be evaluated after launch.

Frequently Asked Questions

Do we need our own model?
That depends on the task and your constraints. Compare available models and simpler approaches before committing to training or hosting one. The decision should account for evaluation results, data requirements, operating cost, and the team available to maintain it.
Will our customers’ data train someone else’s model?
Check the actual data flow and the terms for each service and endpoint. Document what is sent, where it is processed, stored, and logged, how long it is retained, and whether training use is permitted. Verify the applicable agreement and configuration for hosted services and any external services used with your own infrastructure.
What about hallucinations?
A model can produce incorrect answers even when it uses your data. Test those failures, let users check sources, and define when a response needs review or should be withheld. These controls do not eliminate the risk. Our Responsible AI service covers evaluation, permissions, review, and incident procedures.
Can our own team build it?
Yes, if it has the capacity and skills for the agreed scope. The Discovery roadmap includes findings, assumptions, integration needs, and acceptance criteria that your team or another provider can use. Implementation with Sophilabs is optional.
How do we know it worked?
Compare the results with the task as it was done before. Review the agreed measures alongside quality checks, user feedback, operating cost, and incident reports. Use that evidence to decide whether to expand, change, or stop the feature.

For a historical delivery example, see the Amgen voice assistant case. We delivered the assistant before the LLM era. For code and prompt ownership, data access and handover terms, see the Trust Center.

Plan an AI feature for your product.

A two-week, fixed-price Discovery Sprint gives you prioritized opportunities, feasibility findings, and a roadmap your team can use. If the scope is already defined, we can discuss implementation directly.