Buying AI for a business is not a choice between a subscription and building everything from scratch. An existing feature may already cover the job. A configurable platform may need data connections and approval steps. A custom product may be justified when important requirements remain unmet.
Before requesting proposals, define the work that must be completed and what an acceptable result looks like. Otherwise, you may compare a licence for one tool with delivery of an entire workflow, without accounting for the different outcomes and responsibilities.
Talvyn Labs assesses and develops AI Products and Agents around existing business systems. Where an available product fits, the work may centre on integration and review workflows. Custom development should address a demonstrated gap.
Compare three delivery approaches
| Approach | When it can fit | What still needs checking |
|---|---|---|
| An existing AI feature | The job already happens in software whose features cover the main workflow | Quality on real inputs, permissions, plan limits and how results are used |
| A configured platform with integrations | Available components fit, but data, workflow and approvals need configuration | Connector capabilities, maintenance when systems change and operating responsibilities |
| A custom AI Product or Agent | Important requirements remain unsupported or create substantial friction | Delivery scope, acceptance tests, ongoing costs and maintainable handover |
Microsoft Copilot Studio, for example, provides tools for building agents and connecting data or services. A platform still requires decisions about permissions, workflow and evaluation. Available capabilities do not establish that a particular business process will work without configuration.
Custom development does not necessarily mean training a model. A team can use existing models and services while building the data connections, interface and application behaviour the business requires. Evaluate the usable result and the ability to maintain it.
Give each option the same job
Consider an illustrative distributor that wants to prepare quotations from customer messages and documents. Give each shortlisted provider the same authorised requests, catalogue sample and definition of an acceptable draft. Ask to see the journey from intake to staff review.
One tool may extract the requested items but lack access to customer-specific prices. Another may identify product codes without showing the source message behind each choice. Those differences affect review time even when both tools produce polished text.
Include incomplete inputs: similar product names, different selling units and unreadable attachments. A useful system identifies unresolved information and leaves a decision for a person. Filling every field can conceal a failure.
For this workflow, the guide to preparing quotations from LINE and PDFs separates interpretation from verification of commercial terms. Use that distinction when deciding what a proposed solution actually covers.
Questions that reveal whether the solution will operate well
Where does the information come from, and who can access it? Identify sources, update methods and permissions for real users. If branches or customer accounts have different access, test that an unauthorised person cannot retrieve another group's information.
What happens when AI is uncertain? Look for original sources, explicit missing information and a workable handoff to staff. Producing an answer every time does not establish correctness.
Who approves a change? Drafting a message and sending it to a customer have different consequences. The scope must distinguish reading, preparing and executing, with a record of approval where required.
Who maintains connections when another system changes? Ask who owns connectors, accommodates new data formats and investigates failures. Integration work continues after the first successful demonstration.
What can the business take away if it switches providers? Check export rights for data, configuration, history and documentation under the proposed agreement. A subscription does not imply ownership of every component. A custom project does not mean every underlying service can be installed independently.
Answers should be demonstrated where practical and recorded in the proposal. A verbal assurance that a feature “is possible” leaves its cost, constraints and delivery responsibility unresolved.
Compare the cost of running the workflow
Include licences or platform charges, usage charges, data preparation, integration, staff review and ongoing support. Terms depend on the provider and scope, so compare current proposals rather than treating a model's request price as the full cost.
Where usage varies, ask how consumption is monitored and bounded, and what happens when a limit is reached. Retries, failed requests and correction work can matter as much as the successful response. A low unit price is less informative when staff must rebuild much of the output.
Do not turn assumed time savings into guaranteed returns. Trial representative work and measure the full journey to an accepted result, including review and correction. The guide to evaluating and accepting an AI pilot explains how to establish a baseline and record failures.
Write down the reason for custom development
Potential reasons include detailed access rules, conditional approvals, specialised product information or a need to work inside the user's existing screen. First establish that the evaluated options cannot meet the requirement adequately and that the gap matters to the business.
An AI feature within a product used by customers is a different scope from a staff tool. Search or summarisation inside a Customer Portal may depend heavily on interface design, account boundaries and application permissions. See adding AI to existing software when those are central requirements.
It can also make sense to combine approaches: keep an existing system of record, use a suitable platform and develop the review interface or connection that is missing. Ask which parts are configured, which are custom and who will maintain each one.
Prepare for a useful scope discussion
Bring Talvyn Labs a description of the current workflow, shareable input examples, expected outputs, relevant systems and the people authorised to approve changes. Include constraints that must stay in place, such as an accounting system or a team review process.
The outcome should be a reasoned choice, a bounded trial and criteria for proceeding. If a critical source or permission model is missing, resolve that gap before committing to a larger implementation.
Choose an AI approach before committing to development
Bring your workflow and existing systems to Talvyn Labs to assess what available tools cover, what needs integration and what merits custom development.
Assess your AI delivery options