Choose an AI chatbot development company by checking whether it can demonstrate your required workflow, explain the full delivery scope, and accept measurable launch criteria. A persuasive demo is a starting point. Your purchasing decision should also cover what happens when information is missing, a connected system fails, or a customer needs a person.
If you are comparing proposals, ask each company to respond to the same brief. This gives you a basis for deciding whether a higher quote includes necessary work or a lower quote leaves that work to your team.
Use the checklist below to narrow your shortlist, evaluate a pilot, and agree on what you are buying before signing.
Define what you need before shortlisting companies
Write a short brief around the first workflow you want to improve. “Build a customer support chatbot” leaves too much open. A more useful brief identifies the questions it should answer, the systems it must access, and the actions it must not take.
For example, a proposed order-support workflow might let an authenticated customer check delivery status and read an approved returns policy. Refund approval could remain with a support agent. This is an illustrative scope, not a Lucent project claim.
Include the following in your brief:
- Describe the primary users and where they will access the chatbot.
- Identify approved knowledge sources and who maintains them.
- List required connections, such as a CRM, help desk, or order-management system.
- Separate information retrieval from actions that change records or move money.
- Explain when the conversation must transfer to a person.
- State your budget constraints, target launch window, and internal approver.
If a standard platform already covers the workflow, ask whether configuration or integration work would be sufficient. Our comparison of custom chatbot development and pre-built solutions addresses that earlier decision.
Lucent Innovation provides chatbot development services and has a commercial interest in this topic. Apply the same evidence requirements to Lucent as you would to another shortlisted company.
Compare AI chatbot development companies on these criteria
Use this as a proposal-review checklist. It is a recommended buying framework, not an industry certification or a guarantee of project success.
| Evaluation area | Evidence to request | Reason to pause |
|---|---|---|
| Relevant delivery experience | A comparable workflow walkthrough and the supplier’s specific role | Only screenshots or unrelated projects are available |
| Answer quality | Tests using approved content, grading rules, and failure examples | Accuracy is promised without a defined test |
| Integrations | A scoped system list, permission design, and sandbox demonstration | “Integration included” has no defined boundary |
| Human handoff | A transfer showing transcript, context, and fallback handling | Escalation ends at a contact form without an agreed process |
| Security and data handling | Data-flow documentation, access tests, retention details, and third-party dependencies | Responsibilities remain unclear |
| Commercial scope | Itemized deliverables, assumptions, exclusions, and recurring charges | Material work is left to unspecified later phases |
| Ownership and support | Deliverable inventory, support terms, and an exit process | Access and handover depend on unwritten promises |
1. Check experience with a comparable workflow
Ask a company to walk through a project with similar operational requirements. A retail chatbot and an internal knowledge assistant may share technology, but your evaluation should focus on the work relevant to your use case.
Establish what the company delivered itself. Did its team build the integrations, prepare the knowledge sources, test the bot, and support the launch? Or did it contribute only one component?
Request a reference where client permission allows. If confidentiality prevents a named example, ask for an anonymized explanation of the constraints, responsibilities, and acceptance process. Do not treat a lack of public client names alone as a reason to reject the company.
Ask about: “Show us a comparable workflow and explain one problem your team had to resolve before release.”
2. Require a defined answer-quality test
Ask vendors to evaluate the chatbot using a permitted sample of your content and representative customer questions. Agree on what makes an answer acceptable: correctness, support from approved sources, completeness, and appropriate escalation.
OpenAI’s evaluation guidance recommends task-specific tests, representative inputs, human judgment, and continued evaluation as a system changes. These principles are useful for assessing a delivery partner regardless of the model it proposes.
Request unsuccessful examples as well as successful answers. A vendor should explain what the bot does when a policy is missing or conflicting, and how content updates will be tested.
Ask about: “What test results will we receive, and who approves the chatbot’s answers before launch?”
3. Verify integrations and action boundaries
Replace “connects to your CRM” with a written description of what the chatbot can read, create, update, or delete. Include the systems involved, credentials required, and any work your internal team must provide.
Ask for a sandbox demonstration of both completion and failure. For an appointment workflow, the chatbot should distinguish a confirmed booking from an unsuccessful request. Require the proposal to explain how retries, duplicate requests, and unavailable services will be handled.
Keep permissions in the connected systems. OWASP’s excessive-agency guidance recommends limiting tool capabilities and permissions, enforcing authorization downstream, and requiring approval for high-impact actions. A model’s response should not decide whether a customer is entitled to access a record.
Ask about: “Which actions can the chatbot perform, and where are permissions enforced?”
4. Test human handoff as part of the workflow
Specify the conditions for transferring a conversation, including an explicit request for a person, an unresolved question, or an action outside the bot’s permissions.
During the demonstration, check what the agent receives. Request the transcript, relevant customer context, and a clear account of what the chatbot has already attempted. Also agree on what the customer sees outside support hours or when no agent is available.
Ask about: “Show us an escalation from the customer’s request through to the agent receiving the conversation.”
5. Review data handling with the right internal owner
Ask for a data-flow diagram showing what information reaches the application, model provider, connected systems, and logs. Request written details about retention, deletion, hosting locations, subcontractors, and access to conversation records.
Have your security or privacy owner review those arrangements against your requirements. If a supplier cites a certification or report, ask which organization, service, and scope it covers. Do not accept a general “compliant” statement as approval for your particular deployment.
Ask about: “Which parties process our data, and what evidence can our security team review?”
6. Confirm who will deliver and maintain the chatbot
Ask who owns architecture, integration work, testing, and support. Establish which responsibilities sit with the supplier and which require your product, support, or IT team.
Request a plan with dependencies and approval points. Before accepting a launch date, check whether it assumes ready-to-use content, available APIs, approved credentials, and timely feedback from your team.
Ask about: “What must we provide, who is accountable for each milestone, and what happens if a dependency is delayed?”
Compare proposals on scope and ongoing costs
Ask shortlisted vendors to price the same initial workflow and explain every material assumption. Avoid comparing a demonstration budget with a proposal that includes production deployment and support.
| Commercial item | What to clarify before signing |
|---|---|
| Discovery and design | Deliverables, decisions covered, and whether this work can be used if you choose another supplier |
| Knowledge preparation | Included sources, cleanup responsibilities, permissions, and update process |
| Integrations | Named systems, permitted actions, API dependencies, and excluded work |
| Testing and launch | Evaluation evidence, security checks, deployment responsibilities, and acceptance criteria |
| Recurring usage | Model and platform charges, hosting, channel fees where applicable, monitoring, and billing assumptions |
| Support and changes | Coverage hours, incident handling, maintenance, and how new requirements are priced |
| Exit and handover | Transferable deliverables, account access, documentation, export support, and applicable fees |
Request a recurring-cost estimate with explicit assumptions about usage and conversation length. Ask which charges you pay directly and which the supplier invoices. Also request the conditions that would change the estimate, rather than relying on an unexplained monthly figure.
For ownership, identify each deliverable separately: application code, configurations, prompts, evaluation datasets, deployment instructions, and knowledge exports. Distinguish transferable project work from third-party services and licensed components. Have the relevant commercial and legal owners review the written terms.
Use a pilot to decide whether to proceed
Before commissioning a wider build, propose a limited pilot around one valuable workflow. Agree on the permitted data, test environment, evidence required, and decision process before work starts.
The following are illustrative acceptance scenarios. Adapt them to your use case; they are not measured Lucent results or universal release criteria.
| Test scenario | Proposed acceptance behavior |
|---|---|
| An answer exists in an approved policy | The response matches the relevant policy and identifies its source where appropriate |
| The requested information is missing | The chatbot explains the limit and follows the agreed escalation process |
| A customer requests another person’s order | Authorization prevents unauthorized disclosure |
| A connected booking service fails | The chatbot reports failure and does not claim a confirmed booking |
| A customer requests a support agent | The agreed transfer preserves relevant conversation context |
| An approved policy changes | The updated answer passes review and the obsolete version is no longer used for that response |
Set acceptance criteria by task and consequence. An incorrect opening-hours answer and an unauthorized refund should not be treated as equivalent failures.
Our recommendation is to make unauthorized disclosure and unapproved high-impact actions separate release gates. If a pilot fails one of those gates, require a fix and a retest before approving deployment, even if other answers perform well.
How to make the final vendor decision
Compare documented evidence after the pilot, not only the initial sales presentation. Mark each criterion as verified, unresolved, or not met, and record who reviewed it.
| Decision criterion | Your requirement | Vendor evidence | Status |
|---|---|---|---|
| Priority workflow | Define the outcome to demonstrate | Record the walkthrough or pilot result | Verified / unresolved / not met |
| Answer quality | Agree on task-specific acceptance criteria | Record test results and reviewer approval | Verified / unresolved / not met |
| Permissions | Define allowed users, records, and actions | Record access-control test results | Verified / unresolved / not met |
| Commercial scope | Agree on inclusions and recurring assumptions | Reference the final proposal | Verified / unresolved / not met |
| Operational ownership | Name support and content owners | Reference the support and handover plan | Verified / unresolved / not met |
Shortlist a company when its evidence matches your requirements and the remaining risks have an owner. Pause when a critical requirement remains unresolved, regardless of presentation quality or price.
You may also decide not to commission custom development. If an existing product meets the workflow, your knowledge sources are not approved, or nobody can own escalations after launch, consider a smaller implementation or resolve those dependencies first.
Discuss your chatbot project with Lucent Innovation
Comparing development partners for a customer-support, booking, or internal-assistance workflow? Explore Lucent Innovation’s AI chatbot development services and discuss the scope your team needs to evaluate. Bring an anonymized workflow summary, your planned integrations, and the constraints a proposal must address. Use anonymized information until an approved confidential-sharing process is in place.

