Decision matrix
Build or buy: weigh it for your support and sales chat
The defaults weight every criterion equally, which on this matrix is a close call. Change the weights to match your situation and watch where they lean.
1 · Set what matters to you
2 · Where your weights lean
- Helpdesk AI add-onGood fit
- Standalone chatbot SaaSGood fit
- Custom build on your accountsGood fit
Close call between Standalone chatbot SaaS and Custom build on your accounts. When the matrix is this close, the discovery questions below decide it, not the scores.
3 · The matrix, cell by cell
| Criterion · your weight | Helpdesk AI add-on | Standalone chatbot SaaS | Custom build on your accounts |
|---|---|---|---|
| Live quicklyWeight: Medium | Strong fit Often switched on inside the helpdesk you already run, once content is connected. | Strong fit Days to set up: connect a site or documents, style the widget, embed it. | Partial fit Weeks: scoping, ingestion, evaluation and channel work come before launch. |
| Low upfront costWeight: Medium | Strong fit Usually a plan upgrade or a usage fee, not a project. | Strong fit A monthly subscription by tier. | Partial fit A project fee for the build, paid before anything runs. |
| Questions are mostly standardWeight: Medium | Strong fit Built for exactly this: answering from your help-centre articles. | Strong fit Strong for website questions and lead capture. | Good fit Can do it well, but buying is often enough when answers are all it needs. |
| Must act in your own systemsWeight: Medium | Partial fit Limited to the actions and integrations the vendor ships. | Partial fit Actions through webhooks or prebuilt connectors, within the plan’s limits. | Strong fit Any action your systems expose through an API, behind your own guardrails. |
| Conversations stay on your accountsWeight: Medium | Partial fit Conversations and settings live in the vendor’s cloud under its retention rules. | Partial fit Content and transcripts sit with the vendor; check region and retention options. | Strong fit Runs on accounts in your name; you set storage, region and retention. |
| Control over how answers are foundWeight: Medium | Partial fit Retrieval and answer logic belong to the vendor; you tune content and a few settings. | Good fit More settings than an add-on, but retrieval and models are still the vendor’s choice. | Strong fit You choose retrieval, models, citations and the rules for handing over to a person. |
| Little upkeep for your teamWeight: Medium | Strong fit The vendor runs and updates it; you maintain the articles. | Strong fit Vendor-run; your team keeps content current and reads transcripts. | Partial fit Someone must monitor, update and re-test it: your team or a support plan. |
| Cost stays predictable as volume growsWeight: Medium | Partial fit Often billed per conversation or per resolution, so cost rises with success. | Good fit Tiered plans are easy to forecast until you outgrow a tier. | Good fit Model and hosting usage are paid directly; forecasting needs a usage estimate. |
| Cheap to leave laterWeight: Medium | Partial fit History and configuration rarely export in a form another tool can use. | Partial fit Moving means rebuilding flows and retraining staff on a new tool. | Strong fit You hold the code, prompts and data, so you can change builders or bring it in-house. |
Ratings describe typical products in each group, not any named product, and they are our judgement. Use them to structure a shortlist, then test the shortlist on your own questions.
Three ways to get an AI chatbot
A helpdesk AI add-on is the assistant your existing support tool offers as an upgrade. It reads your help-centre articles, answers in the chat or email channels the helpdesk already runs, and passes anything it cannot handle to your agents in the same inbox.
A standalone chatbot SaaS is a dedicated product: you point it at your website or documents, style a widget and embed it. It usually adds lead capture, simple flows and connectors to common tools, and it is configured through the vendor’s dashboard.
A custom build is a system made for your workflow on model APIs and infrastructure in your name. It can be built by your own developers or by a studio like ours, and it does whatever your systems allow through their APIs. You pay for the build and for the usage, and someone has to look after it.
The criteria that usually decide it
Whether it must act or only answer
This is the biggest single divider. If the chatbot only answers questions from published content, buying is hard to beat. Once it must look up an order, change a booking, issue a credit or update a CRM record, you need actions with permission checks, and bought products only offer the actions their vendor has built. A custom build can call any API you expose, which also means it must be designed so the model never decides on its own whether it is allowed to.
Where the conversations live
Chat transcripts contain names, order details and sometimes health or financial information. In a bought product they sit in the vendor’s cloud under its retention and access rules, which may be fine, but should be read before signing. A build on your own accounts lets you choose the region, the retention period and who can read transcripts, and your security team can review all of it.
What it does when it is unsure
Ask every option the same three questions: does it show the source of each answer, does it decline or hand over when it cannot find one, and what stops instructions hidden in a document or web page from steering it? OWASP’s list of risks for language-model applications opens with prompt injection and says it is unclear whether fool-proof prevention exists, so the useful answer is a layered one, not a promise.
Who keeps it working
Bought products are maintained by the vendor; your team keeps the content current and reads transcripts for gaps. A custom build also needs monitoring, re-testing after model updates and small fixes as your systems change. If nobody can own that, either buy, or budget for a support plan from whoever built it.
What it costs as you grow
Bought products bill per seat, per conversation, per resolution or by tier, and the unit matters more than the headline price. Pricing per resolution means your bill rises exactly when the chatbot works well. A custom build pays model and hosting usage directly at the vendor’s rates, which tracks volume closely but needs an estimate before launch.
What leaving would cost
Ask before you buy what you can take with you: transcripts, configuration, flows and the knowledge base. If the answer is a CSV of conversations, a later move means rebuilding everything else. With a build on your accounts you keep the code, prompts and data, so changing builders or bringing the work in-house is a handover, not a rebuild.
Total cost over time
| Cost line | Helpdesk AI add-on | Standalone chatbot SaaS | Custom build |
|---|---|---|---|
| Setup | Connecting content, internal time | Configuration, widget, internal time | Project fee for scoping, build and evaluation |
| Recurring fees | Plan upgrade or per-unit usage | Subscription tier | Optional support plan |
| Usage | Included or per conversation or resolution | Within tier limits, then overage or upgrade | Model, hosting and channel usage at the vendors’ rates |
| Your team’s time | Content upkeep, transcript review | Content upkeep, transcript review | Content upkeep, approving changes |
| Changes | Limited to the vendor’s settings | Within the product’s features | Quoted per change, or covered by a plan |
| Leaving | Rebuild elsewhere | Rebuild elsewhere | Handover; the system stays yours |
The hybrid most teams overlook
You do not have to choose one layer for everything. A common pattern is to keep the helpdesk your agents already use as the inbox and system of record, and build only the AI layer behind it: retrieval over your content, the actions in your own systems, and a handover that opens a ticket with the full conversation attached. Agents keep their tools, customers get answers that can act, and the custom part stays small.
The questions we ask in discovery
- What share of last month’s conversations could be answered from published content alone, and what share needed someone to do something?
- Which systems would the chatbot need to read from, and which would it need to change?
- Which helpdesk or CRM do your agents live in today, and does it already offer an AI add-on?
- What personal data appears in a typical conversation, and who has to approve where it is stored?
- Who on your side will read transcripts and keep content current once it is live?
- How do you expect conversation volume to change over the next year?
- If you stopped using the tool, what would you need to take with you?
How we decide
- Pull a sample of real conversations and tag each one: answer from content, needs an action, needs a person.
- If answers from content dominate and your helpdesk has an add-on, trial that first. We say so even though it means no build for us.
- If actions or data custody matter, scope the smallest custom piece that delivers them, often the hybrid above.
- Price both paths over the period you expect to run the system, including your team’s time and the cost of leaving.
- Put the comparison in writing with the proposal, so you can hold any vendor, including us, to it.
When MiniAI Labs is the wrong choice
If your questions are standard help-centre answers and your helpdesk already offers an AI add-on, use it. A custom build would cost more for the same result, and we will tell you that on the scoping call.
If you want a self-serve product with a trial and a dashboard your team configures alone, buy a SaaS tool. We build and hand over systems; we do not sell a platform or a subscription seat.
If your procurement rules require every vendor to hold SOC 2 or ISO 27001 certification with no exceptions, a certified SaaS vendor is the practical route. We are not certified; we build security-first on your accounts and answer security questionnaires in writing.
Limits of this comparison
- Vendor features, plans and data terms change quickly; check each vendor’s current documentation before deciding.
- Ratings describe typical products in each group, not any named product.
- This page covers text chat. Voice agents add telephony, latency and call-consent questions of their own.
- Neutral default weights are a starting point; your own situation sets the real ones.
Sources
- LLM01:2025 Prompt Injection (opens in a new tab) · OWASP Gen AI Security Project · accessed
- LLM06:2025 Excessive Agency (opens in a new tab) · OWASP Gen AI Security Project · accessed