Use cases by industry15 use cases

AI for B2B SaaS teams, inside the product and behind it

Take a vertical SaaS with a REST API, a couple of thousand customer accounts and a roadmap that already eats every engineer. Customers now ask for two things the docs chatbot cannot give them: features that do the work for them, and a way to use the product from Claude, ChatGPT or Copilot. Many SaaS companies should build this themselves, and I say so when they have the people. This hub is for the ones whose engineers are fully booked and who have not shipped model-backed features before.

Inside the product, the work is an in-app assistant that acts through your own API with the logged-in user's permissions, and an MCP server whose tools are designed for agents rather than mirroring every endpoint. Upwork's 2026 skills report shows where US buyers moved: demand for AI integration work grew 178% in 2025, against 71% for chatbot development.

Behind the product, the leaks are familiar. Senior engineers get pulled into support rotations to read logs, which a second-line support agent does first. Model bills grow faster than usage, and KPMG found that only 26% of large US companies have full, real-time visibility into what their AI costs to run; an AI cost audit is usually the cheapest fix.

Book a 30-min scoping callAI Integration →

What I would build for saas

What makes an AI feature safe to ship to customers

  • The user's permissions, never a service account

    Every tool call runs as the logged-in user through your API, and retrieval is filtered by tenant. A release test that tries to read another tenant's record runs before anything deploys.

  • Writes get a preview and an undo

    Bulk edits, deletions and workflow changes show what will change, wait for confirmation and stay reversible for a window. Destructive tools sit behind a feature flag until the logs show they are used well.

  • Metering that follows the pricing page

    Usage is metered per account through Stripe and capped per plan, so a handful of heavy users cannot quietly turn the AI feature into a margin problem.

  • Customer content is untrusted input

    A ticket body, a file name or a CRM note can carry instructions aimed at the model. Tool results are treated as data and writes need confirmation; the patterns are in the prompt injection guide.

  • Evaluations on every model change

    A set of real tasks with known good results runs on each prompt edit and each model upgrade, because providers retire models on their schedule, not yours.

Where a SaaS company should buy before it builds

  • First-line support

    Intercom Fin at $0.99 per outcome, Zendesk's AI agents and Decagon, which raised $250M at a $4.5B valuation in January 2026, answer documented questions well. Custom starts at second line, where the answer needs logs, account configuration and the issue tracker.

  • Security questionnaires

    Vanta, Drata, Conveyor, SafeBase and Loopio draft answers from SOC 2 and ISO 27001 evidence. DORA and NIS2 have multiplied these questionnaires for anyone selling to banks and critical sectors; build only when answers depend on each customer's deployment.

  • Call notes into the CRM

    Gong, HubSpot Breeze and Salesforce Einstein already do this. A CRM update agent is worth building when the field logic is your own and every proposed change must quote the call it came from.

  • Generic CSV imports

    Flatfile handles ordinary imports. A custom migration pipeline pays only for vertical products with complex objects, where your implementation team currently maps and validates each customer's legacy export by hand.

Software these builds usually connect to

The systems do not get replaced. The build sits across them, reads from them through their APIs, and writes results back.

Intercom / Zendesk / Jira / Linear / Sentry / Datadog / HubSpot / Salesforce / Stripe / Chargebee / Snowflake / Notion

Frequently asked questions

How do I add AI features to an existing SaaS product?

Start with one job users repeat and would happily hand over, such as building a report or bulk-editing records, not a general chat box. Expose it as a small set of tools over your existing API, running with the user's own permissions, with confirmation before writes, metering per plan and an evaluation set of real tasks. Once one feature holds up in production, the same tool layer carries the next one.

Does our SaaS product need an MCP server?

If your customers already work inside Claude, ChatGPT or Copilot, probably yes. HubSpot, Notion, Linear, Asana and monday.com ship official servers, and Claude's directory lists hundreds of connectors. A server generated from your OpenAPI spec is a fine first version; tools designed around tasks, with OAuth scopes, rate limits and per-tenant audit logs, are what agents actually use well.

Why is our AI bill growing faster than our user count?

Usually several causes stack up: prompts that grew with every fix, agent loops without a depth cap, no prompt caching, a large model doing classification a small one could do, and retrieval sending far more text than the answer needs. An audit measures cost per feature and per customer, then cuts with an evaluation set running, so quality is checked while the bill comes down.

We have engineers. Why hire outside help for AI?

Often you should not. If your team has shipped model-backed features with evaluations, tracing and cost controls, keep the work in-house. Outside help makes sense when nobody has done that before, when the roadmap cannot spare two engineers for a quarter, or when a vibe-coded prototype needs production hardening before real users arrive. A fixed-scope build with a clean handover leaves the code with your team.

Also browse: AI in your product · Customer support · Zendesk and helpdesks · HubSpot · all use cases