# Bringing AI into SaaS - A Case Study

Written by Bart Van Hecke · Imported 2026-10-20

**Type:** article

Every SaaS product is being asked the same question right now: where does AI actually belong? Not as a badge on the pricing page, not as a chatbot bolted onto the corner of the screen — but as something that genuinely changes how the product works for the people using it every day.  This is the story of how we answered that question for NxtPeople, an HR platform used by recruitment consultants to screen, match, schedule, and pay candidates at volume. It's not a highlight reel. It's the actual process: the audit that told us where AI would help and where it would just add noise, the one feature we chose to ship first and why, the architecture decisions behind it, and the organizational shifts we're still working through. If you're a product or design lead asking the same "where does AI belong" question about your own platform, this is what that process looks like from the inside.

## The pattern of automation.

As a seasoned Experience Designer, I’ve shaped a broad spectrum of enterprise and SaaS products, always guided by the same ambition: to create experiences that feel effortless, intuitive, and quietly powerful. Like most things in life, UX is not an exact science. It lives on **patterns** — proven, repeatable solutions that dissolve friction, from autosuggestions to progressive enhancement. Today, the most compelling of these patterns is **AI-driven automation**.

While I explore my broader perspective on AI in [*The End of UX*](https://superflow.design/the-end-of-ux/), this article turns to the tangible and the practical: **how we brought AI into NxPeople**, the HR SaaS platform where I work as Product Experience Designer.

[NxPeop](https://www.nxtpeople.com/en)[le](https://www.nxtpeople.com/en) helps recruiters screen, match, schedule, and pay large volumes of candidates — work that, in many organizations, remains surprisingly manual despite years of automation promises. What began as a powerful tool for interim consultants evolved into a feature-rich platform, effective yet sufficiently complex. Now, we’re entering a new chapter: transforming NxPeople into an intuitive, self-service experience, while extending our reach to smaller organizations through [**GoTemp**](https://www.gotemp.be/nl) — a streamlined spin-off built for flexible hiring.

This evolution demands more than new features. It requires rethinking workflows from the ground up, stripping away friction, and weaving in AI only where it truly adds value — not where it merely looks impressive. Inspired by products like Slack and Notion — calm on the surface, intelligent beneath — we’re reshaping recruitment technology for real people, not just power users.

The goal is simple: to build a platform that feels *considered*, *human*, and unmistakably *for its users* — whether they manage thousands of candidates or just a handful.

![](../../assets/images/articles/62510ff38789b2fe.webp)

## Why AI?

### Changing strategies

Although we’ve been exploring AI for some time — from candidate–job matching to workforce scheduling predictions — much of that work lived quietly in prototypes, tucked away in our sandbox. As a small, focused team, our energy naturally went to the essentials. AI was a *nice to have*: often experimental, often exploratory, and frequently entrusted to interns or students building proof-of-concept models.

That moment has passed. The market is shifting, and with it, our clarity. We now see where AI can deliver genuine, measurable value — not as a novelty, but as a force that meaningfully improves the product.

Crucially, we came to understand that AI cannot exist as a surface layer. Over time, it must **become part of the system’s spine** — shaping how data flows, how insights surface, and how decisions are supported. Doing this well demanded a strategy that embeds AI into the product’s foundation, rather than grafting it on as an afterthought.

At the same time, our user base remains firmly rooted in **recruitment consultants in Flanders**. Expanding toward a **self-service model** for end clients introduces a delicate new tension. The question is one of balance: how do we unlock AI-driven efficiency without disturbing the trusted workflows our consultants rely on every day?

## Where AI?

### Audit

We began by mapping the entire workflow end-to-end, looking closely at where friction accumulates and momentum slows. The same pressure points surfaced again and again: repetitive data entry, pattern recognition at scale, scheduling complexity, and the absence of predictive insight. Each of these moments signaled a clear opportunity for automation.

From there, we moved into prototyping. One of the earliest concepts was an AI-driven matching demo that combined the reliability of structured search with the nuance of contextual intelligence. Job requirements were mapped against parsed candidate profiles — stored as JSON and optimized for fast Elasticsearch queries — while an AI layer enriched the results with meaningful context: relevant projects, inferred skills, and even candidate motivations.

An optional LLM-powered rescore step then refined the shortlist further, applying a semantic match score that was both explainable and transparent. Recruiters didn’t just receive a ranked list; they could see *why* candidates surfaced, grounding trust in clarity rather than mystery.

Still, turning promising concepts into production features takes time — and attempting to ship everything at once would only create noise. Our users in the interim sector have developed deeply ingrained, highly manual workflows over years of practice. Any AI-driven improvement must therefore meet them where they are: respecting the habits that keep their work moving today, while gently opening the door to what becomes possible next.

### Our decision: AI Job Description

We chose to begin with something deceptively simple: helping users write better job descriptions. After all, a vacancy text is often the very first handshake between a candidate and a company — and yet it’s where inconsistency, vague language, and unintentional bias most frequently creep in. The result is familiar: mismatches, misunderstandings, and time lost on both sides.

Here, AI proves its value quickly and quietly. Through a combination of prompt chains, role-aware prompting, and structured templates, it generates drafts that align with industry standards, surface the right keywords, and strike a more welcoming tone. The flow remains intentionally simple: users provide their input, AI proposes a draft, and human judgment stays firmly in control — reviewing, refining, and shaping the final result.

To anchor the feature in real-world practice, we spoke with the people who write vacancies every day — those translating client needs into clear expectations. We learned how they calibrate tone, manage nuance, and set boundaries without closing doors. At its heart, a job description is an invitation. AI doesn’t write that invitation *for* our users — it helps them make it clearer, more consistent, and ultimately, more compelling.

## How AI?

### Designing for Adoption

![](../../assets/images/articles/99f7df092979d5ef.webp)

Once we identified a feature where AI could truly add value, the next question became **how** to integrate it in a way that felt natural, helpful, and trustworthy. In short we wanted to answer the question: which design choices work best for my user? Should we introduce a friendly, AI-powered assistant – something like KBC’s “Kate,” but with fewer weather updates – to guide users step by step? This would allow people to keep the familiar manual workflow as a fallback, adopting AI at their own pace. Or should we take the bold route and weave AI directly into the core interface, offering a streamlined, modern experience from day one and phasing out the old way entirely?

Best practices across the tech world show that a gentle transition is often the most user-friendly one – Slack didn’t force new workflows on day one, and Notion’s onboarding remains a masterclass in gradual learning. Giving users the choice between manual actions and AI-assisted flows respects their existing habits while nudging them toward smarter, more efficient ways of working. The real secret, though, is *continuous feedback*. By listening closely to both consultants and end clients throughout this transformation, we can shape features that empower rather than overwhelm.

Further looking at industry leaders – Salesforce Einstein, HubSpot, ClickUp Brain, Google Gemini, Notion AI, Zendesk, Grammarly – we noticed a common pattern: avoid bolting AI onto existing features. Instead, weave intelligence *into the workflow itself*, so it feels like part of the journey rather than a separate tool. This aligns perfectly with a **progressive disclosure** strategy: start with low-risk, high-value enhancements like autocomplete or smart recommendations, then gradually move toward predictive and automated functionality. Offering tiered control – from manual, to semi-automated, to fully automated – helps our users build confidence without feeling pressured.

![](../../assets/images/articles/06daafdf0880716d.webp)

Many modern SaaS platforms now offer an AI assistant that sits quietly at the user’s side, ready when needed. At **NxtPeople**, we share that philosophy. Since many of our users are still getting comfortable with AI, we want it to feel supportive, not intrusive. We studied how others approach this – HubSpot’s Breeze, Google Gemini, and especially Sider, a side-panel assistant that provides guidance without getting in the way. Inspired by these models, we explored a similar assistant for NxtPeople, one that could slide open into a split-screen view and help automate tasks like generating candidate profiles or drafting client emails.

But through conversations with stakeholders, we realized such an assistant only becomes valuable once enough workflows are automated in the background. If it only helps with minimal tasks at the beginning, it feels rather stupid as an assistant and as such, users loose trust, which is crucial if you want to get them acquainted with automation. So for now, we’re starting small and intentional. The first step is introducing a lightweight chat assistant to help users draft job descriptions – an impactful, low-risk entry point. From here, we’ll expand gradually, guided by user feedback and by a simple principle: AI should enhance the work, not complicate it.

![](../../assets/images/articles/0bff82c56c290ff2.webp)

### Choosing the right architecture

To determine how to bring AI into NxtPeople effectively, we then explored the broader landscape of available technologies. We compared highly capable generative AI platforms such as OpenAI’s ChatGPT and Anthropic’s Claude, with more traditional recruitment-oriented tools like [Spott.io](http://spott.io/), [Vincere](https://www.vincere.io/), or [Bullhorn](https://www.bullhorn.com/eu/products/amplify/). The former are powerful, flexible, and strong in contextual reasoning, while the latter offer ready-made, cost-efficient integrations that may lack the depth needed for our domain.

Our evaluation centered on a few critical questions:

– Is the tool available via API and suitable for multi-tenant environments?

– How does it handle authentication, localization, data retention, and GDPR compliance?

– Can it support trigger-based workflows or webhooks?

– What are the usage limits, costs, and scalability options?

– Does it offer audit logs, fallback behaviour, or human-in-the-loop safeguards?

– And above all: does this choice support our long-term vision for new regions, languages, and more advanced features?

We also looked at existing third-party generators like Workable’s or Grammarly’s AI Job Generator, but our developers preferred a framework that allowed deeper, more structured workflow integration.

![](../../assets/images/articles/0cb3528722baa9c6.webp)

### **From Exploration to Execution**

After selecting the foundational tools, the development team began mapping out the technical setup required to build an MVP. They reviewed API documentation, explored request and response structures, and assessed how to leverage Tiptap’s functionality efficiently. The goal: build a first version quickly so that real users—not assumptions—can guide our direction.

The rollout plan mirrors this philosophy. We’ll introduce the AI model to a selected group of users while others encounter it naturally. Using [PostHog](https://posthog.com/), we’ll track engagement, adoption, and meaningful value metrics – such as aiming for 40 users interacting with the AI assistant in the first week and generating at least 20 job descriptions.

We don’t yet know if this is the perfect testing strategy. What we do know is that progress requires action. By experimenting early, learning from real behaviour, and adjusting based on evidence, we’ll ensure AI becomes not just an add-on, but a natural, empowering part of NxtPeople.

### The Price

When selecting AI tools, understanding the underlying pricing model is just as important as evaluating functionality. Many platforms charge a **license fee per user** or an **API fee based on the number of tokens processed**, and costs can escalate quickly if usage isn’t monitored. Take ChatGPT, for example: the license fee is $25 per user per month up to 149 users, meaning 50 users would cost $1,500/month. In addition, API-based consumption charges $3 per million input tokens and $10 per million output tokens. For a typical job description of 300 words—roughly 420 input tokens and 900 output tokens—50 users generating one job description per day would result in 1,500 job texts per month, not including extra tasks like candidate generation that add to token usage. By actively **monitoring consumption**, teams can leverage flexible token-based pricing without unexpected cost overruns, making it an essential consideration when integrating AI at scale.

### **Our choice: Tiptap: A Foundation for Scalable, Human-First AI**

This led us to [**Tiptap.dev**](https://tiptap.dev/open-source-to-platform?gad_source=1&gad_campaignid=22014820935&gbraid=0AAAAAqkAF24GF_SVQqtR2aobxQwdg81yo&gclid=Cj0KCQiAhOfLBhCCARIsAJPiopN6WelPXnLb843caILL6yGGsozmtbrElm5taZmlJiKZknjiNTH77VcaAiyoEALw_wcB), a modular, headless rich-text editor framework. Unlike standard WYSIWYG editors, Tiptap provides full control over both **editor functionality and workflow integration**, while offering prebuilt components and templates to accelerate development. This modular, flexible architecture is ideal for embedding AI assistants and layering automation into workflows, from simple text assistance to advanced tasks like workforce scheduling and candidate management. By starting small and expanding in stages, we can integrate AI gradually without overwhelming users, while maintaining an intuitive, collaborative experience.

With Tiptap, we could start simple and expand gradually, layering AI-driven features over time without disrupting the user experience. It also aligns with a technical approach that processes most tasks on the frontend through an npm package, sending results to the backend only when necessary. This should significantly improve performance compared to our current AI setup, which triggers a backend validation request at every step.

Tiptap also supports future growth. Whereas general-purpose AI APIs would require additional “glue” to fit into structured workflows and collaborative document editors, Tiptap’s framework allows us to scale from basic AI text features to fully embedded, workflow-driven intelligence. This makes it a robust, long-term foundation for our SaaS platform: flexible, modular, and designed to evolve alongside the complex needs of recruitment consultants and end clients.

In short, while simpler tools like ChatGPT can handle basic AI text tasks, Tiptap gives us the **control, extensibility, and scalability** necessary to embed AI deeply into NxtPeople, aligning with our vision of a human-first, progressively intelligent recruitment platform.

![](../../assets/images/articles/ba8ef72f329c3683.webp)

When a user creates a job description with AI, the frontend sends a REST API request to our backend, which combines candidate and employer parameters and forwards them to the LLM (OpenAI). The generated result is then returned to the frontend.

## **Testing AI**

Testing with real users is essential. As we’re in the process of conducting usability sessions focused specifically on AI interactions, we observe how users respond to recommendations, whether they trust automation, and how they react when the AI makes mistakes. These insights reveal issues that theoretical planning alone cannot anticipate.

To ensure AI delivers real value, we’ll track **engagement, quality, and business impact metrics**, and use tools like PostHog to monitor adoption rates, how often users interact with AI features, and whether automation actually saves time compared to manual processes. This enables us to evaluate the **accuracy and relevance** of AI suggestions, measure overrides, and gather user feedback to identify areas for improvement. Finally, we needed to connect AI usage to business outcomes: measure improvements in conversion rates, candidate satisfaction, or response times, and calculate the correlation between AI engagement and customer lifetime value.

What we learned during usage tests before, is that we need to define **concrete use cases and measurable goals**—for example, “adding this feature reduces average time to first response by X%” or “this add-on functionality increases candidate satisfaction by Y%.” By combining rigorous testing, thoughtful fallbacks, and clear metrics, AI can enhance SaaS workflows reliably while keeping human users in control.

## **Building an AI-First Organization: Teams, Processes, and Governance**

Successfully weaving AI into a SaaS product isn’t something you *switch on*. It’s something you grow into. At NxtPeople, we’re very aware that this is where we need to go—even if we’re not fully there yet. What we *are* doing, though, is learning fast, adjusting daily, and building the muscles that will let us get there responsibly.

One of our biggest insights so far is that AI can’t live in a corner. It needs **multidisciplinary teams** around it. Not “AI over there,” but data, ML, product, and design sitting together, shaping features from the start. We’re moving in that direction, knowing it takes time to break habits and silos. In parallel, we’re discovering just how important **infrastructure and MLOps** really are—data pipelines, deployment, monitoring, and cost control aren’t glamorous, but without them AI stays fragile.

We’re also learning that **trust beats cleverness**. Users don’t just want smart outputs; they want to understand what’s happening. Clear labeling, simple explanations, and the ability to push back or opt out matter more than we initially thought. Building feedback loops and thinking seriously about ethics and governance isn’t overhead—it’s table stakes. We’re still figuring out what “good” looks like here, but the direction is clear.

AI is also forcing us to rethink **how work flows through the product and the organization**. It’s not just another feature; it changes sequencing, ownership, and expectations. Concepts like multi-agent systems are exciting, but we’re realistic: before running, we need to walk. Sometimes a solid search or rules-based solution is the right step *for now*, until data volume and usage justify something more advanced.

So yes—we know where we’re heading: toward AI-native teams, systems, and workflows. We’re not pretending we’ve arrived. But every iteration teaches us something, and every small, pragmatic step gets us closer. For us at NxtPeople, building with AI is less about a big reveal and more about **earning the right to go further, one learning at a time**.

## Conclusion

Integrating AI into SaaS isn’t about chasing hype. It’s about **designing systems that are intelligent, transparent, and resilient**—systems that empower users to do more while keeping them in control. From embedding AI expertise directly into product teams, to building modular, scalable infrastructure with tools like Tiptap, to designing graceful fallbacks and explainable interactions, every step must prioritize the user experience. Ethical oversight, feedback loops, and staged feature rollout ensure AI adds value without overwhelming or alienating users. Ultimately, AI shouldn’t replace your product’s humanity; it should **amplify it**, enhancing workflows, boosting efficiency, and making technology feel smarter, yet always human-centered.

## Want this applied to your product? Book a call.

[Book a call](https://superflow.design/contact)