About the author
Nikša Demović
Business Development Manager
Introduction
In software development, building new features for abstract personas is nothing unusual. That tendency has only grown with the pressure to infuse AI into everything, even as teams are warned not to follow that path blindly.
This article tells the story of how we built AI for the employability sector through our collaboration with Candid AI, where our users were as real as it gets: caseworkers juggling high caseloads and participants whose lives are anything but simple. This project pushed us to design not around idealized flows, but around our users’ real-world situations – which is how it should be every time.
The Caseworkers’ Perspective
The challenge was obvious: caseworkers were drowning in paperwork. They had to turn complex, sensitive stories into clear, job‑ready CVs and applications while working under constant time pressure. Formatting, rewriting, and tailoring documents to different roles took a lot of manual effort, and any slowdown delayed people’s progress towards employment. In other words, the bottleneck wasn’t with or without AI, but how to properly support caseworkers without compromising quality or empathy.
“One Size Fits All” Was Never Going to Work
For participants from many different backgrounds, each story needed to be told in a way that was accurate, respectful, and aligned with sector standards. Specific circumstances such as long‑term unemployment, health conditions, caring responsibilities, digital skills gaps, or badly timed career breaks require more than a good‑enough, generic template.
Off‑the‑shelf platforms produced repetitive or impersonal content and simply didn’t reflect how employability professionals really work. That’s why the solution had to be AI trained by real CV writers and able to adapt to highly individual contexts.
The POC Discovery Workshop: What did it Help Us Learn?
We started with a POC Discovery Workshop to align on outcomes and stress‑test ideas before committing to a product direction. This phase helped both Notch and Candid define a few core questions:
- Which parts of the process most needed speed?
- Where was a high level of empathy a must‑have?
- Which tasks could be safely automated, and which needed human review?
That workshop shaped the experiments we ran later and stopped us from chasing shiny but irrelevant features.
Why Did We Use a Two‑Phase Delivery Model?
We split the work into two major phases: a Proof of Concept (POC) and then an MVP discovery and definition phase. The POC’s job was to de‑risk both technical and business assumptions.
By the time we moved into MVP discovery, we already knew which directions had real potential. That meant the workshops in Phase 2 could focus on user journeys, flows, and product decisions instead of debating whether the whole idea was even feasible.
The AI Proof of Concept
The POC was run by a cross‑functional team – AI engineers, business analysts and Candid team including professional CV workers – working in a time-boxed window to explore ten concrete scenarios. Those scenarios were designed around actual casework, not just hypothetical AI demos.
We tried out things like generating CVs from minimal input, reframing seemingly unappealing history into strengths, spotting gaps or inconsistencies, and keeping tone and style consistent across documents. We also explored tailoring for different jobs and industries and used external job data to keep outputs realistic rather than generic.
Every scenario gave us a clearer sense of where AI could make a meaningful difference.
The MVP Discovery: Turning Experiments into a Product
Once we had enough proof that the core ideas worked, we switched from Can we do this? to How should this work for real people? That’s where the MVP discovery workshops came in.
Over four days, stakeholders, advisors, and designers mapped how CVs are really created in the employability sector. This means detecting where time pressure hits hardest, where admin work piles up, and where emotional labor is heaviest. Using tools like personas, empathy maps, customer journeys, and post‑up sessions, we turned scattered observations into structured user flows and a clear MVP feature set.
How Does the Platform Keep Humans in Control?
Built on this foundation, the platform took shape as a human‑led, AI‑powered system rather than an AI‑only solution. Caseworkers capture a client’s key skills, experience, interests, and personality traits, and the system generates a tailored CV – but the crucial part is control. The platform highlights suggested improvements, lets caseworkers reword sections in seconds, and makes it easy to adapt the CV for different roles without starting from scratch. Managers get dashboards that show activity and performance across teams, while the narrative of each participant stays in the hands of the professionals: after all, they know those people.
How Did We Decide What to Build First?
Even with a strong concept, not everything can go into the first release. We used prioritization techniques like MoSCoW to decide which capabilities were essential at launch and which could wait. Fast CV generation, easy tailoring, intuitive editing, and management dashboards made the must-have list.
These priorities sat on top of a validated architecture and a clear roadmap, so the MVP isn’t just a prototype, but a solid first version designed to deliver measurable value and evolve without constant rework.
What Can Other Organizations Learn from Candid’s Approach?
The Candid project shows that businesses benefit most when AI is wrapped around real workflows and real constraints. Starting with structured discovery, testing tightly scoped scenarios, and then designing around real user journeys led to a platform that supports both efficiency and dignity.
If you want to see the results and impact, read our customer story on building with Candid AI to see how this process is being built specifically for the employment sector.
