2026
TalentLab
Giving a rural employment service a shared method for career guidance, with AI doing the administrative work and people keeping the decisions.
AI Integration
Service Design
Research
Product Overview
“The brief was to bring AI into a career guidance service. The first thing I did was stop asking what AI could do,
because that question only ever returns what people already believe AI does.”
Method first, tools second
TalentLab is a project with the employment service of the Lluçanès, a rural county losing its qualified young people to Barcelona. The service helps people who left to study and want to come back, and its two counsellors are psychologists with years of experience in guidance, CVs and interviews.
The goal is not to build software. It is to take the know-how those two people already carry, give it a shared method, and use existing AI tools to absorb the administrative and mechanical parts so the human conversation stays human. Delivery runs to December 2027, in three-week cycles, with something usable at the end of each one.
0
0
Friction Areas Mapped
Consolidated from the field sessions
0
0
Maturity Dimensions
Four levels each, to place any candidate
0
0
Week Cycles
Something usable at the end of each one
The Brief and Its Constraints
The interesting part of a public-sector brief is usually what you are not allowed to touch.
What was fixed from day one
The case management system stays untouched: it is being improved by another department, and anything we build has to fit around it without duplicating data or work. GDPR compliance is strict, with anonymisation and informed consent. And one rule sits above everything: the AI never decides on its own. Every output is reviewed by a person before it reaches a candidate.
Two very different users
The candidate who wants to come back but is still weighing options and prioritises the professional fit, and the candidate who is coming back regardless and prioritises location. The service is modular by design: it can be one session or many, and that is a professional judgement, not a funnel.
The tension I wrote down first
The expectation in a project like this is a single magic centralised tool. The reality is field methodology and adaptation to what already exists. I listed nine of these tensions at the start, budget against premium subscriptions, power against simplicity for non-technical users, a common team tool against two counsellors already using different assistants. Not to resolve them, but so that every cycle decision could be made knowing which balance it was moving.
Research Method
Before proposing anything technical, I went to look at the work.
Ask about the friction, never about the AI
When you ask someone what AI could do for them, the answer is bounded by whatever they currently believe AI does. When you ask what is annoying, what is repetitive, what you cannot see, the whole problem shows up and the decision about what is automatable stays with me. Every question in every session was phrased that way.
Contrast before the sessions
I interviewed a counsellor from another county who uses the same case management system, so I arrived at the sessions with sharper hypotheses than the team could have given me alone. I was explicit that everything from that interview was a hypothesis to validate, not a finding. That mattered: several of those assumptions turned out to be false here.
An analysis board with one rule
The whole discovery ran on a shared board with a strict principle. Only frictions, problems and opportunities go on it. Solutions never get written there, even when they surface mid-conversation, they get parked somewhere else and picked up later. It is the only reliable way I know to stop a diagnosis from closing before it has been made.
What We Mapped
Five artefacts, each one built to answer a different question about the service.
The process map
Three tracks, company, counsellor and candidate, following the whole timeline from a company noticing a need to a candidate being placed. The counsellor track sits in the middle, with the service’s interventions marked on a process that would otherwise happen without them. Making those moments visible is what allowed us to talk about friction precisely instead of generally.
The ecosystem map
Who surrounds the service and with what tools. This is where the borrowed assumptions died. The design tool is not used for CVs at all but for social posts, one resource portal is deliberately excluded for territorial reasons, and the professional network is only accessed from one counsellor’s personal account.
The candidate maturity table
Eight dimensions a candidate can work on, professional narrative, CV, network profile, digital presence, supporting material, market attention, search management and contacts, each with four levels. The finding was that the top level is not the goal for everyone. For many local vacancies it does not pay off and the cost of maintaining it is real. The table places a candidate, it is not a ladder to complete.
The network uses matrix
Every objective you can pursue on a professional network and the minimum required for each. It concluded that the network is useless for hyperlocal search: its geographic taxonomy does not go below large city or county, and result ranking is personalised by the searcher’s own network in a way that cannot be reset. The public job portals already index at municipality level. The network does have value, but earlier, for visibility before a vacancy opens.
The tools table
What the service actually uses, for what, and what it gains and loses with each.
The Six Frictions
Everything on the board consolidated into six problem areas, then prioritised by the people who live them.
How the vote worked
Six areas, written in a uniform format so they could be compared. Each counsellor distributed three points, three to whatever weighs most, two to the next, one to the third. Six points between two voters.
What came out on top
Diagnosis by intuition, with no method, took five. The initial assessment of each candidate depends on the experience and the eye of whoever does it. There is no shared criterion, no guarantee two counsellors facing the same case land in the same place, and no trace of why one route was chosen over another.
Second, the radius problem
With four points. Market knowledge is very fine up close and thins out fast as it moves away from familiar ground. Opportunities outside the county, remote, or in sectors they do not usually touch fall outside the service’s field of vision.
Reading the Vote
The interesting part of that vote is not the ranking. It is the shape.
A sum would have given the wrong answer
The two counsellors do not agree on their maximum score. What they agree on is that those two frictions are worth points at all. Every other area was voted by at most one of them. The decision about where to start does not come from adding up the numbers, it comes from the only real overlap between two people who see the same job differently.
The two zeros are the most structural problems in the brief
Young people arriving with no experience and no local reference for their profession, and a slow-moving rural business fabric that struggles to absorb qualified profiles. Nobody gave either of them a single point. That is not a lapse of judgement. They voted for what can be solved from their own desk, and those two are frictions of the system, not of the daily work. They get handled on a different plane of the project rather than treated as irrelevant.
The First Build
The first cycle attacks the diagnosis, because it is the door everything else passes through. If the initial assessment has no method, everything downstream inherits the same arbitrariness.
Four steps, three of them built
A decision tree with closed questions places the candidate’s situation and leads to a concrete endpoint, on fixed logic with no AI in it at all. Where there is no clear direction, the candidate is routed to an existing validated instrument rather than a new one built for the occasion. Then the profile is crossed against the territory knowledge file, which is the step that carries the differential value. The fourth step, what material each endpoint calls for, is deliberately not designed yet.
Keep the knowledge layer separate from the generation layer
The territory knowledge file, sectors, companies, viable local and remote roles, training paths, is an independent document that the AI only reads at the endpoints of the tree. It never invents data about the territory, and it holds no candidate data at all. This is the architectural decision the whole thing rests on.
What we deliberately did not build
No configured assistant in cycle one. Building the assistant before the knowledge file exists would produce confident, invented answers about a county whose whole value in this project is that the knowledge is real. The second friction gets a low-cost partial answer instead: alerts configured on the relevant portals, which does not solve the radius problem but attacks the habit and tooling part of it and lets us measure the impact cheaply.
What This Was Really About
On paper this is an AI project.
In practice almost none of the first cycle is AI, and that is the point. The work was getting two experienced people to agree on where the pain actually is, and then being disciplined about what not to automate. The decision tree runs on fixed logic. The test already existed. The only place the AI is allowed near is the one place where local knowledge, written down by people who have it, gives it something true to say.
More Works
©2026
2026
TalentLab
Giving a rural employment service a shared method for career guidance, with AI doing the administrative work and people keeping the decisions.
AI Integration
Service Design
Research
Product Overview
“The brief was to bring AI into a career guidance service. The first thing I did was stop asking what AI could do,
because that question only ever returns what people already believe AI does.”
Method first, tools second
TalentLab is a project with the employment service of the Lluçanès, a rural county losing its qualified young people to Barcelona. The service helps people who left to study and want to come back, and its two counsellors are psychologists with years of experience in guidance, CVs and interviews.
The goal is not to build software. It is to take the know-how those two people already carry, give it a shared method, and use existing AI tools to absorb the administrative and mechanical parts so the human conversation stays human. Delivery runs to December 2027, in three-week cycles, with something usable at the end of each one.
0
0
Friction Areas Mapped
Consolidated from the field sessions
0
0
Maturity Dimensions
Four levels each, to place any candidate
0
0
Week Cycles
Something usable at the end of each one
The Brief and Its Constraints
The interesting part of a public-sector brief is usually what you are not allowed to touch.
What was fixed from day one
The case management system stays untouched: it is being improved by another department, and anything we build has to fit around it without duplicating data or work. GDPR compliance is strict, with anonymisation and informed consent. And one rule sits above everything: the AI never decides on its own. Every output is reviewed by a person before it reaches a candidate.
Two very different users
The candidate who wants to come back but is still weighing options and prioritises the professional fit, and the candidate who is coming back regardless and prioritises location. The service is modular by design: it can be one session or many, and that is a professional judgement, not a funnel.
The tension I wrote down first
The expectation in a project like this is a single magic centralised tool. The reality is field methodology and adaptation to what already exists. I listed nine of these tensions at the start, budget against premium subscriptions, power against simplicity for non-technical users, a common team tool against two counsellors already using different assistants. Not to resolve them, but so that every cycle decision could be made knowing which balance it was moving.
Research Method
Before proposing anything technical, I went to look at the work.
Ask about the friction, never about the AI
When you ask someone what AI could do for them, the answer is bounded by whatever they currently believe AI does. When you ask what is annoying, what is repetitive, what you cannot see, the whole problem shows up and the decision about what is automatable stays with me. Every question in every session was phrased that way.
Contrast before the sessions
I interviewed a counsellor from another county who uses the same case management system, so I arrived at the sessions with sharper hypotheses than the team could have given me alone. I was explicit that everything from that interview was a hypothesis to validate, not a finding. That mattered: several of those assumptions turned out to be false here.
An analysis board with one rule
The whole discovery ran on a shared board with a strict principle. Only frictions, problems and opportunities go on it. Solutions never get written there, even when they surface mid-conversation, they get parked somewhere else and picked up later. It is the only reliable way I know to stop a diagnosis from closing before it has been made.
What We Mapped
Five artefacts, each one built to answer a different question about the service.
The process map
Three tracks, company, counsellor and candidate, following the whole timeline from a company noticing a need to a candidate being placed. The counsellor track sits in the middle, with the service’s interventions marked on a process that would otherwise happen without them. Making those moments visible is what allowed us to talk about friction precisely instead of generally.
The ecosystem map
Who surrounds the service and with what tools. This is where the borrowed assumptions died. The design tool is not used for CVs at all but for social posts, one resource portal is deliberately excluded for territorial reasons, and the professional network is only accessed from one counsellor’s personal account.
The candidate maturity table
Eight dimensions a candidate can work on, professional narrative, CV, network profile, digital presence, supporting material, market attention, search management and contacts, each with four levels. The finding was that the top level is not the goal for everyone. For many local vacancies it does not pay off and the cost of maintaining it is real. The table places a candidate, it is not a ladder to complete.
The network uses matrix
Every objective you can pursue on a professional network and the minimum required for each. It concluded that the network is useless for hyperlocal search: its geographic taxonomy does not go below large city or county, and result ranking is personalised by the searcher’s own network in a way that cannot be reset. The public job portals already index at municipality level. The network does have value, but earlier, for visibility before a vacancy opens.
The tools table
What the service actually uses, for what, and what it gains and loses with each.
The Six Frictions
Everything on the board consolidated into six problem areas, then prioritised by the people who live them.
How the vote worked
Six areas, written in a uniform format so they could be compared. Each counsellor distributed three points, three to whatever weighs most, two to the next, one to the third. Six points between two voters.
What came out on top
Diagnosis by intuition, with no method, took five. The initial assessment of each candidate depends on the experience and the eye of whoever does it. There is no shared criterion, no guarantee two counsellors facing the same case land in the same place, and no trace of why one route was chosen over another.
Second, the radius problem
With four points. Market knowledge is very fine up close and thins out fast as it moves away from familiar ground. Opportunities outside the county, remote, or in sectors they do not usually touch fall outside the service’s field of vision.
Reading the Vote
The interesting part of that vote is not the ranking. It is the shape.
A sum would have given the wrong answer
The two counsellors do not agree on their maximum score. What they agree on is that those two frictions are worth points at all. Every other area was voted by at most one of them. The decision about where to start does not come from adding up the numbers, it comes from the only real overlap between two people who see the same job differently.
The two zeros are the most structural problems in the brief
Young people arriving with no experience and no local reference for their profession, and a slow-moving rural business fabric that struggles to absorb qualified profiles. Nobody gave either of them a single point. That is not a lapse of judgement. They voted for what can be solved from their own desk, and those two are frictions of the system, not of the daily work. They get handled on a different plane of the project rather than treated as irrelevant.
The First Build
The first cycle attacks the diagnosis, because it is the door everything else passes through. If the initial assessment has no method, everything downstream inherits the same arbitrariness.
Four steps, three of them built
A decision tree with closed questions places the candidate’s situation and leads to a concrete endpoint, on fixed logic with no AI in it at all. Where there is no clear direction, the candidate is routed to an existing validated instrument rather than a new one built for the occasion. Then the profile is crossed against the territory knowledge file, which is the step that carries the differential value. The fourth step, what material each endpoint calls for, is deliberately not designed yet.
Keep the knowledge layer separate from the generation layer
The territory knowledge file, sectors, companies, viable local and remote roles, training paths, is an independent document that the AI only reads at the endpoints of the tree. It never invents data about the territory, and it holds no candidate data at all. This is the architectural decision the whole thing rests on.
What we deliberately did not build
No configured assistant in cycle one. Building the assistant before the knowledge file exists would produce confident, invented answers about a county whose whole value in this project is that the knowledge is real. The second friction gets a low-cost partial answer instead: alerts configured on the relevant portals, which does not solve the radius problem but attacks the habit and tooling part of it and lets us measure the impact cheaply.
What This Was Really About
On paper this is an AI project.
In practice almost none of the first cycle is AI, and that is the point. The work was getting two experienced people to agree on where the pain actually is, and then being disciplined about what not to automate. The decision tree runs on fixed logic. The test already existed. The only place the AI is allowed near is the one place where local knowledge, written down by people who have it, gives it something true to say.
More Works
©2026
2026
TalentLab
Giving a rural employment service a shared method for career guidance, with AI doing the administrative work and people keeping the decisions.
AI Integration
Service Design
Research
Product Overview
“The brief was to bring AI into a career guidance service. The first thing I did was stop asking what AI could do,
because that question only ever returns what people already believe AI does.”
Method first, tools second
TalentLab is a project with the employment service of the Lluçanès, a rural county losing its qualified young people to Barcelona. The service helps people who left to study and want to come back, and its two counsellors are psychologists with years of experience in guidance, CVs and interviews.
The goal is not to build software. It is to take the know-how those two people already carry, give it a shared method, and use existing AI tools to absorb the administrative and mechanical parts so the human conversation stays human. Delivery runs to December 2027, in three-week cycles, with something usable at the end of each one.
0
0
Friction Areas Mapped
Consolidated from the field sessions
0
0
Maturity Dimensions
Four levels each, to place any candidate
0
0
Week Cycles
Something usable at the end of each one
The Brief and Its Constraints
The interesting part of a public-sector brief is usually what you are not allowed to touch.
What was fixed from day one
The case management system stays untouched: it is being improved by another department, and anything we build has to fit around it without duplicating data or work. GDPR compliance is strict, with anonymisation and informed consent. And one rule sits above everything: the AI never decides on its own. Every output is reviewed by a person before it reaches a candidate.
Two very different users
The candidate who wants to come back but is still weighing options and prioritises the professional fit, and the candidate who is coming back regardless and prioritises location. The service is modular by design: it can be one session or many, and that is a professional judgement, not a funnel.
The tension I wrote down first
The expectation in a project like this is a single magic centralised tool. The reality is field methodology and adaptation to what already exists. I listed nine of these tensions at the start, budget against premium subscriptions, power against simplicity for non-technical users, a common team tool against two counsellors already using different assistants. Not to resolve them, but so that every cycle decision could be made knowing which balance it was moving.
Research Method
Before proposing anything technical, I went to look at the work.
Ask about the friction, never about the AI
When you ask someone what AI could do for them, the answer is bounded by whatever they currently believe AI does. When you ask what is annoying, what is repetitive, what you cannot see, the whole problem shows up and the decision about what is automatable stays with me. Every question in every session was phrased that way.
Contrast before the sessions
I interviewed a counsellor from another county who uses the same case management system, so I arrived at the sessions with sharper hypotheses than the team could have given me alone. I was explicit that everything from that interview was a hypothesis to validate, not a finding. That mattered: several of those assumptions turned out to be false here.
An analysis board with one rule
The whole discovery ran on a shared board with a strict principle. Only frictions, problems and opportunities go on it. Solutions never get written there, even when they surface mid-conversation, they get parked somewhere else and picked up later. It is the only reliable way I know to stop a diagnosis from closing before it has been made.
What We Mapped
Five artefacts, each one built to answer a different question about the service.
The process map
Three tracks, company, counsellor and candidate, following the whole timeline from a company noticing a need to a candidate being placed. The counsellor track sits in the middle, with the service’s interventions marked on a process that would otherwise happen without them. Making those moments visible is what allowed us to talk about friction precisely instead of generally.
The ecosystem map
Who surrounds the service and with what tools. This is where the borrowed assumptions died. The design tool is not used for CVs at all but for social posts, one resource portal is deliberately excluded for territorial reasons, and the professional network is only accessed from one counsellor’s personal account.
The candidate maturity table
Eight dimensions a candidate can work on, professional narrative, CV, network profile, digital presence, supporting material, market attention, search management and contacts, each with four levels. The finding was that the top level is not the goal for everyone. For many local vacancies it does not pay off and the cost of maintaining it is real. The table places a candidate, it is not a ladder to complete.
The network uses matrix
Every objective you can pursue on a professional network and the minimum required for each. It concluded that the network is useless for hyperlocal search: its geographic taxonomy does not go below large city or county, and result ranking is personalised by the searcher’s own network in a way that cannot be reset. The public job portals already index at municipality level. The network does have value, but earlier, for visibility before a vacancy opens.
The tools table
What the service actually uses, for what, and what it gains and loses with each.
The Six Frictions
Everything on the board consolidated into six problem areas, then prioritised by the people who live them.
How the vote worked
Six areas, written in a uniform format so they could be compared. Each counsellor distributed three points, three to whatever weighs most, two to the next, one to the third. Six points between two voters.
What came out on top
Diagnosis by intuition, with no method, took five. The initial assessment of each candidate depends on the experience and the eye of whoever does it. There is no shared criterion, no guarantee two counsellors facing the same case land in the same place, and no trace of why one route was chosen over another.
Second, the radius problem
With four points. Market knowledge is very fine up close and thins out fast as it moves away from familiar ground. Opportunities outside the county, remote, or in sectors they do not usually touch fall outside the service’s field of vision.
Reading the Vote
The interesting part of that vote is not the ranking. It is the shape.
A sum would have given the wrong answer
The two counsellors do not agree on their maximum score. What they agree on is that those two frictions are worth points at all. Every other area was voted by at most one of them. The decision about where to start does not come from adding up the numbers, it comes from the only real overlap between two people who see the same job differently.
The two zeros are the most structural problems in the brief
Young people arriving with no experience and no local reference for their profession, and a slow-moving rural business fabric that struggles to absorb qualified profiles. Nobody gave either of them a single point. That is not a lapse of judgement. They voted for what can be solved from their own desk, and those two are frictions of the system, not of the daily work. They get handled on a different plane of the project rather than treated as irrelevant.
The First Build
The first cycle attacks the diagnosis, because it is the door everything else passes through. If the initial assessment has no method, everything downstream inherits the same arbitrariness.
Four steps, three of them built
A decision tree with closed questions places the candidate’s situation and leads to a concrete endpoint, on fixed logic with no AI in it at all. Where there is no clear direction, the candidate is routed to an existing validated instrument rather than a new one built for the occasion. Then the profile is crossed against the territory knowledge file, which is the step that carries the differential value. The fourth step, what material each endpoint calls for, is deliberately not designed yet.
Keep the knowledge layer separate from the generation layer
The territory knowledge file, sectors, companies, viable local and remote roles, training paths, is an independent document that the AI only reads at the endpoints of the tree. It never invents data about the territory, and it holds no candidate data at all. This is the architectural decision the whole thing rests on.
What we deliberately did not build
No configured assistant in cycle one. Building the assistant before the knowledge file exists would produce confident, invented answers about a county whose whole value in this project is that the knowledge is real. The second friction gets a low-cost partial answer instead: alerts configured on the relevant portals, which does not solve the radius problem but attacks the habit and tooling part of it and lets us measure the impact cheaply.
What This Was Really About
On paper this is an AI project.
In practice almost none of the first cycle is AI, and that is the point. The work was getting two experienced people to agree on where the pain actually is, and then being disciplined about what not to automate. The decision tree runs on fixed logic. The test already existed. The only place the AI is allowed near is the one place where local knowledge, written down by people who have it, gives it something true to say.
More Works
©2026