A decade ago a global staffing and HR services group was writing the integrating platform blueprint behind two business propositions. One cell in the data maturity grid, at level four of five, read: "Suggestions engine to advise best time to move." Six words describing what a reader today would call a career agent. The document was a version 0.9 working draft, not a delivered system, and that is what makes it worth reading now.
The challenge
The group ran across many country markets on legacy systems that were fragmented and had absorbed a lot of local development and tailoring, and it had two business propositions to serve from one platform. The first was relationship-led, aimed at professional talent, and needed depth: personalised profiles, advice, engagement that carried across channels. The second was automated and digital by default, and needed fast match-making above all else.
On the five-level customer experience model, the current state was marked at level one. Experience was pushed from the inside out, built around the services the group wanted to offer, and it varied across individuals and countries. The multichannel experience was recorded as unintegrated and fragmented. One rung up, in the omnichannel row, sat the note that media planning was siloed, fragmenting end-to-end journeys, touchpoints and cross-channel experience.
The data model was built on the same five rungs. Its bottom rung, analyzing, meant local use of data for analysis and reporting: in-country or business-unit reporting, one-off analyses of the impact of new initiatives and of profit margins per company type. Useful work, scoped to one country or one unit at a time.
Against that, the ambition was specific. The relationship-led proposition needed the top of the data scale, self-learning, because multidimensional personal profiles require internal and external sources fused together. The digital-by-default proposition needed level four, pro-active, because an accurate straight-through-processing algorithm requires data integrated company-wide. The slide setting that second target named the binding constraint in its own title: high data quality is what enables automated matching and near-real-time job suggestions. Not better algorithms. Data quality.
The approach
The data model graded its levels by the scope over which data was used, not by which tools were installed: local, cross-functional, cross-business-unit, company-wide, ecosystem-wide. The experience model graded by what the services were organised around, climbing from an inside-out push of services, to services built around market competition, then customer needs, then developing relationships, and finally services as one continuous experience. Grading by scope and orientation rather than by technique made a level assessment hard to fake, and put the difficulty where it sits: integration and governance.
The career agent then appeared as a specification, level by level, rather than a feature request. Both propositions were given the same populated ladder, cell for cell; only the target level differed.
Peer reasoning ran down the spine of it, one rung at a time. Level two compared usage statistics between peer groups. Level three determined likeliness between peers from self-proclaimed skills and preferences. Level four filled in missing profile elements by comparing a person to their peers rather than asking them for more data. Level five gave career advice based on successful moves by peers. Three algorithm classes were named as what the data capability had to enable: smart matching, job suggestions, demand predictions.
Suggestions engine to advise best time to move.
The rest of the blueprint answered how it could be built. Four operating models were evaluated against the five design principles: a central authority, a central mall of services, a command tower, and a hub-and-spoke arrangement. Hub and spoke was carried forward on plainly stated reasoning. Centralised models failed on local fit and were hard to implement in a group that was not already centralised. A model with no single point of accountability relied on goodwill for progress. Hub and spoke put the experience leader accountable for delivery and coordination, kept execution local, and made the central unit contain local representatives, which is the mechanism by which best practice actually moves.
The organisation question was split in two and answered separately: which roles are needed and when, then how to source them. Five skill families were required, spanning creative, technology, data, operations and management, with the majority needed in the first horizon to detail the strategy as well as execute it. Sourcing ran three routes in parallel: short-term contract hires to accelerate development while permanent staff were recruited and existing staff upskilled. Two recorded lessons still read well: hire key roles early enough that they shape the vision rather than inherit it, and do not overlook the potential to retrain existing staff.
The outcome
The deck should be read as what it was, a document still in draft. It set targets, chose an operating model, sized the skill blend and recorded what its authors had learned the hard way elsewhere. It did not report a shipped recommendation engine, and it left open questions visible on the slides, including whether segmentation should be defined globally or country by country. What it produced is a decision record, and the record holds up a decade on.
Start with the agent. When it was written, "advise best time to move" implied a multi-year programme, and the ladder said so by putting the prerequisites on the rungs below it: a 360 degree profile assembled from unstructured touchpoint data and peer-likeliness modelling at level three, taxonomy and folksonomy work running alongside, near-real-time vacancy feeds at level four, and a data estate integrated company-wide before any of it paid out. Today most of that stack is bought rather than built. A language model reads unstructured CV text, job descriptions and conversation transcripts without a folksonomy programme in front of it. Embeddings do peer similarity without waiting for people to self-declare skills. The suggestion that took a level-four data estate to compute now arrives as a sentence, in the second person, from an agent watching the vacancy feed and the person's own job history continuously.
Look closer at what level three was actually asking for. A 360 degree profile assembled from unstructured touchpoint data, in a staffing business, is mostly not text sitting in a database. It is documents and images: CVs in every layout a person can invent, scanned diplomas and certificates, right-to-work paperwork photographed on a phone, signed timesheets coming back from a placement. At the time that meant an OCR project with a taxonomy and folksonomy programme behind it, which is exactly the work the blueprint put on the rungs below the career agent. Computer vision now reads the layout along with the characters, so a certificate becomes a dated claim attached to a profile rather than a PDF nobody opened, and the missing profile elements the ladder wanted filled at level four are frequently already in the candidate's own file.
The three algorithm classes the blueprint named, smart matching, job suggestions and demand predictions, were drawn as capabilities to acquire one by one. They work as one connected pipeline instead. Vacancy feeds land and an agent parses and normalises them, another keeps profiles current from incoming documents and conversation notes, matching runs continuously rather than when someone opens a screen, and the suggestion goes out with its reasoning attached. Every handoff in that chain is a place a person used to re-key something. Where a step is fully specified and reversible, normalising a vacancy or refreshing a profile from a new certificate, the agent runs it autonomously and logs what it did; where it touches somebody's job, a recruiter sees it before the candidate does. Autonomy is a scope decision taken step by step, not a setting on a product, and the effect of taking it seriously is that agents carry a real share of how a branch runs its day rather than sitting beside the work as a tool someone remembers to open.
The staffing answer is the part that collapses hardest, which is worth saying plainly in a case study about a staffing group. The organisation section scoped five skill families across creative, technology, data, operations and management, most of them needed in the first horizon, filled through three sourcing routes running in parallel. Read the same ladder now and most of those rungs are loop and harness engineering rather than headcount. The loop is the agent that reads the vacancy, drafts the suggestion, checks it against the candidate's record and revises before a human sees it. The harness is what decides which systems it may touch, on whose identity, under which country's labour rules, and what it writes to the log. Both are built by a small team that knows the domain, and both can be run against recorded candidate journeys before a single market goes live. A programme that was scoped in horizons is now paced by how fast you can build and evaluate those two things, which is much faster than hiring five skill families and waiting for them to agree.
What has not moved is the constraint the blueprint named. Level four was defined as company-wide integration, and data quality was written into a slide title as the enabler of automated matching. Agents do not relax that requirement. They raise the cost of missing it, because a model grounded on stale or contradictory candidate records produces fluent, confident, wrong advice, and it delivers that advice in the first person to someone deciding whether to change job. The failure mode then was a dashboard nobody trusted. The failure mode now is a plausible recommendation nobody can audit.
The operating-model argument survives intact, with the nouns swapped. Central authority against local execution is now the question of who owns agents, prompts, evaluation sets and guardrails, and the failure modes are the ones the blueprint scored: a central build that ignores local labour law and local hiring practice, or local builds nobody can evaluate or govern. Central capability staffed with local representatives, local ambassadors driving adoption, and one named person accountable end to end remains the right default for an AI platform.
One line does need rewriting. At level five, above the career agent, the blueprint asked the system to predict which talents will search for a job. That is now straightforward to build, which turns it from a modelling question into a consent question. Advising a person when to move is a service to that person. Predicting they are about to leave, for someone else's benefit, is surveillance of them unless they asked for it. The two sat one rung apart on the same ladder and the document did not draw a line between them. Anyone building this now has to: European AI regulation puts employment and worker-management systems in its high-risk tier, where the obligations attach whether or not the prediction is accurate.
For a staffing or talent leader now, the agent is the easy part. The two questions the blueprint opened with still decide whether it works: what scope does our data actually cover, and who is accountable for the experience from end to end. What has changed is the clock. Six words that once implied a programme measured in horizons are now a loop, a harness and an evaluation set, and the remaining schedule is mostly the time it takes to answer those two questions honestly.
- 5
- Levels per maturity model
- Level 4
- Data floor across both blueprints
- 4
- Operating models evaluated