The strongest lever is access

delivers it, and where the boundary lies.

One argument, two delivery channels An AI platform for a whole profession, and the Expertise Amplifier for one theme. Both aim at the same lever, both serve the individual, the team and the organization, and both leave the boundary intact.

·

·

·

L&D buys knowledge where the work needs access

Ask an L&D function what it delivers and the answer is about knowledge: modules, learning paths, a course that brings people up to date. Ask a nurse, an engineer or an advisor what they need at the moment itself and the answer is about something else. The right rule, the right threshold, the agreement that changed last month. Not in advance. Now.

That difference is not a detail. It decides where the money goes. A function that thinks in knowledge buys courses and measures completions. A function that thinks in access designs the work so that what people need is within reach when the question arrives, and measures whether the result moved.

Neither is a matter of taste. The research on what changes behaviour has a clear ranking, and it is uncomfortable for anyone whose product is a course. Knowledge as the target of an intervention does close to nothing. Access does the most.

One more thing changes the arithmetic. Almost everyone in work already holds the licence to operate: the qualification, the diploma, the registration. After that, formal training is genuinely required where the law or the sector prescribes it, and that is the compliance route. Everything else is not a training question at all. It is a question about the cause of a performance problem, which an advisor establishes through performance consulting, with the Performance Consultant AI alongside (see

The request is not the problem

).

This paper puts four things side by side: what the evidence says about access, where the idea comes from, how an organization delivers it without a two-year project, and where the boundary lies. The boundary belongs in the argument, not in the small print.

Sixteen determinants, one clear ranking

A synthesis of meta-analyses across sixteen determinants of behaviour finds a ranking that is hard to argue with. Knowledge, general skills, attitudes and beliefs have negligible effects as targets of an intervention. Habits and social support have medium effects. Access, the material and logistic means that make a behaviour feasible, has the largest (Albarracín et al., 2024).

Read that against a training catalogue and the conclusion is blunt. Most of what L&D sells aims at the weakest determinant on the list, and does so at the highest cost, because the working hours it consumes never reach the invoice.

knowledge

general skills negligible attitudes

beliefs

habits

social support

access

Grouped by how much each determinant moves behaviour when an intervention targets it. The bars show the ranking, not effect sizes.

The idea is older than the technology

Gery described the destination in 1991. What arrived late was the vehicle.

Gilbert

Gery

Five moments

Fifty years in one line. Everything up to and including 70:20:10 aims at one person; what AI adds is augmentation, and a canon a team and an organization can share.

Performance support is not a new word, and moment of need is not a new idea. Gloria Gery gave both their shape in 1991: a system that puts the right knowledge, in the right amount, within reach of the person doing the work (Gery, 1991).

Two things in that definition matter now. It was information, and it was passive. Documents, procedures, step-by-step guidance, mostly for procedural tasks. And it leaned on technology that in 1991 barely existed, so the idea outran its execution for thirty years.

Three models L&D runs on, and all three aim at one person

Five moments of need

Three models, one level. Everything above the individual falls outside them.

Five moments of need

70:20:10 Gottfredson and Mosher took Gery’s idea and

Moved the attention from the course to the work, named the moments at which a person needs

and that was a necessary correction (Arets, support: learning something for the first time,

Jennings & Heijnen, 2015). It still counts what one expanding it, applying it, solving a problem,

person learns, wherever that learning happens. adapting to change (2011). Useful vocabulary for when support is needed.

These three are what the field runs on, and they share one limitation, which is structural. Each describes one person meeting one task. None has a place for the team that has to reach one shared answer, or for the organization that has to keep what it works out, and the environment around the work sits outside all three.

The question does not arrive in the classroom

A course sits weeks before the work. The question arrives inside it: at the bedside, at the machine, in the intake conversation. Everything between those two points is memory, and memory is exactly the part the evidence says not to lean on. Access moves the answer to where the question is.

the question arrives here a course

before the work

at the moment of need

expertise within reach

The distance between the two is where most of the value leaks out of a training budget.

At the bedside

At the machine May a care worker administer this on their own?

The step after the exception, the torque for this The answer sits in a working agreement that

variant, the form that goes with it. Trained once, changed in June. Nobody knows that by heart,

needed twice a year. and nobody should have to.

What is learned and not used, fades

Skills decay, and the loss grows with the time a skill goes unused (Arthur et al., 1998). That is not a criticism of learning. It is a statement about where learning belongs in time. Everything that only comes up twice a year is a poor candidate for a course and a good candidate for access.

Organizations lose knowledge in the same way. What one team works out disappears when the roster changes, unless somebody stores it where it can be found and keeps it current (Argote, 2024). An assistant that answers from a governed canon is one of the few places where that maintenance is somebody’s job by design.

do n ca ne eo m so t ha w

Schematic. What a course leaves behind fades with disuse; what stays within reach in the work does not depend on memory.

What may be looked up, and what may not

The argument only holds with its boundary attached. Some things have to be there without looking: acting under time pressure, reasoning your way through an unfamiliar fault or an unfamiliar patient, coordinating with a team, seeing that a case is unusual before anyone says so. Those need whole-task practice, deliberately designed and built up over time (Van Merriënboer, Kirschner & Frèrejean, 2024). No assistant produces them.

what may be looked up

a rule that changed last month

a threshold, a dose, a torque, a deadline

the step after the exception

which form, which route, which owner

The line between the two is a design decision, taken per task, and it is the first question in any serious conversation about AI in learning.

Two more things stay outside. An assistant produces no record of demonstrated competence, so proving competence remains its own route. And an answer is worth exactly the canon it rests on: an assistant on an ungoverned pile of documents replaces a knowledge problem with a trust problem.

The step after e-learning and performance support

Nothing here breaks with what the profession already does. E-learning solved reach. Performance support moved the answer into the work (Gery, 1991). What changes is what an assistant does with a governed canon: it answers in the language of the question, stays current because one editorial board keeps the canon, and serves the team.

E-learning

Performance support

knowledge in advance

the answer within reach

what it solved

what it solved reach, at low cost per head

support inside the work itself a shared starting point

no working time spent in a classroom

what it left open

what it left open the answer at the moment itself

someone had to know where to look

Practice with feedback, designed on evidence runs alongside all three, and stays necessary

One line, three steps. Practice with feedback is not a step in it, it runs alongside all three and stays necessary.

Good digital learning exists

What changes, and what does not Design whole tasks the way Ten Steps to Complex

The aim is the one Gery wrote down: the right Learning prescribes, and follow the multimedia

knowledge, in the right amount, at the moment of principles (Van Merriënboer, Kirschner &

need. Three things differ. The answer arrives in Frèrejean, 2024; Mayer, 2020). That practice is the

ordinary language, it stays current because one other half of the boundary, not a competitor.

editorial board keeps the canon, and one canon serves individual, team and organization.

Support for one person is where this starts, not where it ends

Classical performance support helps one person finish one task. That is worth having, and in interdependent work it is not enough. A shift rarely fails because one technician cannot find an instruction. It fails because two people hold different versions of the same agreement and neither of them knows it.

A governed canon works at three levels at once. The individual gets the answer at the moment of need. The team gets one answer instead of three, which is what makes coordination possible in the first place. And the organization keeps what one team worked out, because the answer lives in a canon somebody owns and updates rather than in the memory of whoever was on shift.

the individual

the team

the organization

One canon, three levels. The first is performance support as the field has known it; the other two are new.

Two channels, one assistant, one boundary

Access is not a product, it is a property of the work. Two routes deliver it, and the choice between them is technical rather than philosophical. The model behind both is old enough to have been studied: medicine has run a governed, licensed canon for thirty years, and a study across US hospitals finds a small but consistent association with shorter stays and lower risk-adjusted mortality (Isaac, Zheng & Jha, 2012). An association, not a cause.

AI platform (a whole profession)

the knowledge that the work needs

Expertise Amplifier (one theme)

Both routes end in the same place: the assistant people already use, and an answer out of a canon somebody keeps current.

An AI platform

An Expertise Amplifier A composed body of knowledge for a whole

One theme, delivered as a read-only connection profession, extended where needed with the

to an assistant the organization has already organization’s own material. Certification, security,

approved. Light on IT, delivered in weeks, priced management and licences come with it, so it is a

as a licence per year. project of months, delivered by an IT partner.

Four steps, and the first pays for itself

What applies and what has lapsed

01 · Sources Almost every organization finds contradictions and expired documents in this step. That gain arrives with or without AI, which is why it is the first step and not the last.

One file per theme, ordered by the work

02 · Structure What survives step one rarely amounts to a document per question, so the editorial board writes it: one plain file per theme, linked to the files beside it, ordered on questions, tasks and moments rather than on the folder a document happened to sit in. Here an assistant is usable rather than merely available.

Who writes, who checks, when

03 · Stewardship Without that agreement a canon goes stale within a year and the assistant quietly becomes unreliable. Stewardship is a job, not a phase.

Start with the critical tasks

04 · Adoption Begin where it matters most: the tasks that go wrong often, the ones whose consequences are large, the questions people ask every shift. Load the expert knowledge those tasks need, available to everyone around the clock, and keep it current together with the organization. Use follows from usefulness; a launch communication has never made an assistant useful.

An organization that runs these four steps ends up with two things that last: a canon that holds and an editorial board that keeps it current. You replace a platform; you keep a canon.

One editorial board

Read-only

No personal data

Somebody writes while everyone else

The connection reads and never

Nothing about individuals goes into reads, and a log says what changed

writes. What a team records about its

the canon, which keeps the security and when.

own work stays with the team.

review short.

What this changes on Monday

Ask where the question arrives

Split the catalogue in two

Design above the individual

For every request: does this

What may be looked up leaves the

Ask what the team needs to share and knowledge have to be in someone’s

course and becomes access. What

what the organization has to keep. head, or within reach at the moment

has to be there without looking gets

That question is what separates of need? The answer decides the

real practice, and more of it.

access from a help file. intervention.

The two channels, the framework, the author

The two channels

The framework underneath AI platforms: a composed body of knowledge for

Where access meets team and organizational a profession, set up by Tulser and delivered by an

learning, the Learning Conversion Framework IT partner. The Tulser Expertise Amplifier: one

supplies the mechanism:

Learning that crosses theme, delivered by Tulser as an MCP server on

the levels

(Arets, 2026). The four Arets L&D the assistant you already use. tulser.com

Business Models place the same move on one rising line. Both free at tulser.com.

Selected sources

Albarracín, D., Fayaz-Farkhad, B., & Granados Samayoa, J. A. (2024).

Argote, L. (2024). Knowledge transfer within organizations: Determinants of behaviour and their efficacy as targets of

Mechanisms, motivation, and consideration. behavioural change interventions.

Nature Reviews Psychology, 3

(6),

Psychology, 75

psych-022123-105424 Arets, J. (2023, September 7). Is AI the bicycle of the mind?

Arthur, W., Jr., Bennett, W., Jr., Stanush, P. L., & McNelly, T. L. (1998). Performance support powered by artificial intelligence.

eLearning

Factors that influence skill decay and retention: A quantitative Industry

review and analysis.

doi.org/10.1207/s15327043hup1101_3 Arets, J. (2023, September 13). Is AI the bicycle of the mind?

Branch, R. M. (2009).

Instructional design: The ADDIE approach Performance augmentation in corporate eLearning.

eLearning

Learning that crosses the levels: The Learning

technology

. Weingarten Publications. [no DOI available] Conversion Framework

[White paper]. Tulser. [DOI to be added] Gilbert, T. F. (1978).

Human competence: Engineering worthy Arets, J., Jennings, C., & Heijnen, V. (2015).

702010 towards 100%

performance

. McGraw-Hill. [no DOI available] performance

. Sutler Media. [no DOI available]

Cite as

Arets, J. (2026).

The strongest lever is access

[White paper]. Tulser. DOI [to be added]

Also from Tulser:

Learning that crosses the levels

,

The request is not the problem

and

The four Arets L&D Business Models

© 2026 Jos Arets and Tulser. All rights reserved.

You may share this document in unchanged and complete form, with attribution. No adaptation, translation or reuse of parts without written permission.