Today's AI Specialist: The Onboarding Placement Agent. The Agent That Decides Where to Drop You Into the Programme.
Today's AI Specialist: The Onboarding Placement Agent. The Agent That Decides Where to Drop You Into the Programme.
The first session matters more than any other session. The reason is simple: a learner who is placed wrong in the first session disengages before the second. The platform never gets a chance to fix the placement. The first session has to land.
The Onboarding Placement Agent is the agent that runs the first session. She has the 60-second speaking sample the Snapshot Fluency Analyst produces, the full intake form the learner filled in, and any uploaded materials. From that, she has to decide where to drop the learner into the programme.
Wrong placement costs months. Right placement costs nothing.
This is the build story of the Onboarding Placement Agent. Why placement is a different job from scoring, what the agent sees beyond the audio, and the two decisions that decide whether the first session lands.
The problem the agent solves
Most language platforms place learners by CEFR level alone. You take a test, you get a level, you go into the module for that level. The naive approach works for a textbook-style curriculum, where each module is a sequenced set of lessons that build on each other.
It does not work for a coaching platform. The level is necessary; it is not sufficient. A B2 marketing director who needs English for a board presentation in three weeks needs a fundamentally different first month from a B2 software engineer who has been working in English for years and wants to move to C1. Both are B2. Both need different first sessions, different practice priorities, different Sophie session calibrations.
The Onboarding Placement Agent produces a richer object than a level. She produces a placement, which is a tuple of: starting programme module, difficulty calibration for Sophie's first sessions, practice priorities for the first month, and goal scaffolding for the first quarter. The level is one input to the placement. The intake form is another. The agent's job is to assemble all of it into a starting point that fits the specific learner.
What the agent sees
The placement is made against four sources of input.
The first is the 60-second speaking sample. The Snapshot Fluency Analyst produces the CEFR level and four sub-scores. The Placement Agent reads those.
The second is the intake form. The learner's L1, their stated reason for learning English, the specific work scenarios they have flagged ("I have an interview in three weeks", "I freeze in board meetings", "I want to lead conference calls"), the time budget they have signalled ("an hour a day", "fifteen minutes a day", "two hours on weekends"), and the level they self-assessed at on a simple slider before the audio sample. The intake form is dense; most learners fill it out properly because they have just decided to commit and are motivated to be accurate.
The third is any uploaded materials. About 30% of new learners upload something: a CV, a recent email, a slide they will present from, an interview prep document. The agent reads these and uses them to calibrate. A CV at a senior level signals the learner needs senior-register practice. An email full of L1 transfer errors signals the practice priorities should weight grammar more heavily. An interview prep doc signals that simulated-interview Sophie sessions should be in the first month.
The fourth, only available for learners who came through the diagnostic-chat funnel, is the transcript of the diagnostic chat. The diagnostic chat captured a struggle moment and a desired outcome in the learner's own words. The agent uses these to anchor the practice priorities directly to what the learner came for.
The agent has three minutes of cognitive load to assemble everything into a placement. The placement comes back to the learner inside the first session.
Decision one: deliberate conservatism
The single most important design decision in the agent is the bias toward conservative placement. When the four inputs disagree about where the learner should start, the agent places at the lower of the candidate options.
The reasoning is asymmetric cost. Under-placement causes mild boredom. The learner finds the first session too easy, comes back the next day, adjusts up, and is fine. Over-placement causes overwhelm. The learner finds the first session too hard, feels they are not at the level they hoped, and disengages. The disengagement is the failure mode. The boredom is not.
The conservatism shows up in roughly 12% of placements. The agent identifies four learners who could plausibly start at C1 and instead starts them at B2+. One of them adjusts up to C1 on day one. One adjusts up within two weeks. Two stay at B2+ comfortably. None disengage.
If we ran the alternative (placing aggressively, at the highest plausible level), the failure mode would be more disengagement on day one, fewer adjustments, and lower six-month retention. The data on this is unambiguous, even if individual learners sometimes protest the conservatism on the first day.
Decision two: placement is offered, not assigned
The second decision is the framing of the placement to the learner.
The naive framing is "you are at level X, your programme starts here." Authoritative. Algorithmic. The learner accepts it because the system said so. We tested it. Engagement was lower than it should have been, and we eventually traced it to the framing: learners receiving an algorithmically-assigned placement had less ownership of the programme than learners who had chosen something themselves.
The fix was to reframe the placement as a recommendation, not an assignment. The agent now produces a placement with the framing "this is where we recommend starting; you can adjust up or down one module if you disagree." The recommendation is the same. The override is a real option, not a hidden switch.
About 18% of learners adjust on the first day. Most adjust upward: the conservatism surfaces here as the bias most likely to be overridden. About a third of those upward adjustments are themselves adjusted back down within two weeks, when the learner discovers the recommended level was the right one. The override is what lets the agent stay conservative without trapping confident learners.
The interesting finding is that engagement is higher across the board with the override available, even though most learners do not use it. The perception of agency matters more than the exercise of it. Allowing the override is what makes the recommendation feel like a recommendation rather than a verdict.
TL;DR
The Onboarding Placement Agent runs the first session: the one that decides where the learner gets dropped into the programme. Placement is a different job from scoring: it produces a starting module plus difficulty calibration plus practice priorities plus goal scaffolding, not just a CEFR level. The agent reads the 60-second speaking sample, the intake form, any uploaded materials, and the diagnostic chat transcript when available. Two design decisions matter. One: deliberate conservatism. The agent places at the lower of plausible candidate levels, because under-placement causes mild boredom (cheap to fix) and over-placement causes overwhelm and disengagement (expensive to fix). Two: placement is offered as a recommendation with a clear override path, not assigned algorithmically. Engagement is higher across the board when the override is available, even when most learners do not use it. The perception of agency matters more than the exercise of it.
See how the Onboarding Placement Agent was built and meet the rest of the team (/build)
Learning Materials
Key Vocabulary
placement
The act or result of putting someone into a particular position, level, or role within a system.
“The Onboarding Placement Agent decides the learner's placement in the programme.”
tuple
A fixed, ordered group of related elements treated as a single object — often used in data and systems contexts.
“The placement is a tuple of starting module, difficulty calibration, practice priorities, and goal scaffolding.”
calibrate
To carefully adjust something so it operates accurately for a specific situation.
“The agent calibrates Sophie's first sessions to the learner's actual level.”
priorities
The things ranked as most important to deal with first.
“The agent sets practice priorities for the learner's first month.”
scaffolding
A temporary supporting structure — in learning, the framework of intermediate goals that supports the learner until they can stand alone.
“Goal scaffolding for the first quarter is part of the placement object.”
conservative
Deliberately cautious; choosing the safer, lower-risk option when uncertain.
“The agent makes conservative placements when the four inputs disagree.”
overwhelm
The feeling of being unable to cope because the demands are too great.
“Over-placement causes overwhelm; under-placement causes mild boredom.”
disengage
To withdraw mental or emotional involvement from an activity or commitment.
“Learners placed too high disengage before the second session.”
boredom
The feeling of being uninterested because the activity is not stimulating enough.
“Mild boredom from under-placement is cheap to fix.”
authoritative
Speaking or acting with confident, expert-sounding authority — sometimes in a way that closes off discussion.
“The naive framing of placement is authoritative and algorithmic.”
assign
To give a person a particular task, role, or position by decision, often without choice. Contrast 'offer', which leaves the recipient free to accept or adjust.
“The placement is offered, not assigned; the learner can adjust up or down one module.”
agency
The sense of being able to act and make choices that affect one's own situation.
“The perception of agency matters more than the exercise of it.”
perceived
Understood or interpreted as being a certain way, regardless of whether that interpretation is accurate.
“Perceived agency matters even when the override is rarely used.”
uplift
An increase or improvement in a metric, especially one attributable to a change.
“The engagement uplift after introducing the override path was material.”
refresh
To update, renew, or run again — for example, to redo a check on a quarterly basis.
“The placement model is refreshed quarterly against new retention data.”
Grammar Notes
Asymmetric-cost framing ('X is cheap to fix, Y is expensive to fix')
This construction sets two outcomes side by side and assigns each a cost descriptor. The structure is: '[Cause A] causes [outcome A] ([cost A]), and [cause B] causes [outcome B] ([cost B]).' It is the natural English way to justify an asymmetric design choice — leaning toward the lower-cost failure mode. The parenthetical cost phrase ('cheap to fix', 'expensive to fix') compresses the reasoning into the same sentence, so the reader does not need a separate paragraph of justification.
“'Under-placement causes mild boredom (cheap to fix) and over-placement causes overwhelm and disengagement (expensive to fix).'”
Common mistake: Treating both failure modes as symmetric ('both are problems') flattens the argument. The whole point of the framing is that the costs are not equal, and the design leans deliberately toward the cheaper failure.
The 'X is offered, not Y' reframing pattern
This is a contrastive negation pattern: a positive claim immediately followed by a negation of its near opposite, separated by a comma. The structure is: '[Subject] is [the chosen verb], not [the rejected verb].' It is used in English to reframe how a thing should be understood by ruling out the wrong interpretation in the same breath as asserting the right one. Common in design and product writing where the same action could be misread.
“'The placement is offered, not assigned.' / 'Placement is offered as a recommendation with a clear override path, not assigned algorithmically.'”
Common mistake: Inverting the order ('not assigned, but offered') is grammatical but rhetorically weaker — the positive claim should lead, with the negation correcting the likely misreading.
Statistical descriptors and percentage hedging language
English uses hedging quantifiers — 'about', 'approximately', 'roughly', 'around', 'somewhere near' — before a number to signal that the figure is a representative estimate rather than a precise count. The hedge is essential in operational and product writing: it tells the reader the number is real and tracked, but not a guarantee. The hedge can also be expressed as a fraction ('about a third') or a range ('between 10 and 15%').
“'About 18% of learners adjust on the first day.' / 'The conservatism shows up in roughly 12% of placements.' / 'About a third of those upward adjustments are themselves adjusted back down within two weeks.'”
Common mistake: Dropping the hedge ('18% of learners adjust') reads as a precise audited figure and invites pushback if any single week deviates. Over-hedging ('possibly around roughly 18% or so') sounds uncertain and undermines authority. One hedge per number is the right calibration.
Comprehension Questions
- 1.Why does the post argue that placement is a different job from scoring?
- 2.What four sources of input does the Onboarding Placement Agent read before deciding a placement?
- 3.Why is the placement agent deliberately conservative rather than aggressive?
- 4.Why does engagement rise across the board when the override path is available, even though most learners do not use it?
- 5.Think of a time you were placed into something — a course, a level, a team, a role. Was the placement conservative (slightly below where you could have started) or aggressive (slightly above)? What happened next, and what would you have done differently if you were designing the placement?
Run your own diagnostic
Use the same Strategic Council I run my own decisions through. The assessment preview is free. The specific central human intelligence it is based on is verified in person during the call.
Start the free diagnostic →