Adoption with guardrails — built by a security team
Your Team Has AI.Nobody Taught Them How to Use It.
Training, governance, and cost control for companies that bought the licenses and are still waiting for the payoff. We do adoption the way we do security — practical, right-sized, and honest about what it can't do.
40%of desk workers received AI “workslop” last month — output that looks like work and isn't.Two hours to fix. Every time.The cost of untrained AI isn't the license. It's the rework it pushes onto everyone downstream.— BetterUp Labs & Stanford Social Media Lab, n=1,150, September 2025
What You Walk Away With
Artifacts you own — not a slide deck
Every deliverable is built from your work, not from a template. Click any card for detail.
A Prompt & Skill Library From Your Actual Work
Not generic examples — the five things each role does every week
Tested before it goes in
Yours to keep and extend
A Model Routing Policy That Cuts the Bill
Which work goes cheap, which goes expensive
A rule your people can actually remember
Benchmarked on your tasks, not a leaderboard
A Context System That Survives the Session Ending
Project instructions, handoff templates, a save routine
Nobody re-explains the same thing every morning
With a size budget, so it doesn't eat your context
A Skill Catalog With Owners, Ratings, and an Expiry Date
Keeps the library from rotting into 400 near-duplicates
Rating plus usage, so the good one wins the search
Nothing lives forever unreviewed
An AI Policy People Can Actually Follow
Data classification, approved tools, what never goes in a prompt
Enforcement gaps named honestly, in the document
We write these for a living
A Shadow-AI Inventory
What your people are already pasting into what
Usually uncomfortable
Always the most useful page in the readout
By the Numbers
Four numbers that shape how we run these engagements. Every one has a primary source — click for it.
40%
of desk workers got AI “workslop” last month
60%
cheaper to run the same task one model tier down
−24pts
fewer right answers from trained users outside the model's competence (vs. no AI)
93%
of tasks finished by trained users inside it (82% with no AI)
The bill isn't the cost. The cost is the rework nobody invoiced you for.
The Uncomfortable Finding
Training alone didn't catch the wrong answers
758 consultants. Three groups: no AI, AI, and AI plus prompt training.
93%
On work the model was good at, both AI groups did better. The trained group's quality was a little higher than the untrained group's, and both finished more tasks than the group with no AI (93% and 91%, against 82%).
−24 pts
On work the model was bad at, the trained group was wrong 24 points more often than the people with no AI — and lower than untrained AI users too (60% vs. 70.6%, a weaker difference, significant only at the 10% level). Their answers were also rated more coherent and persuasive. Polished, persuasive, wrong.
Why? The authors' reading: over-reliance on AI, a risk reinforced by how persuasive its answers are.
Prompt training on its own didn't make people any better at catching the AI when it was wrong.
So we don't teach prompting alone. Every HyperX engagement teaches the same person how to prompt and how to tell when an answer is out of the model's depth. That isn't a module we added to fill a day. It's the reason the rest of it works.
Dell'Acqua, McFowland, Mollick, Lifshitz-Assaf, Kellogg, Rajendran, Krayer, Candelon & Lakhani. “Navigating the Jagged Technological Frontier.” Organization Science 37(2), 2026. Pre-registered field experiment with Boston Consulting Group.
Everyone sells the training. Almost nobody mentions what it leaves unfixed.
The Offering
Three tracks — choose your depth
Start anywhere. Most companies start with the assessment, then pick a track from what it finds.
1
Prompting Essentials Everyone
Half day · live · up to ~25 people
Get everyone to competent. Mixed roles in one room, working on their own real tasks.
The mental model shift — what this is, and what it isn't
Anatomy of a prompt: context, task, format, constraints. Kills the one-line-question habit.
The trust problem — where it fails, how to verify, what never to trust. This is the module that earns the room. It doesn't get skipped.
Context and continuity — why the model forgets, how to spot it, what to do about it
Iterative refinement — never accept the first draft
Privacy hygiene — what never goes in a prompt, tied to your policy
Hands-on: everyone brings one real task from their actual week and leaves with it done
Takeaways: assessment readout, AUP, routing policy, a live catalog, and a quarterly review cadence.
Right-sizing, out loud: a 30-person company doesn't need a program. It needs Essentials, a policy, and a routing rule. Two weeks, done. We'll tell you that before you buy it.
Every track ends with artifacts built from your work, not ours. The exercises use your real tasks, the skills get built for your team, and the policies name your tools. That's the part a syllabus can't hand you.
The Cost Deliverable
Most of your AI bill is the wrong model doing easy work
Three beats, and then the arithmetic.
The premium is real
The top-tier model costs a multiple of the mid tier. That premium buys real capability — hard reasoning, long complex code, ambiguous judgment.
That isn't what most work is
Summarize this thread. Reformat this into a table. Draft a reply. Classify these tickets. On that work the cheaper tier is indistinguishable.
Nobody routes
Because nobody told them to. The default is whatever the button is set to, and the button is usually set to the most expensive option.
The arithmetic
Published list pricing, August 2026. Ratios, not dollars — vendors re-price.
Tier
Cost vs. top tier
Savings
Top
baseline
—
Mid
40%
60% cheaper
Small
20%
80% cheaper
Move one workload down one tier and you cut its cost 60%. Move it two and you cut it 80%. Nothing about the work changed. Only the routing did.
This is why we quote 40–60% and call it conservative. It isn't our estimate — it's arithmetic on published prices. A team routing everything to the top tier and shifting its volume work down saves more than 60%. We say 40–60% because a real blend keeps the hard work where it belongs.
The routing rule
The actual deliverable is one page your people can remember.
The work
Route to
Why
Summarize, reformat, extract, classify
Small tier
Indistinguishable quality at 20% of the cost. Volume work lives here.
Drafting, analysis, meeting prep, first drafts
Mid tier
The workhorse. Most seats should default here.
Hard reasoning, long code, ambiguous judgment, high-stakes review
Top tier
Worth every cent — for the 10–20% that needs it.
Anything in a loop, on a schedule, or over thousands of records
Cheapest tier that passes the test
Automation is where the wrong default gets expensive fastest.
Honest caveat: savings depend on your current mix. A team already on the mid tier won't see 60%.
The table above is the easy part. Knowing which of your workflows belong in which row takes a benchmark on your actual tasks — not a leaderboard, and not a guess. The assessment tells you your number before you commit to anything.
You won't see it happen. You'll see the mistake it makes ten minutes later.
The Part Nobody Teaches
The model forgets. Your people won't notice.
Every session has a finite window. When it fills, the system doesn't crash — it summarizes older messages away and keeps going.
Summaries are lossy, and the loss is not random. The rule survives. The reason for the rule doesn't. The constraint survives. The exception you carved out doesn't.
What that actually looks like
The half-finished plan
A twelve-step deployment plan, compacted at step seven. The model completes step eight and declares the project finished — dropping verification, rollback prep, and notification. It didn't say it was skipping them. It said “done.” It thought it was.
The TODO that shipped
A config file written with placeholders and a plan to fill them in later. Compaction killed the plan, not the file. The model saw a clean file and committed it. The next service failed at startup parsing the literal placeholder.
The exception that dissolved
“Never edit this directly” became “be careful with system files” in the summary. And “be careful with” is, to a confident model, an invitation to do the thing carefully rather than not at all.
The model doesn't know what it forgot. You can't ask it to check a constraint it no longer remembers having.
Two things go wrong, and they're not the same
This distinction is the whole lesson. Almost nobody has a name for the second one.
Fullness
Depth
What it measures
how close you are to the summarize-and-continue point
how much context you're carrying
The risk
you lose context
the model gets duller — misses details, repeats work, drifts
Nature
Hard. A wall.
Soft. A gradient.
The fix
a deliberate checkpoint, or the recovery playbook
a clean checkpoint, whenever it suits you
Urgency
Now
Not urgent
The window limit isn't the quality limit.
And bigger windows made this worse, not better. On a million-token window you can be deep into the degradation range at 20% full — because 20% of a million is still 200,000 tokens of accumulated context. The bigger window didn't fix the problem. It removed the wall that used to force a reset.
What we teach
Detection
It re-asks something already answered. Suggests an approach you ruled out. States a fact differently than you remember stating it. The “wait, why is it suggesting that?” moment. When you spot one: stop. Treat the last ten minutes as suspect.
Recovery
Stop and verify. Re-read the source docs — don't paraphrase. Check against ground truth, because the file on disk beats the summary in the model's head. Re-state the constraints, since they're exactly what got softened. Resume from a known-good point, not from where the model thinks you were.
Prevention
Scope work to finish before the wall. Check before you start something big. And keep your always-loaded context files small — every line in them is paid for in every session, forever.
The gauge
Two indicators. That's the whole tool.
Hard · allowed to be loud
% Until Compact
How close you are to the wall — measured against the point where summarizing actually kicks in, not against the size of the window. This is the one with an alarm on it.
Soft · a nudge, never an alarm
Context Depth
How much you're carrying. Not urgent, ever — just a signal that a clean checkpoint would sharpen things up.
🟢 Sharp🟡 Warming🟠 Drifting🔴 Deep
And one thing we'll tell you that nobody else will
We built that depth meter with a red danger threshold. Then we measured it against real session history, looking for the point where quality falls off.
There wasn't one. The error rate came out flat.
So we softened the meter so it never reaches alarm red, and we put a card in the tool that says “measured from your history — no knee.” It's a research-informed heuristic, deliberately gentle, and we label it that way. We could have kept the scarier number. It would have sold better.
We built a warning light, checked our own data, found the danger wasn't where we'd drawn it, and shipped the correction.
That's the same instinct behind everything else on this page.
The thresholds are the easy part. Knowing which of your workflows actually run long enough to hit them — and which of your people are quietly working through it every day — is what the assessment finds.
The Second-Year Problem
400 skills, no catalog, and nobody trusts any of them
Year one, everyone's thrilled. People build custom skills, agents, saved prompts. Year two there are hundreds. Twelve do nearly the same thing. Four are subtly broken. The person who wrote the good one left. Nobody can find anything, so everyone writes their own — which makes it worse.
This isn't new. It's policy sprawl, script sprawl, and the shared drive with nine versions of the same spreadsheet. We've fixed that before. The fix is governance, not enthusiasm.
Sprawl isn't too many skills. It's too many skills of unknown quality.
What we stand up
A catalog with a required shape
Name, one-line description, owner, what it's for, what it's not for, last-reviewed date, worked example. No metadata, no entry.
Ratings and usage — let the org sort it
A rating after use, a usage count (the honest signal — what people run beats what people praise), and a one-click worked / didn't work. Search defaults to a blend, so the good one wins.
Tiers, so trust is visible at a glance
🥇 Golden — reviewed, owned, tested. Use this one. ✅ Community — someone's working skill, ratings visible. 🧪 Experimental — no promises. 🗄️ Archived — superseded, hidden from search. Plus a promotion path with a named reviewer, so quality is a process rather than an opinion.
An expiry date on everything
Unreviewed two quarters → flagged. Unused two quarters → archived. Sprawl is what happens when nothing is allowed to die.
Duplicate detection at intake
Before you publish, the catalog shows the three closest existing skills and asks whether you'd rather improve one. Most sprawl is honest duplication by people who couldn't find the original.
It lives where they already work
A repo, a SharePoint list, a Notion database, a board. Not a twelfth destination nobody opens.
Without a catalog
With one
400 skills, unknown quality
40 golden, the rest tiered and searchable
Everyone rebuilds the same thing
Duplicate check at intake
Author left, skill rots
Every skill has a named owner
“Which one works?”
Rating + usage answers it at a glance
Nothing is retired
Expiry and auto-archive
The catalog design isn't the hard part. The first pass is — going through what you already have and deciding which ones are golden, which are duplicates, and who owns each survivor. That's a week of somebody's judgment, and it's the week that makes the rest work.
The class is the easy part. Making it survive contact with a Tuesday is the work.
How We Work
Four steps, and you can stop after the first
1
Assess
1–2 weeks · scoped and priced up front
What's in use, by whom, on what data, at what cost, under what policy.
Shadow-AI discovery — including what you didn't approve
Current model mix and spend
Your prompt and skill inventory, and its real condition
Policy and control gaps, named honestly
It stands alone. You can stop here with something useful and decide later.
2
Teach
1–3 days
Essentials for everyone. The Builder track for the people who'll carry it after we leave.
Both use your real tasks. Nobody sits through a worked example about a fictional company.
3
Build
2–4 weeks
Library built from your real work. Catalog stood up. Routing policy written and tested. Policy finalized.
This is where the artifacts get made — and where the class turns into something that outlives it.
4
Sustain
quarterly
Review the catalog. Re-check the bill. Retrain the drift out. Add the next department.
Routing decays. Catalogs rot. Habits slip. A quarterly touch is what stops the whole thing quietly reverting.
Start with the assessment.
Two weeks. Scoped and priced up front, no commitment past it. You'll get:
What AI your people are actually using — including what you didn't approve
What you're spending, and what routing would save you
Your prompt and skill inventory, and its real condition
A right-sized plan: what you need now, what can wait, what you don't need at all
Everything on this page is public. The methods are in our articles. The tools are on GitHub. What we sell is knowing which of it applies to you — and that part starts with looking at what you actually have.
/adversarial-review
Independent reviewers attack a deliverable before it ships, because same-session self-review is blind to its own gaps.
/sanitize-review
Catches sensitive detail before content leaves the machine — including the dangerous case: harmless facts that together identify one organization.
/save-session/save-full
Structured session records that survive a context reset and hand off between people intact.
/prune
Shrinks a bloated always-loaded context file by relocating what drifted into it — measured in bytes, because line counts hide the problem.
A prompt and skill library built from your actual work
Deliverable — Essentials and Builder tracks
Why generic libraries fail
Generic prompt libraries don't get used, because the examples aren't anyone's job. People open them once, don't see their own work, and never come back.
We build yours from real tasks pulled out of real weeks — the five things each role does every week, in their words.
Every entry gets tested before it goes in. That's the step everyone skips, and the reason most libraries rot inside a year.
“If nobody recognizes their own job in the library, you built a document, not a tool.”
40–60%
typical reduction in AI spend from routing alone
Arithmetic on published list pricing, August 2026
How it works
Most AI spend is the wrong model doing easy work. Moving a workload one tier down cuts its cost 60%. Two tiers, 80%.
Nothing about the work changes. Only the routing does.
We benchmark on your actual tasks, not a public leaderboard — then write the default per role, so it isn't left to each person's habit.
Savings depend on your current mix. A team already defaulting to the mid tier won't see 60%. The assessment tells you your number first.
“Nobody routes, because nobody told them to.”
Every day
Your people pay the five-minute re-explaining tax without noticing
Deliverable — context system
What we build
Every session starts from zero. The fix isn't better prompting — it's a file. Project instructions that load automatically, so nobody re-explains the basics.
But the moment you build that file, you inherit a second problem: it loads into every session, forever, and it grows. So we set a size budget and a health check that measures the right thing.
A handoff template per project — current state, goal, key files, what works, what's open. It works whether the next session is a colleague or the same person after a reset.
And one rule that prevents most of the rot: each fact gets exactly one home.
“A context file that grows unchecked is a file that makes every session worse, forever.”
Year two
is when the skill library becomes the problem it was meant to solve
Deliverable — skill catalog
What the catalog enforces
A required shape. Name, description, owner, what it's for, what it's not for, last-reviewed date, worked example. No metadata, no entry.
Rating plus usage. Usage is the honest signal — what people actually run beats what people say is good. Search sorts on a blend, so the good one wins.
Visible trust tiers — Golden, Community, Experimental, Archived — with a promotion path and a named reviewer.
An expiry date on everything. Sprawl is what happens when nothing is allowed to die.
“Sprawl isn't too many skills. It's too many skills of unknown quality.”
Honest
A policy that names its own enforcement gaps
Deliverable — AI Acceptable Use Policy
What's in it
Data classification, approved tools, and a plain-language list of what never goes into a prompt.
Written to be followed — short enough to read, specific enough to act on. Not a fifty-page document nobody opens twice.
The part that makes ours different: where the policy says “must” and no technical control exists, we say so, in the document.
We write security policy for a living. The AI one is the same discipline applied to a newer problem.
“A policy that pretends to be enforced is worse than one that admits what it can't reach.”
Already
The question was never whether to adopt AI. Your people adopted it.
Deliverable — shadow-AI inventory
What the inventory finds
Which tools are in use, by which teams, on what kind of data — including the accounts nobody approved and nobody is paying for.
Personal accounts holding work product. This is the one that surprises people, and it's the one with real exposure attached.
Where consolidation onto approved tooling is easy, and where it will meet resistance because the unapproved tool is genuinely better.
It is usually uncomfortable. It is always the most useful page in the readout.
“The only open question is whether it's happening on tools you can see, under rules you wrote.”
$9M
a year in lost productivity, for a 10,000-person company
BetterUp Labs & Stanford Social Media Lab, n=1,150 US desk workers, September 2025 (Harvard Business Review)
The numbers
40% of desk workers received AI “workslop” in the past month — content that looks plausibly polished but offloads the real thinking onto whoever receives it.
Two hours to resolve each incident, on average.
$186 per month, per affected employee — an invisible tax that never appears on an invoice.
The researchers named the pattern, which is half the battle. You can't govern a failure mode nobody has a word for.
“The cost of untrained AI isn't the license. It's the rework it pushes onto everyone downstream.”
80%
cheaper if the work moves two tiers instead of one
Published list pricing, August 2026 — ratios, not dollars
The arithmetic
The mid tier runs at 40% of top-tier cost. The small tier at 20%.
On summarizing, reformatting, extraction and classification, output quality is indistinguishable between tiers. That's most of the volume.
This is why the 40–60% headline is conservative. It isn't our estimate — it's division.
We quote ratios rather than prices deliberately. Vendors re-price; the shape of the ladder holds.
“Nothing about the work changed. Only the routing did.”
−24 pts
the trained group's correctness penalty on work outside the model's competence
Dell'Acqua et al., Organization Science 37(2), 2026 — pre-registered field experiment, n=758 (373 on this task)
What the experiment found
Consultants were randomized into three arms: no AI, AI, and AI plus a prompt-engineering overview.
On a task chosen to sit outside the model's capability, the control group was correct 84.5% of the time. The AI groups dropped — untrained by about 14 points (70.6%), trained by about 24 (60%).
Their answers, right or wrong, were also rated more coherent and persuasive: up 25.1% over the no-AI group for the trained arm, 17.9% for the untrained. Polished, persuasive, wrong.
The authors' interpretation: over-reliance on AI (“falling asleep at the wheel”), a risk they say persuasive AI answers reinforce.
“This is why verification has to be taught alongside prompting — not after it, and not optionally.”
93%
task completion for the trained group — against 82% for the group with no AI
Dell'Acqua et al., Organization Science 37(2), 2026
Training works — inside the boundary
On work within the model's competence, both AI groups produced significantly better work than the group with no AI, and the trained group did a little better than the untrained one.
On quality, the paper's own conclusion: the difference between trained and untrained was statistically significant at or near the 5% level in every specification, which “underscores the importance of the added overview.”
So the training helps on quality. It just isn't a free pass — which is the entire argument for pairing it with governance.
“Training is necessary. On its own, it is not sufficient — and the gap is where the damage happens.”
