A backlog grows when demand outpaces capability. Cutting hours with AI clears some of it. It does nothing about the reason it keeps coming back. The better question for any AI plan is not “how much work can agents absorb?” It is “what should each person be able to contribute next year that they cannot today?”
That shift matters for two backlogs. One is the work queue. The other is the training request queue, where every gap becomes a course to build. Both fill faster when organizations skip the diagnosis. Before you build that course, or automate that task, decide what capability the work actually requires.
Picture a junior analyst, eighteen months in. Her first year was unglamorous: reconciling reports, cleaning data, drafting the first version of a variance note, getting corrected by a senior colleague. Tedious work. Also the work where she learned what “wrong” looks like.
This year, agents do all of it. Faster. Cheaper. Cleaner. The dashboard shows hours saved. Her manager is pleased. Her utilization is down, so someone proposes a bigger portfolio of accounts for her to oversee.
What nobody asked: where does she now build the judgment to spot the number that looks right and is not? She reviews outputs she never learned to produce. She approves what she cannot yet challenge.
This is a composite scenario, not a customer story. Any team that has automated entry-level tasks will recognize it.
Most AI business cases are written in hours. Same work, fewer hours. That is a legitimate gain. It is also a plan that treats people as a cost line.
The risk is not hypothetical in the research on automation. A 2025 mixed-method review of AI-induced deskilling in medicine separates two threats. Deskilling is the degradation of competence from reduced practice. Upskilling inhibition is the loss of chances to build new or advanced skill, especially among trainees, through over-reliance on decision support. The evidence there comes from clinical settings, so treat it as a warning about mechanism, not a forecast for your finance or operations team. The mechanism, though, travels: when the practice disappears, so does the learning.
For an enterprise, the consequence shows up late:
Speed now. Capability lost later. That is the trade hidden inside a pure efficiency plan.
The alternative is not to refuse automation. It is to make two decisions together instead of one.
If you make only the first decision, the freed time gets absorbed by the backlog or removed from the budget. If you make both, automation becomes the funding mechanism for growth.
The design question is simple to say and hard to answer well: what should this person be able to contribute next year that they cannot today? Answering it requires knowing what the work demands, which is a diagnosis problem, not a content problem.
Take the junior analyst again. A deliberate path has four stages. Each stage is a change in what she is accountable for, not a course she completes.
| Stage | What she does | What agents do | What proves the stage |
|---|---|---|---|
| Execute | Produces a deliverable with an agent alongside, and learns the shape of a correct one | Drafts, flags anomalies | Her work needs fewer corrections on the cases that matter |
| Judge | Reviews agent output, explains what is wrong and why | Produces output, including planted and real errors | She catches and explains problems before a senior reviewer does |
| Accept | Decides whether output is good enough to release, and states the risk she is taking | Provides supporting analysis on request | Her release decisions hold up in quality review |
| Own | Is accountable for the business result of the analysis, including changes to how agents are used | Operates inside guardrails she helped set | The business outcome she owns moves, and she can show how |
Two details make this work.
First, “judge” cannot be skipped. People do not leap from watching an agent to owning an outcome. They need practice making calls on realistic cases and getting feedback on their reasoning.
Second, the path belongs to the work, not to a training catalog. You define what each stage requires from the real work, then decide how people get there. Sometimes that is a course. Often it is a change to how work is assigned, who reviews what, or what the manager asks in the weekly check-in.
This is where the backlog metaphor bites. Learning teams are drowning in requests: build a course on the new agent tool, a module on prompt practice, a refresher on data quality. Each looks reasonable. Together they become a queue nobody can clear, and completing the queue proves nothing.
The U.S. Office of Personnel Management describes a needs assessment as a way to identify gaps between required and current performance, clarify the critical behaviors and why they are absent, and consider integrated solutions. Training is chosen only if it is the best answer. A July 2026 GAO guide on human capital likewise tells agencies to assess the skills and competencies they need to meet goals and to identify workforce gaps.
In plain buyer language, diagnosis asks four questions:
If you are mapping gaps for a function that is adopting AI agents, a practical look at onboarding automation and capability gaps shows how workflow changes surface the skills teams need next. For our own approach to this step, see our skill gap analysis playbook for HR leaders.
You will not build everything. A cyclical model of enterprise capability from a University of Minnesota paper integrates leadership, assessment, prioritization, learning objectives, measurement and design as a continuous loop rather than a one-time project. The useful lesson is that prioritization sits before design.
A workable filter for each request:
Requests that fail the first two do not enter the queue. That alone shrinks most backlogs.
Once you know the capability is a human one, and that it involves judgment, the design changes. Knowledge recall is cheap to test and cheap to forget. Judgment needs rehearsal.
The evidence on how to do that is consistent:
For the junior analyst, this means a set of realistic review cases where the agent is wrong in different ways, a debrief on how she reasoned, and a harder case next. That is the “judge” stage made concrete. Our guide to scenario-based training design covers the build in detail.
We get one request more than any other: build us a course. We understand why. A course is visible, schedulable and easy to report on.
But we start earlier. Before anyone builds, we want to know what capability the work requires, whether learning is the right enablement play, and what the organization’s own evidence says. That evidence is strategy, frontline friction, quality findings and the expertise of the people who do the work well. It is never a template and never a generic content library.
AI changes the stakes. An agent can generate a course in minutes, which means the backlog of “things we could build” becomes infinite. Speed of production is not the constraint anymore. Knowing what deserves to exist is.
So we hold a simple position. AI should be governed, grounded in your evidence, and pointed at growth. Use it to take on the routine. Use the capacity it frees to move people into harder judgment, with practice that mirrors real decisions, and measure the result against business outcomes rather than completions.
Completing a course is not the same as being able to do the work. We build for the second.
— Brian
Using AI to design and run development carries its own risks. The Association for Talent Development’s AI governance framework points to practical steps: align AI use with organizational goals and measure its impact, inventory tools and data flows, assess privacy, bias and fairness risks, establish human oversight and employee recourse, and audit regularly while collecting employee feedback.
Two habits keep that manageable. Keep a named human accountable for every design decision an AI assists. And keep the content current: capability programs drift as tools and processes change, so treat them like living assets. A discussion of version control for learning content shows how software teams handle that continuity.
The closing test for any AI plan: measure what your people can now contribute, not just what AI saved.
Hours saved is a fine number. It is also a number every competitor with the same tools can report. What differentiates you is the answer to a different question: what decisions are people trusted with this year that they were not trusted with last year?
The GAO guide recommends that leaders develop performance indicators and use valid, reliable data to assess how training contributes to goals, with benefits that can include productivity, customer satisfaction, quality or lower cost. It also says programs should build in measures of effectiveness at the design stage. Do that early. A measure added after launch is usually completion.
Practical indicators of contribution growth, defined with the business owner before the work starts:
Pick the few that match the role. Set the baseline first. Our view on why completion-style models fall short is in does the Kirkpatrick model measure capability.
Finance leaders will ask for the arithmetic. Efficiency cuts cost once: the saving arrives, then it is in the baseline. Capability compounds. Each person who can take on a larger contribution raises what the organization can deliver without a proportional rise in headcount. The GAO guide asks leaders to identify training costs and associated benefits and to link learning objectives to organizational results. Frame the AI investment the same way: cost saved plus capacity created, each tied to an owner.
We built the Cognistry Platform for this sequence: diagnose the capability the work requires, ground the design in your own evidence, structure the response, and measure back to the business. Signal supports the diagnosis. Forge structures the learning architecture. Sim gives people decision practice in realistic situations. If you are weighing us against an LMS, start with LMS vs capability platform. If you want to see the approach applied to your own roles, start from our overview.
No. The risk appears when automation removes the practice where people build judgment and nothing replaces it. Research on deskilling comes largely from clinical settings, so apply it as a design caution. Replace the lost practice with deliberate scenarios and review responsibilities.
Run a cause analysis before any build. Ask whether the gap sits with the person or with the environment: tools, process, ownership, incentives. OPM’s guidance frames training as the choice only when it is the best solution after that assessment.
Measure changes in the work: quality findings, escalation patterns, time to independent work and business results someone owns. Define them with the business owner at design time, and capture a baseline before launch.
Faster production fills the backlog faster. Diagnosis-first work shrinks it by removing requests that learning cannot solve and focusing practice on the judgment the work demands.