By Dr. Ravinder Tulsiani
The budget review six months in
Six months after an AI rollout, an executive team sits down to review the investment. Tool usage is up. Some analysts use AI daily. Others have barely changed how they work. Managers cannot say with confidence whether an AI prepared report can be trusted without rework. Executives ask a direct question: why has productivity not moved?
The natural response is more training.
But the analysts already know how to use the tool. What the organization has not decided is how their work should change, which decisions they remain accountable for, and how their work will be checked.
This article is for two readers in that review. If you are an executive approving an AI training budget, your decision is what work changes and what people remain accountable for. If you lead learning, your decision is what capability is therefore worth building, and what evidence will show it is being used at work. It applies where AI tools are already available, adoption is uneven, and a large training investment is being considered before work, accountability and measures have changed.
The insight is direct: teaching people to use AI is the easier part. Deciding how people and AI should work together is the leadership decision that determines whether training is worth buying at all.
Decide what work changes before you decide what to learn
Take an illustrative analyst team. Its work includes collecting information, preparing reports, identifying patterns, and advising leaders.
AI can increasingly support the first three activities. The analyst’s value shifts to interpreting findings, evaluating uncertainty, challenging assumptions, and helping leaders make a decision. That shift is the point. If the team is simply trained on the tool, without reconsidering its responsibilities, it is being prepared to perform tasks that technology is already changing.
This is workforce strategy, not a training needs analysis. A training needs analysis starts from existing jobs and asks what skills are missing. A workforce strategy starts from redesigned work and asks what people must still be able to do.
The benefit of starting with work is concrete: you stop training people for outputs AI will generate, and you start training them for the judgment that makes those outputs useful. The cost of skipping this step is also concrete: a team trained on time, a budget spent, and a job description that no longer describes the work.
This does not mean every job changes at the same pace. Where work is stable, existing job requirements may still be a sound basis for planning. The check applies where AI is already changing the tasks, which is the condition this article addresses.
Stop rewarding usage. Start checking decision quality.
AI adoption is often measured by tool access, usage frequency, training completion, and employee confidence. Those measures show whether people are engaging with the technology. They show much less about whether performance has improved.
Return to the analyst team. One analyst uses AI for every report and produces work that needs extensive correction. Another uses AI selectively and produces a clearer recommendation that a leader can act on. On a usage dashboard, the first analyst looks more successful. In business terms, the second analyst has contributed more.
The payoff of changing the measure is direct. If you reward usage alone, you will encourage activity with little connection to outcomes. If you check decision quality, you will see where AI helps, where it creates rework, and where a person must step in.
Check three things at work, not on a dashboard:
- Whether the output is accurate enough to act on without extensive correction
- Whether the decision it supports is better, faster, or less risky than before
- Whether risk has increased, for example through an error that no one caught
If you cannot answer those questions from live work, you do not yet have evidence of value, however high usage is. That gap is useful. It tells you what to measure next, instead of letting you mistake activity for progress.
Preserve judgment on purpose, not by accident
One of the most consequential questions in AI adoption is what people should continue to do, even when technology can perform a similar task. This is not only a technical question. It involves accountability, trust, professional judgment, values and the consequences of a wrong decision.
For the analyst team, the question is sharp. AI can summarize feedback, identify patterns, and prepare a draft recommendation. Should it determine a forecast, recommend a course of action, or close a decision without meaningful human review? The answer depends on more than what the tool can do. It depends on who is accountable if the recommendation is wrong.
Human capability can fade when it is no longer exercised. An analyst who routinely accepts AI summaries may lose the ability to evaluate them. A person cannot provide meaningful oversight of a decision they no longer understand well enough to challenge.
Use a preservation test before you automate a judgment, and repeat it for any decision with serious consequences:
- If AI is wrong, who is accountable, and can that person spot the error?
- What work must that person still do themselves to keep that judgment sharp?
- How will you check that judgment, and how often?
If any answer is blank, do not treat that judgment as covered. Either keep the judgment with a person and build practice into the work, or change the work so the check is real. For the analyst, that could mean reviewing a sample of source data, challenging the assumptions behind a draft, and signing off on the recommendation, rather than forwarding an AI summary.
Naming capabilities is not enough. Critical thinking, professional judgment, and the ability to recognize errors matter only where they are required, practiced and checked. The test above turns that principle into a decision.
Change the conditions, not just the skills
Imagine the analyst team is given approved tools, practical workshops, and ongoing learning resources. The analysts complete the training and understand the technology.
Yet their daily responsibilities, approval processes, performance expectations, and reporting requirements remain unchanged. They are still evaluated on the same activities, using the same processes, within the same constraints.
Under those conditions, adoption will depend largely on individual initiative. Some analysts will experiment. Others will return to familiar practices. The organization may read inconsistent adoption as resistance or as a skills gap, when the real issue is that it has not redesigned the conditions in which analysts work.
That is why AI transformation needs coordination across technology, operations, human resources, risk and business leadership. Learning has an important role, but it cannot by itself resolve workflow design, decision authority, incentives, governance, or accountability. Those decisions need to be made alongside workforce development, not after training has already been delivered.
The benefit of making those decisions first is that training lands in work that is ready for it. The cost of making them later is a second round of training, after the first round has already taught people to work around unchanged processes.
A five step sequence before you commission training
Do not commission a large AI training initiative on the basis of a general case for adoption. Work through these five steps, in order. Each step names an owner, the evidence to look for, and the rule that stops you guessing.
1. Name the outcome. Owner: executive sponsor.
Write the specific business outcome AI is expected to improve, and how you will know it has improved. The benefit is that you have something to judge training against, instead of judging it by usage. The rule is simple: no named outcome means no large training commission yet.
2. Name the work change. Owner: business leader, with operations.
List which analyst activities AI will do, which it will assist with, and which will remain with a person. The benefit is that training can be aimed at the new work, not the old job. The rule is simple: if no activity changes, what you are buying is a tool demonstration, not a workforce strategy.
3. Name accountability. Owner: executive sponsor, with risk.
List the decisions a person remains accountable for, and who checks each one. The benefit is that consequential decisions do not drift to the tool by default. The rule is simple: no named checker means no rollout for consequential decisions yet.
4. Name the capability in the redesigned work. Owner: learning leader, with the business leader.
State what a person must be able to do at work, as shown in live work, not in a self report or a test alone. The benefit is that you train only capability that will be used and checked. The rule is simple: if a capability will not be used at work, do not buy training for it.
5. Name the conditions that must change. Owner: operations, with human resources.
State the workflow, approval, measure and incentive changes that make the new capability useful. The benefit is that people are not trained for work the system still punishes. The rule is simple: if measures still reward the old way of working, fix the measures first, or expect training to be ignored.
If any step is blank, that step is not done yet. That blank tells you what to decide next. It also tells you who needs to decide it, so the conversation does not fall back to learning alone.
What this sequence avoids
The return on this sequence is not a percentage claim about training. It is the cost avoided by not buying the wrong capability.
It avoids training analysts to produce outputs AI can generate. It avoids paying for high usage that creates rework. It avoids losing the judgment needed to challenge AI when it is wrong. And it avoids a second training cycle when the real barrier was an unchanged process, measure, or approval.
Workforce investment always has an opportunity cost. Time spent building one capability is time not spent building another. When technology changes how work is distributed between people and machines, that cost matters more, because a capability that was valuable last year may not be valuable in the redesigned work.
Organizations do not need to predict the future perfectly. They do need to avoid treating today’s job requirements as permanent where work is already changing, and they need to be explicit about the uncertainty that remains.
The organizations that benefit most from AI will not necessarily be those that train the largest number of employees or achieve the highest tool usage rates. They will be those that make better decisions about how technology changes work and prepare their people accordingly.
Microsoft’s 2026 Work Trend Index, based on a survey of 20,000 AI users across 10 countries, found only 19% in its Frontier group, where individual capability and organizational readiness reinforce each other.
Before asking how to train employees to use AI, decide what you want employees to accomplish with it, what decisions they must remain accountable for, and how their roles need to change.


