Inside Workday's $1.1 Billion AI Learning Bet
On July 22, Workday gave learning teams a number that could empty months from a production calendar: up to 98%.
That was the reported reduction in content creation time for some organizations using the AI authoring capabilities in Workday Learning powered by Sana. A PDF or slide deck can become an interactive course in minutes. The system can propose objectives, draft lessons and quizzes, translate material, assign learning paths and give each employee an AI tutor.
For an L&D team with a backlog, the claim lands like relief. A compliance update that took weeks could go live in hours. A product launch could reach several regions at once. A new manager could ask a question about local policy instead of searching a long module.
The next morning, the harder meeting starts.
The chief learning officer can show that course production got faster. Then the questions spread around the table. The CFO wants to know what Workday’s $1.1 billion purchase bought. A business leader asks whether error rates fell. The manager wants to know who will make time for practice. The employee has the most personal question: does this path create a real chance to enter the role it describes?
None of those questions fits inside a course-authoring benchmark.
Four months before the learning launch, Workday CEO Aneel Bhusri acknowledged the worker behind the benchmark. “A lot of low level HR work is going to get replaced by agents,” he said at a March press conference covered by ITPro. He said that Workday and the rest of the industry needed a plan for displaced employees, then pointed to retraining as part of the response.
The same company is selling the agent that removes routine work and the learning system meant to help people prepare for what follows. That makes the quality of the retraining more than a product metric.
Workday knows the distinction. Its launch release says that completion rates reveal little about capability. Joel Hellermark, Sana’s founder and now Workday’s chief AI officer, put it plainly: checking a box does not build a skill. The product connects learning data with Workday’s records about roles, skills, organizations and locations. It can relate a program to performance, retention, internal mobility or even regional safety incidents.
Putting those records together could make learning easier to evaluate. It also moves the argument from the L&D dashboard into decisions that affect pay, assignments and careers.
When course production becomes cheap, companies lose an old excuse. They can no longer say that personalized training is too slow to make. Now they have to prove that learning changed work. They have to give managers time to observe practice, define what good performance looks like, open roles that trained employees can actually enter and separate useful evidence from convenient activity data.
The software can put a tutor beside the employee, but the company still has to supply the assignment, the manager’s feedback and the opening into which a newly trained person might move. People also have to decide whether a training signal deserves to affect pay or promotion.
Workday is putting AI-generated content, workforce data, performance signals and career movement inside one enterprise system used by more than 11,500 organizations. Buyers are choosing how evidence about learning will travel into decisions about work and careers. Course design is only the first layer.
July 22 put a tutor inside the HR system
Workday’s release gives the tutor a specific employee question: “How do I handle a customer data request in Germany?” Instead of sending the person through a long generic course, the system can retrieve the relevant lesson or policy explanation. It can also suggest what to learn next using the employee’s role, skills, organization and location. A manager who moves regions might see local policy and leadership material. Someone preparing for another role might receive a path aligned with that role’s requirements.
Behind that exchange, a course creator gets an AI editor. Existing presentations, PDFs and course files can become structured lessons. The editor proposes an outline, objectives and knowledge checks. A writing assistant can simplify policy text, draft questions and adapt material for a region or job group. Translation can run from the same source rather than through a sequence of local projects.
Administrators can automate assignments and alter learning paths when an employee joins, changes role or moves to another region. Workday says some customers report compliance reporting that is up to five times faster and learner engagement that is three times higher than with legacy systems.
Those are vendor-reported results, and the qualifiers matter. “Up to” is not an average. Higher engagement does not establish skill transfer. Faster reporting says little about whether the underlying behavior changed. The release does not publish baselines, sample sizes or a controlled comparison for the headline figures.
Still, the workflow is concrete. L&D has long carried a production problem that other business functions rarely see. A policy owner writes a document. An instructional designer turns it into a course. Legal and compliance review it. Regional teams translate it. Administrators build audiences and assignments. Managers chase completion. A reporting team prepares evidence for an audit. By the time the material reaches everyone, the policy, product or tool may have changed.
AI can compress much of that chain. Workday says organizations using the authoring tools have moved multiweek production cycles into hours. The product also allows multiple contributors to work on a course, which can shorten the review path between subject experts, learning teams and compliance owners.
Speed has real business value. A pharmaceutical company cannot wait a quarter to update safety instruction. A bank cannot teach a changed control after the control is already live. A software company introducing a new AI assistant cannot rely on a generic annual module. Timely material reduces the gap between a changed process and an employee’s first attempt to use it.
Speed also multiplies weak material. A vague policy can become a vague course in minutes. An inaccurate slide can produce an inaccurate quiz. A source document written for lawyers may remain useless to a frontline worker after the system simplifies it. Translation can spread the same hidden assumption across dozens of markets.
Source quality moves to the top of the deployment checklist.
Before authoring starts, somebody has to certify which document governs the work, who owns the policy, when it was last reviewed and which local exceptions apply. After the AI drafts a course, a subject expert still has to test its examples against real cases. If the course prepares an employee to make a decision, the review should include people who see the decision fail.
A tutor inside an HR system can make knowledge easier to reach. Its answer is only as dependable as the source, permission and escalation path behind it.
Workday’s availability note makes that boundary visible. The integrated product is globally available for Workday HCM customers, while Sana Learn remains a standalone option. Some regulated and sovereign deployments cannot yet use Sana components. A global buyer therefore needs more than one launch plan. It needs a map of where the tutor can operate, which content can enter it and what employees do when their environment remains on the older system.
Distribution can speed up while accountability stays local.
$1.1 billion bought more than a course factory
Workday announced its agreement to acquire Sana in September 2025 for approximately $1.1 billion. At the time, Sana’s products had served more than one million users across hundreds of enterprises. Workday described a much larger base of 75 million users.
The acquisition release did not frame Sana as a narrow learning-management add-on. Workday called it a new front door for work. Sana brought enterprise search, agents, course generation and personalized tutoring. Workday brought data about people, money, roles and business processes.
The deal bought both sides of that connection.
Workday completed the acquisition on November 4. Its later annual report puts the accounting behind the headline: $1.1 billion in acquisition-date purchase consideration, including $1.0 billion in cash and the fair value of Workday’s existing Sana stake. The filing assigned $903 million to goodwill, which Workday attributed mainly to Sana’s assembled workforce and expected integration benefits. The title of this article refers to that total consideration, not a $1.1 billion cash payment.
Gerrit Kazmaier, Workday’s president of product and technology, described the plan as a single interface for the systems, data and actions employees rely on. Hellermark went further, saying Sana wanted to build the user interface for AI. The language was promotional, but the product boundary was clear. A course factory improves one departmental workflow. An employee interface can answer a policy question, find a file, guide a task, run a workflow and recommend a skill path.
The 2025 release included striking customer examples. Workday said an electric-vehicle manufacturer increased learning engagement by 275%. A European installation distributor with 7,500 employees cut course creation from four months to four days. A fintech company reduced a three-week content cycle to three hours. These examples came from the seller and did not include the detail needed to calculate an independent return, but they show the economic target: remove content-production labor and place more knowledge inside a searchable, adaptive system.
For Workday, the strategic payoff can come from several places. Sana can make Workday Learning more competitive. Search and agents can pull employees back into Workday more often. Personalized learning can make skills data richer. That data can support talent planning, internal recruiting and performance workflows. Each connection increases the cost of replacing one part of the stack.
Customers face a different calculation.
They should not use Workday’s $1.1 billion acquisition price as a proxy for their own return. Their business case depends on contract cost, implementation work, content migration, integration, access controls, change management, manager time and the value of outcomes that actually move.
Content savings are the easiest line to estimate. Count the hours instructional designers, subject experts, translators, administrators and reviewers spend on a course today. Compare that with the new workflow. Include the cost of validating AI output. Separate work eliminated from work merely shifted to another person.
Compliance speed is also measurable. How long does it take to identify an affected population, publish an update, complete assignments and produce evidence? Faster reporting can reduce audit preparation and help leaders see gaps sooner.
Skill development is harder. A course can be produced in an afternoon and still produce no useful change. The return appears only when an employee performs a task better, avoids an error, takes on a larger scope, moves into a needed role or helps a team redesign a workflow.
That delay creates a budget trap. The L&D team can prove its own efficiency quickly, so finance books the authoring savings. Business performance improves later, if at all, and depends on managers and job design outside L&D. The buyer may conclude that the platform delivered because courses became cheaper, even though the workforce problem that justified the purchase remains.
The business case needs separate dates for production and capability. Finance can measure hours to create, translate, assign and report during the first release cycle. Practice, observed performance, role readiness and internal movement take longer. A quarterly review should show both, because a product may make the first release look efficient while the employees remain no more prepared for changed work.
Workday’s larger opportunity is to help customers connect the dates without pretending that two nearby records prove one caused the other.
Cheap content moves the bottleneck to practice
Once a learning team can make ten versions of a course, it has to decide which version deserves to exist.
The temptation will be volume. Every policy owner can request a module. Every product group can publish an update. Every job family can receive a personalized path. AI makes the marginal cost of drafting another lesson look close to zero.
Employees do not receive marginal time.
They still have a workday, a manager and a queue. A hundred relevant courses can be less useful than one protected hour spent solving a real problem with feedback. Personalized recommendations can become personalized overload if the system keeps finding gaps but the company never decides which gap matters now.
Recent workforce research shows why the distinction matters. McKinsey’s HR Monitor 2026 surveyed about 1,300 HR professionals and 5,500 employees across ten countries. Twenty-four percent of employees reported no training participation. More than half received feedback once a year or not at all. Only 11% of organizations took a long-term view of workforce planning.
More course material will not repair those gaps.
An employee can have a perfect AI tutor and still lack a useful assignment. A manager can receive a skills dashboard and still have no time to coach. HR can identify a future capability and still run headcount planning one quarter at a time. The learning platform can reduce friction without supplying the missing operating rhythm.
McKinsey’s April report on the AI upskilling challenge sharpens the point. Seventy-five percent of U.S. workers in its survey expected AI to change their role within five years, while only 45% had gone through a recent upskilling program. The report argues for peer learning and role modeling in daily work because a fixed course about one model can age before the program ends.
Heather Stefanski, McKinsey’s chief learning and talent officer, co-wrote the report from the perspective of someone running the firm’s own programs. McKinsey’s lighthouse teams used real client work, visible role models and peer coaches. The firm reports that participants’ AI proficiency and usage rose by roughly 35 percentage points, from about 15% to about 50%, with gains persisting months after the sprint. People saw a workflow applied, tried it in context and shared the method with nearby teams.
A customer-service employee learning to handle escalations needs a difficult case, an angry customer and a policy boundary. A procurement analyst learning supplier risk needs to compare two real vendors and defend a recommendation with consequences.
Useful practice produces a work sample and an observable decision. Someone accountable for quality gives feedback, then the employee tries again.
An AI tutor can support each step. It can prepare a scenario, explain a mistake, offer another case and help an employee reflect. It can lower the cost of individualized repetition. But someone still has to select the right practice, decide what evidence counts and protect time for the employee to do it.
Quiet, repeatable practice is the product’s strongest case. It can reach an employee who would not ask a senior colleague the same basic question five times. A local manager does not have to invent every scenario, and a global L&D team can update examples without rebuilding the whole program. The tutor can extend scarce coaching, provided the company does not mistake the extension for the coach.
That changes the L&D job.
Instructional designers will spend less time formatting slides and more time validating sources, designing practice and defining assessment. Subject experts will review generated examples and maintain the edge cases the tutor should not answer alone. Learning administrators will become campaign and data operators. L&D leaders will negotiate with business managers for assignments, review time and openings tied to the skills being taught.
Course creation can fall by 98% while the total cost of skill transfer remains stubborn. The remaining cost is more human and closer to operations, so the budget has to follow it there.
Completion cannot prove judgment
Workday’s promise to place learning beside business and people data could matter longer than the 98% authoring figure.
The launch release offers a sensible example. Instead of judging a safety program only by completion, leaders could examine incident data in the relevant regions. They could also relate learning to skill development, performance, retention and internal mobility.
An L&D team can ask whether employees who practiced the procedure made fewer errors and whether trained workers passed a work sample for an adjacent role. It can check whether managers observed the target behavior, whether internal candidates moved into open jobs and whether the result lasted for three months.
It also creates a measurement hazard.
Data living in the same system does not mean one event caused another. High performers may take more courses because they already have supportive managers. Employees selected for an advanced program may have been close to promotion before training began. A team with fewer safety incidents may have better equipment and lower workload. Retention can rise because pay changed, not because a learning path felt relevant.
Companies need stronger evidence than either a completion rate or a simple correlation.
Start with a baseline. Record how the work performs before training: error rate, cycle time, quality score, escalation rate, manager assessment or work-sample result. Choose one or two measures that match the behavior the program is meant to change.
Define the practice. Specify the assignment in which the employee will use the skill. “Apply AI” is not an assignment. “Use the approved assistant to draft a supplier-risk summary, verify every cited source and present the recommendation to procurement” is observable.
Use a comparison when possible. A staged rollout can compare teams entering the program at different times. Repeated work samples can show whether performance improves. Manager ratings should use the same rubric and calibration examples.
Wait long enough. A quiz given five minutes after a lesson measures recall. A work sample one week later tests application. A manager observation after a month starts to reveal transfer. A sustained business result may need a quarter.
Record adverse movement as well as gains. If cycle time falls but rework rises, the skill did not transfer cleanly. If employees complete more learning but cannot access stretch assignments, the system may be generating aspiration without opportunity. If recommendations rely on stale skills data, employees may be routed toward the wrong path.
Microsoft’s 2026 Work Trend Index gives a useful target for AI-related learning. In its survey of 20,000 AI users across ten countries, 50% named quality control of AI output as a skill becoming more important, and 46% named critical thinking. Eighty-six percent said they treated AI output as a starting point and remained responsible for the thinking.
Those skills do not show up in a login count.
A company has to watch a person inspect an answer, identify a weak assumption, correct the work and own the final decision. The evidence can be a reviewed work product, a scenario score, an observed customer interaction or a documented exception. It should be close enough to the job that an employee recognizes the task.
Measurement also needs a boundary. Learning telemetry should not quietly become performance evidence without a declared purpose, a validated relationship to the job and a way for employees to understand or challenge the signal. A person who asks the tutor many basic questions may be learning deeply, entering a new domain or working in a second language. A person who completes every module quickly may have learned little.
A login count misses the moment when a person spots a weak assumption, fixes it and takes responsibility for the decision.
Managers still own the transfer
On Monday, an employee finishes an AI course. On Tuesday, the same queue and the same assignments are waiting. Unless the manager changes the work, the newly learned skill has nowhere to go.
Software can recommend a path and L&D can build a practice. The manager controls the assignment, watches enough of the result to judge quality and gives feedback while another attempt is still possible.
Microsoft’s research quantifies part of that effect. In a separate study of 1,800 workers cited in the Work Trend Index, active manager modeling was associated with a 17-point lift in reported AI value, a 22-point lift in critical thinking about AI use and a 30-point lift in trust in agentic AI. Workers whose managers created psychological safety around experimentation were 1.4 times more likely to be frequent users of agentic AI.
Manager behavior mattered more than another message telling employees to adopt.
Boston Consulting Group found the same gap from a different angle. Its AI at Work 2026 survey covered 11,749 workers across 14 markets. Seventy-two percent said expectations for their skills had changed, but only 36% believed they had received adequate upskilling. Among regular frontline AI users, 42% said AI saved at least eight hours a week. Sixty-six percent received limited or no guidance on what to do with the saved time.
That unused time is a management decision waiting to happen.
If a support employee saves four hours, the manager can fill the space with more tickets, remove a position, assign complex cases, schedule practice or give the employee a project that builds an adjacent skill. The learning platform can suggest a path. The manager controls the work that makes the path real.
Managers need a small weekly routine.
At the start of a learning sprint, manager and employee choose one work outcome. They name the task, the quality bar and the evidence to collect. The employee gets an assignment within days, not after finishing a long curriculum. The manager reviews a sample using a short rubric. They agree on the next attempt and record whether the employee can take on more scope.
This routine requires capacity. A manager with 14 direct reports cannot provide deep observation for every skill every week. The company has to decide which capabilities deserve manager time, when peers can review and when an expert cohort should run calibration. It may need to count coaching hours in the program budget.
Without that budget, AI learning creates a familiar transfer.
L&D reports faster production. Employees receive more recommendations. Managers inherit more assessment and coaching. Finance sees lower content cost but does not see the review burden. The program appears efficient because the expensive human work moved out of the L&D cost center.
Manager evidence also needs limits. A single manager’s impression can preserve favoritism under a more modern label. Rubrics should describe observable behavior. Reviewers need examples of weak, acceptable and strong work. High-stakes decisions should use more than one signal. An employee should know what skill is being assessed and have another chance to demonstrate it.
An integrated HR system can place the target skill, practice record, manager feedback and next role in one workflow. It can remind a manager to observe a second attempt. More revealingly, it can show that an employee completed a path but never received the promised assignment.
The last signal may be more valuable than the completion rate. It tells HR that the learning program failed at work, not in the course.
A learning-to-work evidence table
A buyer can test an AI learning program with one operating table. Each row starts with a business change, not a course title. The examples below are hypothetical, and the table does not imply that Workday supplies every field.
| Business change | Role or task affected | Target skill | Practice assignment | Manager evidence | System signal | Movement or stop decision |
|---|---|---|---|---|---|---|
| Customer-support agent handles routine cases | Support staff take harder escalations | Diagnose ambiguity and repair trust | Resolve three complex cases with a senior reviewer | Case reasoning, tone, policy accuracy, escalation choice | Rework, repeat contact and customer outcome | Expand scope after two clean reviews; pause if rework rises |
| AI drafts supplier analysis | Procurement analysts review more generated work | Verify sources and weigh commercial risk | Produce a supplier-risk recommendation with cited evidence | Source quality, missing risk, decision logic | Review cycle, exception rate and approved recommendation | Move toward category ownership; stop automation if unsupported claims pass |
| Coding assistant increases output | Engineers review more code | Test design and defect detection | Review an AI-generated change seeded with known faults | Defects found, severity judgment, test coverage | Escaped defects and rework | Grant wider review responsibility after calibrated samples |
| New policy reaches several regions | Local managers interpret changed rules | Apply policy to regional cases | Work through two local scenarios and one employee question | Correct rule, local exception and escalation | Query themes, error pattern and case outcome | Release policy authority only where local validation passes |
| AI saves administrative time | Coordinator prepares for an adjacent role | Workflow design and stakeholder judgment | Redesign one recurring process and run a controlled pilot | Problem definition, handoffs, exception handling | Cycle time, failure rate and stakeholder response | Offer stretch work or internal interview; stop if time savings simply become more volume |
| Employee targets an internal move | Candidate lacks one required capability | Role-specific applied skill | Complete a work sample used by the destination team | Calibrated score from destination manager and peer | Skill evidence, opening status and interview result | Advance to internal process; do not keep assigning learning when no path exists |
The table keeps fields separate that platforms often collapse. Content explains the task; practice exposes behavior. A manager judges the attempt, while system data records what happened around the work. The final column forces an actual decision about scope, career movement or a weak program that should stop.
Each row also reveals an owner. L&D can design the practice. The business manager owns the assignment. A domain expert owns the quality standard. HR owns fairness and career rules. IT and security own access. Finance owns the claim that the program changed value.
No single owner can certify the full row. The table should travel with the program from purchase to renewal. At procurement, a buyer can ask whether the platform supports each signal without forcing every sensitive record into one audience. During implementation, the team can configure permissions and handoffs. At quarterly review, leaders can see which rows produced movement and which produced only content.
An employee should know the role or task being prepared for, the practice they will receive, the evidence a manager will consider and what happens if they meet the bar. “Personalized learning” becomes credible when personalization changes an opportunity, not only a recommendation.
Some rows will remain compliance rows. The outcome may be correct behavior and fewer incidents rather than promotion. That is fine. The discipline is the same: start with the behavior, collect evidence close to work and avoid treating completion as the result.
The table will make certain programs look smaller. It may cut a catalog of 300 AI-generated courses down to six workforce changes the company can support this quarter. Those six changes carry more organizational weight when each one has practice, observation and a movement decision. Three hundred completions can disappear inside a dashboard.
Career paths decide whether retraining is real
Learning becomes a career-ladder issue when AI removes the work that used to teach people.
BCG reports that 67% of workers say AI has taken over simpler tasks and left them with more complex work. That sounds productive at the team level. For a newer employee, it can erase the repetition that built context before harder judgment arrived.
The World Economic Forum and PwC put the exposure in demographic terms. Their June report on AI and entry-level work says more than one in three young workers globally are employed in occupations with medium-to-high exposure to AI-driven task change. Its framework covers job access, job design, talent pipelines and education alignment.
An AI tutor can help rebuild part of the missing practice. It can create cases, answer questions privately and let a learner repeat a task without consuming every minute of a senior employee’s time. That is useful for employees who hesitate to expose a gap in a group setting.
A simulation does not grant organizational trust by itself. A destination team still has to accept the evidence, and a hiring manager has to believe that the practice resembles the role. The company needs a vacancy, project or stretch assignment. A trained employee needs priority over an external candidate often enough for internal mobility to feel real.
The OECD’s June report A Skills-First Labour Market offers two practical examples. Deutsche Telekom has built more than 500 role profiles and uses annual self-assessment followed by manager validation to identify skill gaps and personalize training. Coca-Cola Europacific Partners maps three to five key skills to each role and gives employees a career platform for development and internal opportunities. The OECD reports that more than half of permanent roles at the company are filled through internal mobility.
The examples matter because they connect four objects: a role, a small set of skills, evidence checked by another person and an actual move.
Many corporate learning systems stop after the second object. They infer a skill and recommend content. Employees accumulate badges. Job descriptions remain broad. Managers still hire through personal networks or familiar credentials. The learning layer looks intelligent while the internal labor market behaves as before.
Personalization can widen access to a career path, especially for an employee whose manager would not have suggested it. It can also repeat old assumptions. A recommendation grounded in the current role, recorded skills and manager history may keep showing a worker more of the work they already do. Employees need to see the skills attached to a role, correct stale records and express an interest that the system did not infer.
Workday has pieces of the longer chain. Its release says recommendations can align with the requirements of a future role. Learning data can sit beside internal mobility and performance data. Workday has also described Sana learning alongside HiredScore’s internal-mobility capabilities.
The buyer must decide how those pieces connect.
For each priority role, define a short skill set and a work sample. Let employees see the standard before they begin. Ask destination managers to calibrate examples. Give qualified internal candidates a real interview or project. Track where the process breaks: no practice, no manager review, no open role, rejected evidence or a pay band that makes movement unattractive.
The company should also publish what learning cannot promise.
Completing a path may not guarantee a job. It can guarantee consideration, feedback and a transparent standard. A scarce role may have more qualified internal candidates than openings. A company can state that honestly. What it should avoid is recommending months of learning for a role it has no plan to fill internally.
Recommending months of preparation for a role the company will not fill internally turns development into unpaid hope. Career paths make the AI learning investment testable. If skills improve but internal movement stays flat, leaders can inspect whether openings, manager incentives or pay rules blocked the path. If mobility rises but performance falls, the assessment was weak. If retention improves among employees who receive practice and movement, the company has a stronger result than engagement alone.
The employee experiences the system through a simpler question: did learning change what I am trusted and paid to do?
Procurement should price the human system
Workday’s launch gives procurement teams a polished list of capabilities. The buying process should add a less polished list of operating costs.
The product can use role, skills, organization and location to personalize recommendations. During security review, buyers should map which data is required, who can see it, how inferred skills are corrected and which records can influence a career decision. A tutor needs access to be useful. Permissions still have to follow the underlying source.
Content needs a similar lineage check. Every generated course should point to an authoritative source, an owner, a review date and a path for local exceptions. An employee who receives an uncertain answer needs to know when the material was last validated and where to escalate it.
Then the implementation spreadsheet should leave the software column. Estimate the work samples, sandbox environments, customer cases and supervised assignments needed for priority skills. Add the hours subject experts will spend validating examples. Add manager time for observations and calibrated review. A platform contract without those hours buys distribution without transfer.
Opportunity belongs in the same spreadsheet. List the roles, projects and expanded duties available to employees who meet the bar. If a program targets a role with no openings, call it readiness rather than mobility. If redeployment is part of an automation plan, reserve internal interview capacity before affected jobs disappear.
Baselines and stop signals should exist before launch. Completion, engagement and tutor usage can diagnose a program, but they do not prove better work. A buyer also has to budget for the older system that remains. Workday says some regulated and sovereign environments cannot yet use Sana components, so a global organization may carry duplicated administration and inconsistent employee experiences during migration.
These costs show where the product ends and the employer’s work begins.
Workday has made the production side of learning dramatically more credible. A learning team can move from a blank page to a live course in hours. Employees can receive an explanation in context. Administrators can update assignments as roles and locations change. Data can connect the program to the rest of the employee record.
That removes friction that companies have tolerated for years. It also makes old excuses harder to defend. Personalized courses expose generic career paths. Skills records expose reliance on completion. A recommendation for a future role exposes whether destination managers accept the evidence. Saved authoring time leaves finance with a plain question about where the time and money went.
The $1.1 billion Sana bet will be judged first by product adoption and revenue. Customers should use a tougher measure.
The useful test is whether the tutor helped an employee perform a changed task, a manager witnessed the skill and another team trusted the evidence enough to change the person’s scope, pay or route into a new role. The business result also has to improve without shifting invisible review work to managers.
At the quarterly CHRO and CFO review, the packet should name at least one changed task, the employee who practiced it, the manager who observed the attempt and the staffing decision that followed. If the last page ends with completion and engagement, the career claim is still waiting for evidence.
Published July 23, 2026.