# Learning Analytics: What It Is and How to Use It at Work

> Learning analytics explained: SoLAR's 2026 definition, the four types, one workplace course taken from report to decision, real cases and two templates.

**Author:** Carolina Martin  
**Published:** 2026-11-19  
**Updated:** 2026-11-16  
**Categories:** Instructional Designers, L&D Teams

---

Picture a regional bus operator eight weeks after launching a short online course on its new accessible boarding procedure. The report says 94% of drivers finished it, and the operations manager has a list of ramp-fault calls and accessibility complaints that has not shrunk. In this hypothetical example, the completion figure says who finished. It does not tell anyone what to change.

I'm an instructional designer at Coursebox, and I test every release, including the reports where learner results land. Coursebox makes an [AI LMS with built-in learner reporting](https://www.coursebox.ai/ai-lms), so I have a commercial interest in how teams use training data. This guide explains what learning analytics is under the field's current definition, the four types, how to walk one workplace course from a report to a decision, and what to watch when the learners are your own staff. It includes three real cases and two editable Word templates.

## What is learning analytics?

Learning analytics is the collection, analysis, interpretation and communication of data about learners and their learning, used to answer a specific question about how people learn and then to change the course, the coaching or the work practice that follows.

[View the full-size learning analytics diagram](https://cdn.sanity.io/images/3qyhkhue/production/6b23e2a0db06bb06c9175167473afefd326415a5-1536x1024.webp)

That plain version follows the definition the Society for Learning Analytics Research (SoLAR) adopted after a review that began in March 2025. Its [2026 final report](https://www.solaresearch.org/wp-content/uploads/2026/08/SoLAR_Position-Paper_Final_August-2026.pdf) defines learning analytics as "the collection, analysis, interpretation and communication of data about learners and their learning that provides theoretically relevant and actionable insights to enhance learning and teaching."

The older wording still circulates. In 2011 the field described itself as "the measurement, collection, analysis and reporting of data about learners and their contexts, for purposes of understanding and optimising learning and the environments in which it occurs." SoLAR revised it because, in the report's words, it had become hard for authors and reviewers "to agree on what was in or out of scope". The task force formed in March 2025 and presented the new definition in May 2025. After more community feedback, including a 2026 survey, the final report kept the definition unchanged and made minor changes to the scope statement.

Three changes affect how a workplace team uses the term:

- **The words measurement and reporting have been replaced in the definition. Those activities remain useful, but data collection and reporting alone do not meet the revised definition.**
- **Interpretation and communication are in.** Someone has to explain what the data means to the people who can act on it.
- **Insights must be actionable and tied to theory.** The analysis should rest on an idea about how people learn, and it should lead to something a person can do.

The report adds that this work "often involves descriptive methods for analysis, and frequently uses data visualization, dashboards, or data storytelling to communicate insights." Dashboards are a channel, not the point.

| Part of the definition | What it means | The question it asks of your training | What it looks like at work |
| --- | --- | --- | --- |
| Collection | Gathering data about learners and their learning | What do we record, and from which systems? | Course completions, question answers and supervisor check sheets |
| Analysis | Finding patterns in that data | Where do results differ, and by how much? | One question missed far more often than the rest |
| Interpretation | Explaining what a pattern means for learning | Why might people be getting this wrong? | The course shows one version of a task that some staff never perform |
| Communication | Getting the explanation to the people who act | Who needs to know, and in what form? | A one-page note to the course owner and two supervisors |
| Actionable insight | A change someone can make to improve learning | What will we change, and how will we know it helped? | A new practice step and a check on the job eight weeks later |

### Is it learning analytics or a report?

SoLAR's scope statement gives a practical test. It treats as out of scope "work focused solely on aspects of methodology and data manipulation that does not discuss, work towards or include closing the loop to stakeholders". Closing the loop means bringing the findings back to the people who can change something.

It also says that work comparing "departments, institutions, regions or countries (for example management dashboards...)" is "usually classed as academic analytics". A dashboard ranking sites by completion rate would usually count as academic analytics under that test. It becomes learning analytics when someone uses it to ask why people at one site are struggling to learn something, and changes the training in response.

```json
{
  "_key": "849dce486172",
  "_type": "calloutBox",
  "body": [
    {
      "_key": "d6eeab87f888",
      "_type": "block",
      "children": [
        {
          "_key": "4ca4b68c10c3",
          "_type": "span",
          "marks": [],
          "text": "SoLAR wrote its in-scope and out-of-scope lists to guide papers for its journal and conference, not to police workplace practice. I use them here as the field's own test for a fuzzy term. They do not make a completion report useless; they show that it is the first step, not the whole job."
        }
      ],
      "markDefs": [],
      "style": "normal"
    }
  ],
  "markDefs": null,
  "title": "The scope statement was written for researchers",
  "variant": "note"
}
```

```json
{
  "_key": "460f08158a99",
  "_type": "blogCaseStudy",
  "body": [
    {
      "_key": "2120b7819d5e",
      "_type": "block",
      "children": [
        {
          "_key": "ea963b326ac4",
          "_type": "span",
          "marks": [],
          "text": "In late 2016, Boeing engineers took an MITxPRO online course on the architecture of complex systems, with Boeing paying the fee. Researchers from Indiana University and Boeing wanted to see how the engineers actually moved through a large course."
        }
      ],
      "markDefs": [],
      "style": "normal"
    },
    {
      "_key": "5fe6674e8774",
      "_type": "block",
      "children": [
        {
          "_key": "283c29027b49",
          "_type": "span",
          "marks": [
            "strong"
          ],
          "text": "What they did:"
        },
        {
          "_key": "9161a1cf09b3",
          "_type": "span",
          "marks": [],
          "text": " the team analysed the course's event logs and turned them into visual maps of learner paths, module use and performance by module."
        }
      ],
      "level": 1,
      "listItem": "bullet",
      "markDefs": [],
      "style": "normal"
    },
    {
      "_key": "e830544fd974",
      "_type": "block",
      "children": [
        {
          "_key": "ce06149439e3",
          "_type": "span",
          "marks": [
            "strong"
          ],
          "text": "Evidence and result:"
        },
        {
          "_key": "9ba9b77fc4ad",
          "_type": "span",
          "marks": [],
          "text": " 1,565 engineers were active, \"generating nearly 31 million click event records\". Learners used 310 of the course's 551 lowest-level modules, and in week one those who later disengaged interacted \"with fewer problems and videos\", "
        },
        {
          "_key": "0e72d3b80513",
          "_type": "span",
          "marks": [
            "46800b5a6c7e"
          ],
          "text": "the team reported in PLOS ONE"
        },
        {
          "_key": "e5f44b078448",
          "_type": "span",
          "marks": [],
          "text": "."
        }
      ],
      "level": 1,
      "listItem": "bullet",
      "markDefs": [
        {
          "_key": "46800b5a6c7e",
          "_type": "link",
          "href": "https://doi.org/10.1371/journal.pone.0215964"
        }
      ],
      "style": "normal"
    },
    {
      "_key": "1f18afc3e711",
      "_type": "block",
      "children": [
        {
          "_key": "538007202267",
          "_type": "span",
          "marks": [
            "strong"
          ],
          "text": "Limitation:"
        },
        {
          "_key": "a4d1da4dd77d",
          "_type": "span",
          "marks": [],
          "text": " one course and one run, described only. The paper did not test whether any redesign improved learning, and its own counts of registered engineers differ slightly between the abstract and the methods."
        }
      ],
      "level": 1,
      "listItem": "bullet",
      "markDefs": [],
      "style": "normal"
    },
    {
      "_key": "9527ee994342",
      "_type": "block",
      "children": [
        {
          "_key": "fd58ec2a983b",
          "_type": "span",
          "marks": [
            "strong"
          ],
          "text": "Lesson:"
        },
        {
          "_key": "5b9fea1bd2b7",
          "_type": "span",
          "marks": [],
          "text": " logs can show which parts of a course nobody uses and where early disengagement starts. They cannot say why, or whether a change will help, so the analysis has to end in a question for someone who can act."
        }
      ],
      "level": 1,
      "listItem": "bullet",
      "markDefs": [],
      "style": "normal"
    }
  ],
  "heading": "Case study: Boeing engineers in an online systems engineering course, USA",
  "image": {
    "alt": "A person at a wooden table works on a laptop showing a slide with a technical diagram",
    "photographer": "ThisIsEngineering",
    "photographerUrl": "https://www.pexels.com/@thisisengineering",
    "url": "https://cdn.sanity.io/images/3qyhkhue/production/68b6b5a5433266e0c22ef7c597556662c214d887-1600x1067.webp"
  },
  "imagePosition": "right",
  "markDefs": null
}
```

## The SoLAR report and other sources: which to read

Three primary sources cover the definition and its use at work:

- **2026:** [Reimagining Learning Analytics](https://www.solaresearch.org/wp-content/uploads/2026/08/SoLAR_Position-Paper_Final_August-2026.pdf), SoLAR's final report on the new definition, with the out-of-scope list on pages 7 and 8.
- **2026:** SoLAR's page on [what learning analytics is](https://www.solaresearch.org/about/what-is-learning-analytics/), which gives the four methods in one line each.
- **2019:** [An Infrastructure for Workplace Learning Analytics](https://doi.org/10.18608/jla.2019.62.9) by Adolfo Ruiz-Calleja and colleagues, in the Journal of Learning Analytics, on applying the field outside formal education.

If you only read one, read pages 2 to 8 of the SoLAR report. It sets the old and new definitions side by side and explains what changed.

## The four types of learning analytics

Here is the hypothetical example in full. A regional bus operator has 90 drivers across three depots: North (35 drivers), Harbour (30) and Westfield (25). It has introduced a new accessible boarding procedure covering three tasks: deploying the ramp, securing a wheelchair, and what to do when the ramp will not deploy. Every driver must complete a 20-minute online course that ends with six questions.

The course owner wants drivers to handle the ramp and its faults correctly on their first attempt, without leaving a passenger waiting. Eight weeks after launch, two kinds of data exist: the course report, and the operations team's ramp-fault calls and accessibility complaints by depot. The procedure owner approves all course content; this example gives no operating instructions.

The example suits the four types because the report alone looks fine, yet the job data says something is still going wrong. SoLAR's page lists four methods, and each one asks a different question of the same course.

### 1. Descriptive: what happened

SoLAR describes descriptive analytics as work to "aggregate data to reveal trends from past performance." Most LMS reports sit here.

- **Question for the bus course:** who finished, and which questions did drivers miss?
- **Data you need:** completions and answers to each question from the course report.
- **What someone does with it:** the course owner sees that 85 of 90 drivers finished (94%), but 48 of those 85 missed the question on what to do when the ramp will not deploy.

### 2. Diagnostic: why it happened

Diagnostic analytics sets out to "investigate data to uncover why trends occurred." It usually means splitting the data or joining it to another source.

- **Question for the bus course:** are the misses spread evenly, or concentrated somewhere?
- **Data you need:** answers to each question by depot, ramp-fault calls by depot from operations, and which bus types each depot runs.
- **What someone does with it:** 24 of Harbour's 29 finishers missed the question, against 24 of 56 at the other two depots. Harbour runs the oldest buses, whose ramp works differently from the one shown in the course. The course owner now has a likely reason, which the Harbour supervisor can confirm with drivers.

```json
{
  "_key": "e214e2b627d5",
  "_type": "blogCaseStudy",
  "body": [
    {
      "_key": "104123f66908",
      "_type": "block",
      "children": [
        {
          "_key": "278685da450c",
          "_type": "span",
          "marks": [],
          "text": "Every accredited ear, nose and throat surgery residency programme in Taiwan records its competency decisions on one national platform. Senior residents were not reaching the expected level of independence, and the question was whether the trainees or the training system explained it."
        }
      ],
      "markDefs": [],
      "style": "normal"
    },
    {
      "_key": "9d41c6d75a48",
      "_type": "block",
      "children": [
        {
          "_key": "53d844119625",
          "_type": "span",
          "marks": [
            "strong"
          ],
          "text": "What they did:"
        },
        {
          "_key": "b45532b50525",
          "_type": "span",
          "marks": [],
          "text": " researchers analysed the platform's committee decisions on how much supervision each resident needed for each professional task, and modelled what was associated with residents reaching the expected level."
        }
      ],
      "level": 1,
      "listItem": "bullet",
      "markDefs": [],
      "style": "normal"
    },
    {
      "_key": "b8b9f91dc336",
      "_type": "block",
      "children": [
        {
          "_key": "3b2bfa95ec95",
          "_type": "span",
          "marks": [
            "strong"
          ],
          "text": "Evidence and result:"
        },
        {
          "_key": "0dd590e6205a",
          "_type": "span",
          "marks": [],
          "text": " across 3,504 decisions for 292 residents, targets were met in 2,558 (73%). Decisions made where the committee reviewed residents less often had lower odds of meeting the target (adjusted odds ratio 0.41), "
        },
        {
          "_key": "8d2c4e3cd6f2",
          "_type": "span",
          "marks": [
            "9c1301509aa0"
          ],
          "text": "according to the study in JMIR Medical Education"
        },
        {
          "_key": "643816a60315",
          "_type": "span",
          "marks": [],
          "text": "."
        }
      ],
      "level": 1,
      "listItem": "bullet",
      "markDefs": [
        {
          "_key": "9c1301509aa0",
          "_type": "link",
          "href": "https://doi.org/10.2196/95066"
        }
      ],
      "style": "normal"
    },
    {
      "_key": "4839fedeae75",
      "_type": "block",
      "children": [
        {
          "_key": "faf7156d9dda",
          "_type": "span",
          "marks": [
            "strong"
          ],
          "text": "Limitation:"
        },
        {
          "_key": "df8cea969030",
          "_type": "span",
          "marks": [],
          "text": " the authors state that \"the cross-sectional design precludes causal inference\", and no change to review frequency was tested."
        }
      ],
      "level": 1,
      "listItem": "bullet",
      "markDefs": [],
      "style": "normal"
    },
    {
      "_key": "68073d610d67",
      "_type": "block",
      "children": [
        {
          "_key": "8af42fc496b7",
          "_type": "span",
          "marks": [
            "strong"
          ],
          "text": "Lesson:"
        },
        {
          "_key": "0464e9957698",
          "_type": "span",
          "marks": [],
          "text": " the pattern pointed at something the programmes control, how often the committee met, rather than at individual trainees. Diagnostic work is most useful when it looks at the training system as well as the learners."
        }
      ],
      "level": 1,
      "listItem": "bullet",
      "markDefs": [],
      "style": "normal"
    }
  ],
  "heading": "Case study: Otolaryngology residents across Taiwan",
  "image": {
    "alt": "Two doctors in scrubs and white coats sit at a glass desk looking at a laptop together",
    "photographer": "Thirdman",
    "photographerUrl": "https://www.pexels.com/@thirdman",
    "url": "https://cdn.sanity.io/images/3qyhkhue/production/df2ab5db6a6a80555f81e29bd1912130f9a8e894-1600x1066.webp"
  },
  "imagePosition": "right",
  "markDefs": null
}
```

### 3. Predictive: what is likely to happen next

Predictive analytics uses "models to forecast future trends and outcomes." In large programmes that can mean a statistical model; in a team of 90, a simple, stated rule is usually more honest.

- **Question for the bus course:** which drivers are most likely to struggle on their first real ramp fault?
- **Data you need:** the diagnostic result and the bus roster showing who drives the older buses.
- **What someone does with it:** the working assumption is that Harbour drivers who missed the question and are rostered on the older buses face the highest risk. It is a rough forecast, so the Harbour supervisor checks it with those drivers before anyone acts on it.

### 4. Prescriptive: what to do about it

Prescriptive analytics aims to "recommend actionable strategies to optimize learning." This is where the loop closes.

- **Question for the bus course:** what change will help drivers handle a fault on the older ramp?
- **Data you need:** everything so far, plus the procedure owner's approved steps for the older ramp.
- **What someone does with it:** the course owner adds the older ramp's approved fault steps and a practice question to the course. The Harbour supervisor runs a five-minute practical check at shift handover. Eight weeks later, the operations manager compares ramp-fault calls and complaints by depot with the previous eight weeks.

```json
{
  "_key": "5e63fd9257ee",
  "_type": "calloutBox",
  "body": [
    {
      "_key": "b79b743ae1f4",
      "_type": "block",
      "children": [
        {
          "_key": "ad7fcf75a78c",
          "_type": "span",
          "marks": [],
          "text": "Start with the change you might make, such as rewriting a module or adding a practical check, and the person who would make it. Then decide which of the four questions you need answered. Otherwise the report sets the agenda, and it can only show what it happens to measure."
        }
      ],
      "markDefs": [],
      "style": "normal"
    }
  ],
  "markDefs": null,
  "title": "Write the decision before you open the report",
  "variant": "tip"
}
```

## How the four types relate

Diagrams sometimes draw the four types as a ladder, with prescriptive at the top as the goal. SoLAR's page lists them as methods, not stages, and its report notes that the work often involves descriptive methods. In the bus course, a descriptive report plus one diagnostic split led straight to a prescriptive change; a predictive step added little beyond a short list of drivers to check first.

What links them is the question. Each type answers a different question about the same course, and the decision you face sets which ones you need. A team that only ever reads descriptive reports is not doing something wrong, but it stops before the step SoLAR now puts in the definition: interpreting the pattern and getting it to someone who will act.

## How learning analytics is used in workplace training

Workplace questions tend to fall into three groups, and naming the group helps you find the right data:

- **About the learning experience:** where people stall, which questions they miss, which parts of a course nobody opens. In the bus course: the missed ramp-fault question.
- **About the learners:** who needs more practice, coaching or a different route through the material. In the bus course: Harbour drivers on the older buses.
- **About the programme and the business:** whether training changed how work is done, and for which teams. In the bus course: whether ramp-fault calls and complaints fall at Harbour after the change.

In 2019, Ruiz-Calleja and colleagues wrote that "the field of workplace LA is still in its early stages of development", and noted that, "in contrast to formal education, workplace learning is often driven by personal interest or by problems" that come up at work. That second point is why job data, not only course data, carries so much of the useful evidence in a workplace.

## Who benefits, and what each person can act on

Learning analytics is useful only when the person who receives the finding can do something about it. For the bus course, each role gets a different slice:

| Person | Question they can act on | Data they need | What they can change |
| --- | --- | --- | --- |
| Driver | Which step do I need to practise? | My own answers and the supervisor's check | Practise the older ramp's fault steps before the next shift on that bus |
| Depot supervisor | Who on my team needs a practical check? | Answers by driver for my depot, and the roster | Run the five-minute check at handover |
| Course owner | Which part of the course is not working? | Answers by question and by depot | Rewrite the module and add a practice question |
| Operations manager | Did the change reduce faults and complaints? | Fault calls and complaints by depot, before and after | Keep, extend or drop the practical check |
| Procedure owner | Is the approved procedure clear for every bus type? | The course owner's finding | Issue the older ramp's steps as a separate job aid |

The learner appears first on purpose. Analytics that only flows upward, to managers, misses the person most able to fix a gap in their own practice.

## Learning analytics tools: LMS reports, LRS and BI

Most workplace teams already hold three kinds of data, each stored in a different type of tool.

- **LMS reports:** completions, progress, scores and answers to each question inside your courses. This is the descriptive layer. Our guide to [LMS reporting metrics](https://www.coursebox.ai/blog/lms-reporting-metrics) covers which reports to run and how often.
- **A learning record store (LRS):** a database for xAPI statements, which can record learning activity outside a course package, such as a supervisor's observation or time in a simulator. SCORM is the older course-package standard many LMSs use to record completion and scores, and anyone who's built a SCORM package knows it reports what happens inside the package and little else. The [xAPI vs SCORM](https://www.coursebox.ai/blog/xapi-vs-scorm) guide compares the two standards.
- **Business intelligence (BI) tools and job systems:** HR records, operations logs, quality data and customer feedback. These hold the evidence of whether work changed, and a BI tool is often where course data and job data finally meet.

Choosing tools comes after choosing questions. If you are comparing LMS platforms on their reporting, [LMS analytics](https://www.coursebox.ai/blog/lms-analytics) lists the metrics to look for.

```json
{
  "_key": "a360c6726989",
  "_type": "blogCaseStudy",
  "body": [
    {
      "_key": "c8e29336820d",
      "_type": "block",
      "children": [
        {
          "_key": "ee2c46954475",
          "_type": "span",
          "marks": [],
          "text": "Junior doctors in their second and third years of internal medicine training at the University of Virginia handle patient messages and calls through the electronic health record inbox. Inbox work contributes to burnout, and the training programme wanted objective evidence of how residents managed it."
        }
      ],
      "markDefs": [],
      "style": "normal"
    },
    {
      "_key": "30b0f9b819fb",
      "_type": "block",
      "children": [
        {
          "_key": "d82492823809",
          "_type": "span",
          "marks": [
            "strong"
          ],
          "text": "What they did:"
        },
        {
          "_key": "019a72a80697",
          "_type": "span",
          "marks": [],
          "text": " the programme built an individual report for each resident from electronic health record data on inbox efficiency and quality, and used it in structured one-to-one coaching across two training periods."
        }
      ],
      "level": 1,
      "listItem": "bullet",
      "markDefs": [],
      "style": "normal"
    },
    {
      "_key": "fc6f7a2c7c8d",
      "_type": "block",
      "children": [
        {
          "_key": "cc2a2be9b057",
          "_type": "span",
          "marks": [
            "strong"
          ],
          "text": "Evidence and result:"
        },
        {
          "_key": "56c6a2e6358a",
          "_type": "span",
          "marks": [],
          "text": " residents cut their turnaround time for patient calls by 2.2 days, and time spent in patient calls fell from 0.99 to 0.74 minutes a day. Perceived burnout, confidence and efficiency did not significantly change, and 6 of 7 faculty found the data at least moderately useful, "
        },
        {
          "_key": "67418d2b8adf",
          "_type": "span",
          "marks": [
            "9eb10770a09a"
          ],
          "text": "the authors reported in the Journal of General Internal Medicine"
        },
        {
          "_key": "d89df8cb26ea",
          "_type": "span",
          "marks": [],
          "text": "."
        }
      ],
      "level": 1,
      "listItem": "bullet",
      "markDefs": [
        {
          "_key": "9eb10770a09a",
          "_type": "link",
          "href": "https://doi.org/10.1007/s11606-026-10720-z"
        }
      ],
      "style": "normal"
    },
    {
      "_key": "eeaec65496de",
      "_type": "block",
      "children": [
        {
          "_key": "428f2968328a",
          "_type": "span",
          "marks": [
            "strong"
          ],
          "text": "Limitation:"
        },
        {
          "_key": "5916f8794774",
          "_type": "span",
          "marks": [],
          "text": " one site, measured before and after with no control group, and the analytics were combined with coaching, so neither can take the credit alone. The abstract does not report how many residents took part."
        }
      ],
      "level": 1,
      "listItem": "bullet",
      "markDefs": [],
      "style": "normal"
    },
    {
      "_key": "e30242dc7c43",
      "_type": "block",
      "children": [
        {
          "_key": "a8bb466a608f",
          "_type": "span",
          "marks": [
            "strong"
          ],
          "text": "Lesson:"
        },
        {
          "_key": "f1a50d3ef079",
          "_type": "span",
          "marks": [],
          "text": " the data that changed practice came from the job system, not a course, and it worked through a coaching conversation rather than a dashboard. Measured work improved while how residents felt about it did not, so track both."
        }
      ],
      "level": 1,
      "listItem": "bullet",
      "markDefs": [],
      "style": "normal"
    }
  ],
  "heading": "Case study: Internal medicine residents' inbox coaching, University of Virginia, USA",
  "image": {
    "alt": "Hands typing on a white keyboard on a desk beside a stethoscope",
    "photographer": "Vitaly Gariev",
    "photographerUrl": "https://www.pexels.com/@silverkblack",
    "url": "https://cdn.sanity.io/images/3qyhkhue/production/80f18b84f2d73250282915ca5d46ddea93a9b7ca-1600x900.webp"
  },
  "imagePosition": "right",
  "markDefs": null
}
```

## Personalised and adaptive learning: what the data can drive

Adaptive learning can put a prescriptive decision into a course: the system uses evidence of learning to change what appears next. Personalisation is broader and may also use a learner's role, preferences or choices. A simple version sends a learner who misses a question to extra practice before they continue.

In the bus course, an adaptive rule could send any driver who misses the ramp-fault question to the practice for their bus type. That is useful, but it only adapts to what the course records. It cannot see the roster, the fault calls or the supervisor's check, so the larger decisions in the example still sit with people. The [adaptive learning design](https://www.coursebox.ai/blog/adaptive-learning-design) guide covers how to plan the rules.

## Ethics and privacy when the learners are your employees

On a campus, learning analytics is about students. At work, the learners are employees whose course data sits next to their performance, pay and discipline records, and that changes the risk.

Sharon Slade and Paul Prinsloo's [2013 paper on the ethics of learning analytics](https://doi.org/10.1177/0002764213479366), written for universities, names ethical challenges including "location and interpretation of data; informed consent, privacy, and deidentification of data; and classification and management of data." Their sociocritical view highlights "the role of power, the impact of surveillance, the need for transparency". Power and surveillance weigh more heavily when the organisation analysing the data also employs the people in it. George Siemens, writing in the [same 2013 journal issue](https://doi.org/10.1177/0002764213498851), warned that "privacy and data ownership will become increasingly important for all participants in analytics projects."

For the bus course, that turns into practical rules:

- **Tell drivers what is recorded and why** before the course launches: answers to each question, used to improve the course and plan practice.
- **Analyse by depot for course decisions.** The course owner does not need named results to rewrite a module.
- **Restrict named results to the driver, their supervisor and any authorised support role that needs them for the stated purpose.**
- **Do not join complaints to named quiz results** unless a written policy allows it and drivers know.
- **Set how long each record is kept,** and check your privacy obligations and any workplace agreements with whoever owns them in your organisation.

```json
{
  "_key": "754739890311",
  "_type": "calloutBox",
  "body": [
    {
      "_key": "9dbf73cb3568",
      "_type": "block",
      "children": [
        {
          "_key": "4779afd4b88b",
          "_type": "span",
          "marks": [],
          "text": "A wrong answer in a training quiz is a sign that the course or the practice needs work. If staff learn that quiz answers feed performance reviews or discipline, they have a reason to guess, look answers up or rush, and the data stops telling you what people know."
        }
      ],
      "markDefs": [],
      "style": "normal"
    }
  ],
  "markDefs": null,
  "title": "Do not turn course data into a performance file",
  "variant": "warning"
}
```

## How to use learning analytics to answer a training question

Run on any course, the same steps make one workflow:

1. Name the decision you might make and the person who would make it.
2. Write one question that would inform that decision.
3. Choose which of the four types the question needs.
4. List the data that answers it, and where that data lives: the LMS, an LRS or a job system.
5. Check the ethics: who will see the data, at what level of detail, and what learners have been told.
6. Run the analysis, then interpret it with someone who knows the work.
7. Communicate the finding to the person who acts, with the action and a date.
8. Check whether the change worked, using job data where you can.

Step 8 is evaluation. For how to plan it, the [Kirkpatrick model](https://www.coursebox.ai/blog/kirkpatrick-evaluation-model-it-actually-begins-at-level-4) treats on-the-job behaviour as Level 3, and the guide to [evaluating training effectiveness](https://www.coursebox.ai/blog/how-to-evaluate-the-effectiveness-of-a-training-program) covers the design of the check.

### The question planner, filled in for the bus course

The planner applies the workflow one question at a time. The figures and weeks come from the bus example; replace them with your own. The filled example and a blank version are in the download.

| Question | Type | Data source | Who acts | Action and date | How we will check |
| --- | --- | --- | --- | --- | --- |
| Which questions did drivers miss? | Descriptive | Course report | Course owner | Review answers to each question in week 8 | Answers by question |
| Are the misses concentrated at one depot? | Diagnostic | Course report by depot, fault calls by depot | Course owner with Harbour supervisor | Talk to Harbour drivers in week 9 | Supervisor's notes |
| Which drivers face the highest risk on a real fault? | Predictive | Diagnostic result, bus roster | Harbour supervisor | List drivers for a practical check in week 9 | Roster check |
| What change will help with the older ramp? | Prescriptive | Approved procedure for the older ramp | Course owner, procedure owner | Add steps and a practice question by week 10 | Fault calls and complaints by depot, weeks 10 to 17 |

**Learning analytics question planner**
A Word planner for turning a training decision into one question, the type of analysis it needs, the data, who acts and how you will check the result. Includes the filled bus-driver example and a blank version.
- [Download the planner (.docx)](https://cdn.sanity.io/files/3qyhkhue/production/7e6c4b3885182331a3f185641d6fe1026da1c584.docx?dl=)

When results come back, give feedback to learners first and quickly. A driver who missed the ramp-fault question needs the right steps before their next shift on an older bus, not a dashboard entry. If the question is whether drivers still know the steps months later, the guide to [knowledge retention](https://www.coursebox.ai/blog/knowledge-retention) covers delayed checks.

```json
{
  "_key": "832d0f9ac3cf",
  "_type": "calloutBox",
  "body": [
    {
      "_key": "49af022a2279",
      "_type": "block",
      "children": [
        {
          "_key": "43507258ef13",
          "_type": "span",
          "marks": [],
          "text": "A finding without an owner and a date is a report. In the planner, the \"Action and date\" column is the one that turns analysis into learning analytics, so fill it in before you run the numbers."
        }
      ],
      "markDefs": [],
      "style": "normal"
    }
  ],
  "markDefs": null,
  "title": "Put a date on the action",
  "variant": "tip"
}
```

## What course requests show about starting from data

Learning analytics begins with a question about how people are learning, and often with data the team already holds. I wanted to know how often training requests start there, so I looked at requests typed into the Coursebox course generator.

The sample covered 18 August to 15 November 2026: one first prompt per person, English browsers, staff excluded. An AI model labelled the requests, I collapsed identical prompts into one and removed prefilled example prompts that many people sent unchanged, and then I read every remaining hit by hand. That left 68 workplace course requests.

```json
{
  "_key": "63cf7102a600",
  "_type": "statCallout",
  "label": "workplace course requests described what learners currently get wrong; none mentioned how an earlier run of the course went or a business measure the course should move",
  "markDefs": null,
  "source": "Coursebox course requests, 18 August to 15 November 2026; one 90-day sample, English browsers, staff excluded, one request per person, duplicates and example prompts removed; AI-labelled and hand-checked; an upper bound",
  "value": "At most 1 of 68"
}
```

The one possible request listed behaviour problems the course should address.

A request describes the course someone wants built, so teams may hold data they did not type, and a classifier can miss wording. The finding cannot show that teams lack data. It does suggest a practical step: when you brief a course, write down what learners currently get wrong and how you know, because that sentence is the starting point for any later analysis.

## Learning analytics vs LMS reporting

The two are often used as if they meant the same thing. LMS reporting describes what happened inside your courses; learning analytics uses that and other data to explain a learning problem and change something.

| Question | LMS reporting | Learning analytics |
| --- | --- | --- |
| What it is | Standard reports on enrolment, progress, completion and scores | Analysis of learning data, interpreted and passed to someone who acts |
| Best for | Compliance records, chasing incomplete training and spotting obvious problems | Finding why people struggle and testing whether a change helped |
| Weak spot | Shows what happened, not why, and only inside the LMS | Needs a clear question, joined data and someone with time to interpret it |
| Use it when | You need a record or a quick status check | A report has raised a question you cannot answer from the report alone |

A healthy completion rate can still hide where a course is not working. [Ginda and colleagues' PLOS ONE study of the Boeing course](https://doi.org/10.1371/journal.pone.0215964) gives its own completion figure for the engineers:

```json
{
  "_key": "4fd4b159c256",
  "_type": "statCallout",
  "label": "completion rate for Boeing engineers in one MITxPRO online course run, fall 2016; all but 255 registered engineers passed (the paper gives 1,608 registered in its abstract and 1,611 in its methods)",
  "markDefs": null,
  "source": "Ginda, Richey, Cousino and Börner, PLOS ONE, 2019",
  "value": "84.1%"
}
```

The same logs showed learners used only 310 of the course's 551 lowest-level modules. The completion rate recorded who passed; the module data raised the design questions a report alone would not.

Start with the report. Move to analytics when a figure like the 48 missed answers in the bus course raises a question you need to answer. The [training ROI](https://www.coursebox.ai/blog/training-roi) guide covers the separate step of putting a money value on the result.

## Limitations of learning analytics

- **The evidence base for workplaces is thin.** When I looked for verified workplace cases with a measured result, most came from health professions training.
- **Patterns are not causes.** The authors of [the Taiwan study in JMIR Medical Education](https://doi.org/10.2196/95066) state that their design "precludes causal inference". Most workplace analysis is observational in the same way.
- **Combined programmes blur credit.** In [the Virginia inbox coaching study](https://doi.org/10.1007/s11606-026-10720-z), analytics and coaching were delivered together, so the study cannot show what either did alone.
- **How data is produced shapes what it says.** At the University of Alberta, [an analysis of 6,104 assessment forms for 186 family medicine residents](https://doi.org/10.4300/JGME-D-21-00832.1) found that supervisors who worked with a resident continuously recorded different skills from those who supervised occasionally.
- **More data is not more insight.** Logs grow quickly, and volume alone does not explain anything: [the Boeing study](https://doi.org/10.1371/journal.pone.0215964) counted nearly 31 million click records and still did not test whether any redesign helped.

## Using learning analytics with an AI course tool

I ran the bus course through Coursebox as a test on 16 November 2026. My prompt asked for a 20-minute course for 90 drivers on the accessible boarding procedure, asked for answers to be tracked question by question so results could be compared across depots, and asked the builder to tell me what it could not build.

The builder declined to write the procedure. Its reply said no ramp, wheelchair or failure steps had been supplied and that it should not invent safety-critical instructions. That is what I would want. Of the three routes it offered, I chose a course shell for my own content, and it built five pages with one knowledge check.

Then I opened the reports. The course report, under the course's More menu, lists each learner's completion percentage, with search, a status filter and a download. It has no group or depot filter. My shell had no learners yet, so I checked the per-course view on the Analytics page using an existing course in the account. It shows lesson completion and drop-off, highlights the biggest drop with a suggestion to split that lesson or add examples, and lists the hardest questions by first-attempt correct rate. That answers the descriptive question in the bus example: which question most drivers miss first time.

The Groups tab sounded like the place for depots, but it counts community posts, comments and likes, not training results. Comparing depots means downloading the course report and joining it to a depot list yourself, a step I did not test. The fault calls and bus types never live in the LMS at all.

My advice: before launch, keep a list of which depot each driver works at, ready to join to the export, because the depot split happens after the download.

**Learning data inventory and ethics check**
A Word worksheet for listing each source of learning data, who can see it, how long it is kept and what it is used for, with ethics questions adapted for employee data. Includes the filled bus-driver example and a blank version.
- [Download the inventory (.docx)](https://cdn.sanity.io/files/3qyhkhue/production/db615da7984e0fd569b79564f370af5d322b91df.docx?dl=)

## Putting learning analytics to work

Learning analytics is not a dashboard or a tool category. It is the habit of asking a specific question about how people are learning, answering it with the data you hold, and getting the answer to someone who will change the training. If you run courses in an LMS, you already hold the first ingredient in its reports, and your job systems often hold the evidence of whether the training worked. Pick one course where completion looks fine but the work has not improved, write the decision you might make, and fill in one row of the question planner before you open the report. Then set the date when you will check whether the change helped.
