Observe invisible work
16 min read
Lesson objectives
- Design a bounded observation of invisible work, calculate transferred effort and test a competing explanation while protecting worker and customer data.
Please sign in to use all the tools
Once you sign in, the reading tools are yours: larger or smaller text, typeface, line spacing, focus mode, highlighting and bold, and notes saved to your account and gathered in My study.
You arrive at a service center in the morning and see a dashboard reporting healthy response times. Sitting beside an employee for ten minutes reveals two screen transfers and a request for help interpreting an exception. Those minutes are missing because they occur before the ticket is opened. The manager asks whether a drafting assistant will solve the problem. Before answering, you need to see the work as it happens, including search, waiting, correction and informal help. Observation provides evidence about task design. Its purpose is to improve the process in which people work, with a clear explanation of what you record and how it will be used. Agree a limited observation period and collect only the information needed to understand the workflow. Choose a task with visible start and end points, then observe where effort moves and who bears it. A purchase becomes one option to test, and its expected benefit becomes tied to work you can describe.
Start with a narrow field question: where does employee time go between receipt of a complete request and approval of a correct reply? This differs from asking about employee satisfaction or model quality. Define complete receipt and how approval becomes observable. If a document is missing, record document waiting separately from active work. If the employee stops to help a colleague, record the help as an activity serving another request. Use categories that observers can recognize: reading a document, searching guidance, requesting interpretation and editing a response. Have two observers classify one example before fieldwork. Where they disagree, revise the definition and try again. Agreement reduces variation introduced by observers, but the category scheme remains a local practical decision. Review it when work appears that the initial categories failed to cover.
Choose between following a request end to end and taking time snapshots. Following a request reveals handoffs, sequence and waiting between people. It may take days if the request stays open. A time snapshot records the visible activity at a selected instant and helps estimate how work is distributed across the day. It can miss the reason for the activity and what came before it. Combine the methods if useful: snapshots across shifts, followed by selected requests to explain causes. Choose observation times before seeing results. Quiet periods alone produce a different picture from peak demand. Volunteers alone may overrepresent confident or highly skilled employees. Cover different roles, shifts and request types, and name the groups you could not observe. Findings describe the work actually sampled. Confidence improves when independent sources support the same account.
Separate effort from elapsed time. An employee drafts for six minutes, waits an hour for approval and edits for two minutes. Active effort is eight minutes; elapsed time is sixty-eight minutes when the full hour falls between those activities. An assistant that removes three drafting minutes saves effort. It may reduce service time by only those three minutes if the rest stays unchanged. Giving approval authority to an available supervisor may affect customer waiting much more. Draw a timeline for each followed request. When two requests overlap, assign active effort to the request being served. Keep waiting visible without counting it again as employee effort. This distinction helps answer both finance questions about capacity that can be reassigned and service questions about improvements customers will notice.
Manage the effect of being observed. Employees may explain more, postpone personal-tool use or select easier cases in front of you. Explain the purpose, recording boundaries and who will read the notes. Describe tasks rather than naming individuals when that is sufficient. Agree how a serious incident discovered during observation will be handled; organizational policy may require immediate containment. After a case, ask why a particular step occurred and separate the explanation from what you directly saw. “Opened three screens” is an observation. “The system causes distraction” is an interpretation that needs testing. Preserve the distinction so colleagues can challenge an explanation without denying the event. If your presence changes behavior, record that limitation and seek logs or interviews to test the picture.
Look for invisible help. An experienced employee may resolve colleagues’ questions between calls. Others then appear highly productive while the expert looks slow. Ask which questions keep returning and record anonymized examples. If they retrieve a documented rule, better search may help. If they involve balancing exceptions or financial responsibility, delegation or judgment training may be more appropriate. An assistant can shift this balance: staff produce more drafts and ask the expert to verify more of them. Track consultation volume, duration and backlog during a pilot. Tie expansion to available review capacity and a trained substitute for absence. The adoption decision then captures transferred workload instead of recording a saving in one department and leaving the added burden elsewhere invisible.
Set boundaries that protect data. You usually need request type, workflow stage, approximate duration and a coded waiting reason. Names or screenshots may be unnecessary. If a document copy is needed, state why, who stores it and when it will be disposed of under organizational policy. Show staff sample observation notes in advance. Sensitive steps can sometimes be explained using a synthetic case. Process-improvement notes should not silently become individual appraisal. A changed purpose needs an explicit decision and appropriate review. Aggregate management reporting at task level when that meets the purpose. Restrict access to exceptional case detail when investigation requires it. Observation becomes easier to interpret when people understand the design question and the boundaries governing their information.
Compare observation with system records and the employee account. A system may show quick dispatch while an employee describes a later correction made after an off-platform call. Ask for the sequence and date, then verify the correction through an authorized record. Conversely, an employee may describe a long search whose minutes include screen-loading delays. Each source reveals one part and misses another. Record agreements and disagreements and specify the next question for each conflict. When evidence cannot be checked, call the conclusion a hypothesis. A manager can fund a small measurement exercise rather than buy a tool on one story. A simple, low-harm, reversible repair may also be reasonable, provided follow-up tests its effect. Match the commitment to evidence strength and the consequences of an incorrect explanation.
Make weighting reproducible. Suppose morning receives seventy percent of requests and evening thirty. Drafting appears in half the morning snapshots and a quarter of evening snapshots. The weighted estimate is forty-two and a half percent: multiply one half by seventy percent and add one quarter multiplied by thirty percent. The unweighted mean is thirty-seven and a half percent and treats the shifts equally. Choose weights for the question you wrote and state the assumption that request volumes remain stable. If complexity differs, compare request classes within shifts as well. When a reasonable alternative weight changes your conclusion, investigate further. That sensitivity belongs in the decision before the team converts an activity estimate into a large financial promise.
Before recommending action, test a competing explanation. If search appears slow, ask whether the underlying policy is contradictory. If you suspect insufficient skill, try a clear source with another employee before commissioning organization-wide training. One defect can produce several visible activities: someone may repeatedly read a document because they fear signing without authority. Clarifying authority addresses the nearer cause; faster summarization addresses a symptom. Ask a colleague to derive another explanation from the same facts, then choose a small observation that distinguishes them. This challenges attachment to a preferred tool and turns a decision meeting into a comparison of testable actions. Record the evidence that would change your recommendation so a later result can prompt a coherent revision.
The NIST AI RMF Map function supports understanding purpose, deployment setting, user expectations and affected parties. Measure supports methods suited to evaluation questions and their limits. This lesson’s observation protocol, weighting and time categories are an instructional design based on those ideas. Adapt details to your organization, worker rights and data policy. On your next workday, take one request you are authorized to examine and map what happens before its ticket opens and after it closes. Ask its owner about a step the system misses and identify a safe way to verify it. Show the manager one difference between the record and the work, with a small measurement that tests its cause. That concrete output adds information an impressive technical demonstration may overlook, however persuasive its promises.
Before comparing observers, verify that they observe the same unit. One may record the whole request’s approval wait while another records only the employee waiting before moving to another task. These definitions answer different questions. Use one clock for the case and one for active work, showing how each starts and stops. Review three shared observations and record disagreements and causes. If an observer misunderstood a definition, coach them and reclassify before combining data. If the disagreement reveals a new activity, revise the dictionary and explain how that affects earlier comparisons. Keep a temporary category for activities needing explanation rather than forcing ambiguous events into a convenient box. Review that category daily until it yields specific knowledge. Use a synthetic calibration example when a real event would reveal personal details beyond the observers’ need.
Prepare a report with three connected layers: decision question and observation limits, findings with denominators, and a testable next step. Include one mechanism-revealing case and one counterexample so the reader can see what the interpretation may miss. Name missing data and their decision effect; unrecorded help makes a saving uncertain. Match the resource request to the stage: another observation day, a limited integration repair or an assistant trial on a specific task. Identify the owner, review date and stopping evidence. This allows a manager to act proportionately now and revisit results later. It also lets employees check that their work was represented fairly and that uncertainty was preserved.
Lesson summary
Observation captures effort, handoffs and hidden help while separating events from interpretations before selecting a repair or tool.