Guide
Before an AI project, map the work your teams actually do, not just the documented process: combine interviews, a survey and daily check-ins, and sort time into six task families, each with its margin of error. The hours an agent could free remain to be confirmed, written down task by task with an owner.
Key takeaways
- A process map says how work is meant to happen; before an AI project, you need to know where the time goes.
- Combine three methods: interviews explain, a survey covers the whole team, daily check-ins quantify.
- Sort everything into the same six task families, and give each share with its margin of error.
- Freeable time is still to be confirmed: write it down task by task, with its source and its owner.
- The map describes tasks and roles, never an individual’s performance.
Documented process or actual work: why the official map is not enough
Learn more: documented process or actual work: why the official map is not enough
A process map describes the intended sequence: who does what, in what order, with which tool. It does not say how long each step takes, or what teams do around it: re-keying data between two tools, chasing, checking, workarounds. Yet that is precisely the time an AI agent can take over.
So before an AI project, the question is not “what is our process?” but “where does the time go, task by task, and how sure are we?” The answer you need has two parts: a breakdown of time by task family and by role, each share with its margin of error; and a list of concrete tasks, with their frequency, duration and volume.
Knowing whether the company uses AI does not answer that question. In 2025, 18% of French enterprises with 10 or more employees (excluding agriculture, finance and insurance) said they used at least one AI technology, and 15% of those with 10 to 49 employees (Insee). That tells you who uses a tool, not where it saves time.
| Aspect | Process map | Map of the actual work |
|---|---|---|
| What it describes | The intended sequence of steps | The time actually spent, by task and by role |
| Where it comes from | Procedures, process owners | The people who do the work |
| What it shows poorly | Re-keying, waiting, workarounds | The exact order of the steps |
| What it is for before an AI project | Setting the scope | Sizing the time at stake and choosing the tasks |
Three ways to measure: interview, survey, daily check-ins
Learn more: three ways to measure: interview, survey, daily check-ins
No single method is enough. Interviews explain, a survey covers the whole team, daily check-ins quantify. Their biases differ, which is exactly why you combine them.
Time reported from memory has known limits. Writing about weekly working hours, Robinson and Godbey, cited by Jacobs in the Monthly Labor Review (1998), note that people work it out in a few seconds, may give the answer they think is expected, and are not always sure what counts as work. Daily check-ins sidestep memory: at moments spread through the day, you record what the person is doing at that instant, and the share of observations in each category gives the share of time, within a margin.
A bias remains when people record their own activity. In a study of nurses at an Australian hospital, documentation accounted for 19% of self-reported observations but 8% of observations made by a third party (Ampt et al., 2007). Third-party observation is more reliable, but it needs an observer on site for the whole period.
| Method | What it measures | Main bias | Cost |
|---|---|---|---|
| Guided interview | Tasks, tools, pain points; for each task, frequency × duration × volume | Self-reported and from memory; depends on who is interviewed and how the interview is run | Highest per person: one interview, then a reviewed summary |
| Survey | How a typical week splits across six families (100%), perceived drudgery, volumes, for the whole team | Given from memory, rounded; answers people think are expected | Low: a few minutes per person |
| Daily check-ins | An approximate split of time across families, with a calculated margin of error | Being self-reported, it still depends on what people choose to record, and on when they are asked | A few seconds per observation, but repeated throughout the collection period |
Six task families to sort what you observe
Learn more: six task families to sort what you observe
To add up interviews, survey answers and observations, you need the same categories everywhere. SHIFTLABS uses six, shared by all three measures. Daily check-ins add a seventh, “break / not working”: without it, every observation would be forced into a family.
A review of the literature on these check-ins (known there as work sampling) notes that categories, and the training of the people who record, vary from one study to the next, which makes results hard to compare (Blay et al., 2014). Fix the families before you start, and give everyone examples of what belongs in each.
| Family | Example tasks | What to record |
|---|---|---|
| Data entry and checking | Re-keying orders, matching invoices, checking fields | Daily volume, time per item, source and target tool |
| Information search and consolidation | Finding a document, compiling figures from several sources | Number of sources, frequency, time per search |
| Communication and meetings | Internal email, meetings, team check-ins | Recurring share, share that only passes information on |
| Reporting and documents | Dashboards, minutes, periodic reports | Frequency, whether a template exists, time spent on formatting |
| Customer and supplier relations | Answering requests, follow-ups, quotes | Volume, share of repetitive requests, response time |
| Expertise, analysis and decision | Preparing a case, analysing, arbitrating, advising | Share of time spent preparing rather than deciding |
From time spent to hours freed: the figure to confirm, put in writing
Learn more: from time spent to hours freed: the figure to confirm, put in writing
The map measures time spent. It does not measure the time an AI agent will free: that figure remains to be confirmed, and it must be presented as such.
Getting from one to the other takes three lines, each with its status.
- Time spent per family = measured share × hours worked by the team. Status: measured, with its margin.
- Freeable share = the part of that time an agent can take over, task by task. Status: assessed, by the business owner and an expert, from the tasks described in interviews.
- Hours actually redeployed = freeable hours × realisation rate. Status: unconfirmed until measured after go-live.
What the map must not become: a tool for assessing individuals
Learn more: what the map must not become: a tool for assessing individuals
A map of the work describes tasks and roles, not people. If teams believe their answers will be used to judge their productivity, they will give the answers they think are wanted, and the map will be wrong.
Say so before the first question, and stick to it in practice. The trust centre sets out what the employer receives, with the source of each statement.
- The unit of analysis is the task and the role. Opportunities target tasks, not jobs: a SHIFTLABS interview summary is written without any HR recommendation or recommendation to cut a job.
- Summaries passed to the employer are pseudonymised: linked to the stated role, without the first name, the full conversation or the interview date, reviewed by the employee and then validated by a SHIFTLABS expert.
- Nothing is passed on below a threshold: 5 validated interviews for summaries, 5 validated responses for survey and daily check-in averages. Free-text survey answers are not passed to the employer.
- In a small team, these safeguards are not always enough: a role can be enough to identify a person. In that case, group roles together before reporting back.
- The “break / not working” category is there to keep the shares accurate, not to count anyone’s breaks.
A blank grid to fill in
Learn more: a blank grid to fill in
One row per family, for a role or a team. Fill in the measured share first and only then the freeable share: done the other way round, you end up looking for figures that confirm your initial hunch.
| Task family | Share of time (measured, 95% margin) | Concrete tasks: frequency × duration × volume | Freeable share (value used, source, owner) |
|---|---|---|---|
| Data entry and checking | …% (…% to …%) | … | …% · source: … · owner: … |
| Information search and consolidation | …% (…% to …%) | … | …% · source: … · owner: … |
| Communication and meetings | …% (…% to …%) | … | …% · source: … · owner: … |
| Reporting and documents | …% (…% to …%) | … | …% · source: … · owner: … |
| Customer and supplier relations | …% (…% to …%) | … | …% · source: … · owner: … |
| Expertise, analysis and decision | …% (…% to …%) | … | …% · source: … · owner: … |
| Break / not working | …% (…% to …%) | Not applicable | Not applicable |
Frequently asked questions
What is process mapping for an AI project?
It is a description of the work actually done within a scope: the tasks, the time they take by role, their frequency and their volume. It is used to choose where an AI agent will take over time, and later to check that it has.
Is a survey enough on its own?
For a first order of magnitude, yes. But a breakdown given from memory is still approximate: for reported working hours, the calculation is known to be done quickly, the answer may be the one people think is expected, and the boundaries of work are blurred (Jacobs, 1998). Use daily check-ins to verify the families that weigh on the decision.
How many observations do daily check-ins need?
It depends on the precision you need. The margin shrinks with the square root of the number of observations: you need roughly four times as many to halve it. For a share of 24%, the 95% interval runs from 16.7% to 33.2% with 100 observations, and from 20.1% to 28.4% with 400 (Wilson interval). Below about a hundred observations, read the shares as an order of magnitude.
Are self-reported daily check-ins reliable?
More than memory, less than third-party observation. In the study by Ampt et al. (2007), documentation accounted for 19% of self-reported observations against 8% of third-party observations, while other categories matched. So read each share with its margin, cross-check it against the interviews and have discrepancies validated.
Do you need a process mining tool?
Not to start with. Process mining rebuilds a process from a system’s event logs: it only sees what those logs record, not email, spreadsheets or phone calls. At SHIFTLABS, it can be added for larger scopes, by importing a CSV export of the event log, extracted by your IT department.
How long does mapping the work take?
It depends mostly on the daily check-ins, which have to run long enough to gather the number of observations you need; interviews and the survey run during the same period. At SHIFTLABS, all three methods fit within the 2 weeks of data collection of the AI assessment (from 50 employees, allow about 1 more month beforehand for the works council information and consultation), and each conclusion is validated before the readout.
Sources
The sources (4)
- Insee, Insee Première no. 2120 (21 July 2026), “Les technologies de l’information et de la communication dans les entreprises en 2025”, accessed 18 October 2026
- Jerry A. Jacobs, “Measuring time at work: are self-reports accurate?”, Monthly Labor Review, U.S. Bureau of Labor Statistics, December 1998, accessed 18 October 2026
- Ampt, Westbrook, Creswick, Mallock, “A comparison of self-reported and observational work sampling techniques for measuring time in nursing tasks”, Journal of Health Services Research and Policy, vol. 12, no. 1, 2007 (Macquarie University record), accessed 18 October 2026
- Blay, Duffield, Gallagher, Roche, “Methodological integrative review of the work sampling technique used in nursing workload research”, Journal of Advanced Nursing, 2014 (University of Canberra record), accessed 18 October 2026
Further reading
Keep reading
- AI use cases in business: how to find and prioritise them
- ROI of an AI project: how to calculate the gain and the payback period
And in your organisation?
An AI agent interviews your teams about their work and an expert validates every deliverable. You know where AI will save you time, with costed use cases and a roadmap.