Wegner (1987) introduced the idea of transactive memory: a group remembers more than its members individually because each knows which of the others holds what. Lewis (2003) developed a field measure of such systems built on specialisation, credibility and coordination. The implication is that expertise, for organisational purposes, lies in being recognised and consulted as well as in what a person knows. Borgatti and Cross (2003) specified the conditions under which one person seeks information from another: knowing what the other knows, valuing it, being able to gain timely access and judging the cost of asking to be acceptable. A map of expertise must therefore capture a directed relation, from the person who asks to the person who is asked, and should distinguish the topic.

Survey measurement of networks has a known weakness. Bernard et al. (1984) reviewed studies comparing reported with observed communication and concluded that informants' accounts of whom they had interacted with over a given period were seriously inaccurate. Marsden (1990), surveying network measurement more broadly, noted that respondents report stable, typical relations better than specific episodes, and that recognition from a list yields more complete data than free recall. Surveys are thus poor at recording the volume and timing of contact and comparatively good at recording judgements, such as whom one regards as an authority. Their other limitation is burden. A roster covering every colleague grows with the square of group size, and response rates suffer.

Trace data reverse this profile. Digital systems keep records of interaction as a by-product, and organisational researchers have used such records to study structure at a scale no survey could reach (Yang et al., 2022). Comparisons of the two sources are instructive. Quintane and Kleinbaum (2011) and Wuchty and Uzzi (2011) each compared e-mail records with self-reported ties and found a real but partial correspondence: traces recover the relations people report, but they also contain much routine contact that respondents do not regard as a relationship. Traces record that an exchange took place. They are silent on whether it was a request for advice, an instruction or an administrative notice, and they omit whatever happens face to face.

The exercise begins with a written statement of purpose, which should be limited to understanding and supporting the flow of expertise and should exclude the appraisal of individuals. Employee representatives and the data protection officer are consulted before any data are drawn. Within the European Union the General Data Protection Regulation applies: the principles of purpose limitation and data minimisation in Article 5, the need for a lawful basis under Article 6, and the data protection impact assessment required by Article 35 for processing likely to result in a high risk (European Union, 2016). Borgatti and Molina (2005) observed that network studies in organisations cannot offer anonymity in the usual sense, since the data consist of named relations, and recommended an explicit agreement with management on what will be disclosed and to whom. Such an agreement is made here.

Only header information is extracted: pseudonymised identifiers of sender and recipients, a timestamp and the channel. Message bodies, subject lines and attachments are never accessed. A recent window of some months is taken, long enough to smooth over holidays and project peaks. Messages to large distribution lists, automated notifications and meetings above a small size are removed, because they signify no relation between the parties. The key linking pseudonyms to persons is held separately by a custodian who is not an analyst.

The cleaned records are reduced to counts of contact between each pair of people. A tie is retained only where contact is reciprocated and recurs across several weeks, a filter intended to remove the routine traffic that the comparison studies found to be unrelated to reported relationships (Wuchty and Uzzi, 2011). The thresholds are set after inspecting the distributions and are recorded. The result is an undirected, weighted network of working contact. It says nothing yet about expertise.

Each participant receives a short questionnaire listing the colleagues with whom the network shows regular contact, with space to add others. Two or three questions follow, adapted from the information and advice items used in organisational network surveys (Cross and Parker, 2004): to whom the respondent turns when stuck on a problem in a named domain, whose judgement the respondent would want before a consequential decision, and, optionally, who has helped the respondent to learn the job. Domains are chosen with the business in advance and rotated between waves so that no single wave is long. Participation is voluntary, and a respondent may decline to be included in the map as a nominee.

Survey nominations are directed ties labelled by domain. They are laid over the interaction network. Three kinds of tie result: those present in both layers, which are the best attested; nominations with no trace in the metadata, which usually indicate advice sought in person; and frequent contact without nomination, which indicates coordination and not expertise. Keeping the three distinct is the main analytical gain over either source alone.

Standard measures are then applied to the directed advice layer (Wasserman and Faust, 1994). In-degree, the number of colleagues who name a person, identifies recognised experts in each domain. Betweenness, in the sense defined by Freeman (1978), identifies those who connect groups that would otherwise not reach one another, the brokers who span what Burt (1992) called structural holes. Community detection on the interaction layer, for which the method of Blondel et al. (2008) is widely used, shows clusters of working contact, and comparing them with the advice ties reveals groups that depend on a single outside expert or that consult nobody beyond themselves.

Before anything is reported, the patterns are shown to small groups drawn from the population and they are asked whether the map matches their experience. This step guards against artefacts of the thresholds and catches experts whom neither layer has registered, such as a long-serving technician consulted entirely by word of mouth. It also honours an obligation: people whose relations have been charted are entitled to see what has been made of them.

Reports to management are aggregated to units and domains. Individuals identified as heavily consulted are told so privately and are named to management only with their consent. Small groups are suppressed or merged, following the reasoning of Sweeney (2002) that a record is protected only when it cannot be distinguished from a sufficient number of others. Actions that the map may support include pairing an overloaded expert with a successor, introducing isolated groups to relevant experts and recording knowledge held by a single person. The exercise is repeated at intervals, since networks shift with projects and staffing.