Analytics and expert collaboration: How individuals navigate relationships when working with organizational data

Access

Collaboration among experts depends on their meaningful access to each other. Recognizing access as a prerequisite acknowledges the relational nature of expertise. Organizational knowing emerges in problem solving communication among organizational members (Canary and McPhee, 2011; Cross and Sproull, 2004; Kuhn and Jackson, 2008). Experts are not experts just because of what they know but because of what other team members think they know, and being an expert involves the practice of expertise in front of different audiences (Barbour et al., 2016; Leonardi and Treem, 2012; Treem, 2012). Expert relationships are not merely a basis for information exchange, but knowledge-intensive work depends on them being able to reach each other and exchange information and ideas (Carmeli et al., 2015; Fu, 2015; Hollingshead and Brandon, 2003).

Trust

Trust in others’ expertise is also integral to collaboration among organizational experts (Huang et al., 2013; Sankowska and Söderlund, 2015). In fact, disincentives for sharing information (Hollingshead et al., 2011), such as fears that knowledge shared may be unacknowledged or misused, suggest that collaboration among organizational experts is fragile (Empson, 2001). Exploration, reinterpretation, reconceptualization and the like involve not just having access, but also relationships that are thought to be to be close and trustful (Mom et al., 2015; Tsoukas, 2009).

Connection

Likewise, relationships with organizational experts should enable opportunities to engage in rich conversations. We conceptualized connection as the ability to have difficult conversations and develop shared meaning that should in turn enable organizational learning (Barge and Little, 2002). Tsoukas (2009) argued that relational engagement in dialogue involves a suspension of assumptions that leaves actors open to influence and thereby supports the conceptual combination, expansion and reframing needed for knowledge creation. Robust connection should make possible the sort of generative jamming that Eisenberg (1990) argued is rare, valuable and can result in novel approaches to problems.

In summary, we expected that the rigor of participants’ interaction with experts and the trust they held in them would enhance their ability to generate interpretations of data. We hypothesized that:

Furthermore, we expected the effects of these relationships to compound one another such that connection and trust would increase the strength of the relationship between expert access and optioning. We therefore hypothesized that:

To explore the functioning of these relational dimensions in practice, we also asked:

Project autonomy

Traditional uses of data allow organizations to make decisions within particular functions or units (e.g. finance, R&D, marketing), thereby also establishing managerial control over those units’ decisions. Suffusing analytical capability across functional boundaries has the potential to throw this control into disarray (Liberatore and Luo, 2010). Functions with important data or data acumen may be asked to work on new problems where they have not previously had authority to make decisions. Units without data or data acumen may need to seek them from sources where they have no authority or influence. For example, in a study of the use of analytics in news rooms, Hanusch (2016: 13) found instances of useful application, but also ‘signs of accelerating processes of functional differentiation within journalism’ that also involved changing journalistic practice and audience news values.

Conceptualizing analytics as a new organizational function, brings to the fore the practice of analysts who have a distinctive expertise (Treem, 2012) and identity (Alvesson, 2001) that centers on mastery of the professional work associated with doing analytics well (Davenport and Patil, 2012). Like other professionals, these analysts may need freedom to exercise their professional craft as they see fit, and doing so would be in part what would mark them as having distinctive organizational expertise (Leonardi and Treem, 2012; Treem, 2012). Data analysts may need to be able to look at data in ways that are guided by their own judgment—to explore (Mom et al., 2015; Tsoukas, 2009). They may need to be able to ask questions and make calls that other parts of the organization may see as unorthodox or risky (Davenport and Harris, 2007). We conceived of this freedom in this study as project autonomy, namely, the extent to which individuals could exercise their own judgment in their work with data and ask what might seem like unorthodox or unusual questions. We hypothesized that:

Furthermore, relationships with experts should also support optioning in part by enabling project autonomy. For example, Cross and Sproull (2004) found that information relationships supported the validation and legitimation of particular approaches to problems. Carmeli et al. (2015) found that relational information processing supported ideation and creativity especially when it fostered openness and acceptance, creating space for unorthodox questions. Relationships with experts should allow for optioning in part by allowing the participant the freedom to approach organizational problems as they see fit. We hypothesized that:

and we asked:

Procedures

Reliability and validity

To facilitate reliable and valid measurement, existing measures were used when appropriate, and new measures were used with careful attention to relevant theory. Furthermore, early drafts of the questionnaire were reviewed by leaders at FSC and by a pair of scholars and practitioners familiar with FSC and analytics. FSC senior leadership limited the number of questionnaire items as a condition of access, but the reliability indicators for the multiple item measures reached orthodox requirements (see Table 1). The use of mixed methods also provided a useful check of the validity of the measurement approach.

OPEN IN VIEWER

Table 1.

Descriptive statistics.

	Mean	SD	N	α	2	3	4	5	6	7	8	9
1 Optioning	4.74	0.68	53	−	.62	.31	.41	.31	.20	.32	–.03	–.04
2 Project autonomy	5.06	0.62	54	0.73		.40	.51	.10	–.03	.32	–.14	–.04
3 Access	4.75	0.88	54	0.70			.51	.40	.14	.41	.03	.09
4 Connection	4.93	0.72	52	0.76				.63	.15	.43	–.01	–.02
5 Trust	5.18	0.73	52	0.79					.17	.40	.08	−.39
6 Interdependence	5.22	0.84	54	−						–.05	.06	.30
7 Data accessibility	3.30	1.16	54	0.90							.38	–.11
8 Time scarcity	4.53	1.25	54	0.82								–.05
9 Comm. frequency	75.50	60.12	53	−

SD = standard deviation; Comm. frequency = communication frequency. Index means, standard deviations, number of participants responding, Cronbach’s alpha, and zero-order correlations. Bolded correlations are significant (p < .05); correlations in italics are not.

Interview data

After the completion of the second wave, the first author conducted follow-up, in-depth interviews with the participants from the first and second waves. The interviews took place approximately seven months after the survey data collection had concluded. All participants from the first and second waves were contacted by phone and email to participate in follow-up conversations about their work with data. Twelve individuals (27.9% of waves 1 and 2) participated in these interviews. Participants in the third and fourth waves were not available to the research team for interviewing. Interviews from the first and second waves were scheduled after participants completed the questionnaire to avoid effects on the survey results and to help make the survey data collection procedures consistent across waves. As a condition of research access, recording the interviews was not permitted. Instead, the interviewer took notes to capture responses and key phrases, verbatim when possible, and notes were later elaborated with the goal of reflecting the responses as completely as possible (Emerson et al., 2011).

To protect the identities of the participants and FSC, we used pseudonyms for participants, organizational divisions and FSC; replaced specific descriptions with more generic language; and obscured details that might be revealing. Throughout each wave and during interviews with participants, the first author took notes (15 handwritten pages during observations and 42 pages during interviews). The interviewer asked participants to reflect on their work with data in general and on the specific projects that were the focus of the survey.

The semi-structured interview guide consisted of three open-ended questions (‘How do you use analytics in your work?’, ‘What are the dilemmas you experience in this work?’ and ‘How do you address those dilemmas?’), and during discussions each question was accompanied by follow-up questions designed to elicit abstract and narrative accounts of their experiences with data. We ceased recruitment efforts when we observed a sharp decline in novel accounts of problems and strategies. Interviews lasted approximately an hour (median ≈ 57 minutes), and ranged from 40 minutes to more than two hours.

The structure of the interview questions was meant to reflect a focus on problems and dilemmas in their work with data, their strategies for solving problems, and their reasoning on why a particular strategy would address a given problem (Craig and Tracy, 1995). This approach, in combination with the survey data, was designed to shed light on how, participants employed communicative and relational strategies, techniques or formats in their collaboration.

Analysis approaches

The hypotheses were tested using the survey data. We used procedures detailed by Hayes (2013), specifically, ordinary least squares (OLS) regression with bias-corrected bootstrap resampling. For hypotheses focused on direct relationships, we report the zero-order correlations and the coefficients from the most complex conditional process modeling. For the tests of mediation, we report bootstrapped coefficients, standard errors, 95% confidence intervals (CIs), and the R² of the model of direct and indirect effects to indicate explanatory power.

To address the research questions, we conducted an iterative analysis of the interview data, using open and focused coding procedures. Guided by the research questions, the first rounds of open coding identified participants’ accounts of the problems participants experienced in working with data and in the special projects under study and then strategies they employed to address them (Craig and Tracy, 1995). The first author presented these open codes with examples to the research team, and in concert with the results of the hypothesis testing. Through discussion, the research team categorized these codes by comparing each example with the emerging categories, adding categories as needed until the open codes had all been assigned. Throughout this discussion, the first and third authors contributed additional observations from their experiences working with the participants. During the discussions, the first author took notes about emerging themes, and later produced research memos describing the categories of problems, strategies and the underlying ideas about analytics and communication implied in their responses. The research team met again to discuss the logic underlying the relationships between problems and strategies in the data, and the first author produced another research memo capturing that conversation. The team discussed and elaborated upon the memos in an additional analysis meeting, and with these research memos in hand, the fourth author reviewed the notes looking for negative cases and alternative examples. Given that the quantitative analysis preceded the qualitative analysis, a distinct effort was made to seek data that would challenge the results, in order to help account for any selection bias that may have existed in the coding. The team reviewed and refined the memos to answer the specific research questions and to elaborate and clarify the survey results.

RQ1: Working with data and optioning in analytics

Overall, most participants agreed that they had been able to develop options for interpreting information in their projects (M_optioning = 4.74, SD = 0.68, see Table 1). They reported engaging with organizational experts in this project work (approximately 6–7 times per month, M_{commfrequency} = 6.29 times per month; SD = 5.01), and tended to agree that this work depended on connecting with organizational experts (M_{interdependence} = 5.22, SD = 0.84). Thus, there was evidence that optioning was a facet of analytics work and occurred through interactions with experts across the organization. In describing the work involved in analytics, the participants highlighted challenges they encountered in their work with data for their specific projects and in general, including the accessibility of data, the volume and complexity of data, the pace of their work with data. They also mentioned dilemmas that were interwoven with autonomy in particular such as existing agendas for data collection and analysis, hierarchies and data ownership, and legal and regulatory frameworks relevant to their work with data.

These points are illustrated by some of the specific quotes captured during the interviews. For example, analytics work could not occur if relevant data were inaccessible. As Charley explained ‘data is the latest buzz around here,’ but he later added that before asking any questions or conducting analysis, ‘there’s gotta be data . . . Some don’t exist . . . we’ve never asked for it. Some we won’t collect.’ Participants’ accounts confirmed the importance, but also the difficulty, of getting to data in the first place (M_{dataaccessibility} = 3.30, SD = 1.16). Participants described the problem of accessibility in terms of the absence of data and of difficulty locating or retrieving existing data from ‘legacy systems’ only understood by a few. Walter explained, ‘The data is out there, but you have got to ask for it.’ Accessing data required participants to find and communicate with organizational members outside of their respective business units, and with whom they were unfamiliar. This effort could be problematic because there were not always people dedicated to answering such requests, and, as Alex commented, seeking data from another unit or finding the owner of data was viewed as ‘extra work.’

Participants also explained that having voluminous data could itself be problematic. Weber noted, ‘we have too much data, and most of it is useless. We get lots of reports that are interesting, but I can’t use them to change behavior.’ He explained that sales managers ‘have access to a lot of data, but they don’t use it.’ Even in instances where they were accessible in a material sense, data could be functionally inaccessible if workers could not interpret, organize or apply them. Rebecca recalled meetings with one department where they enthusiastically told her that they have a ‘ton of stuff,’ but then added that accessing the data required a particular technical skill. Later in the interview, she told the story of a program that had been generating data for 10 years, but with a glitch: ‘It took a skilled person six weeks to diagnose the problem . . . she’s since retired . . . when you pull data from so many sources, there will be misfires.’ Even when they could access the data, there were challenges of interpretation and application.

Another obstacle participants confronted in their analytics work was the existence of hierarchies, siloes and feelings of data ownership that made working with data more difficult. FSC has many thousands of employees, and is an organization of staggering complexity. Participants explained that hierarchies made finding the right person and getting access permission more difficult. Data were housed in many places across the organization, and different business units had different approaches to data collection. Participants noted that differences emerged because (a) the units wanted answers to related but slightly different questions, (b) they used different language for the same data, and (c) reconciling differing approaches involved organizational politics.

Given this context, analytics work and optioning involved digging deeper into data to move beyond straightforward readings of analyses and to see business problems in new ways. Weber explained that one reason standard data reporting units ‘cannot produce useful information,’ is because ‘some people don’t know what they want.’ He felt that part of his job was to see problems beyond a single unit. Analytics work required communicating about a problem in a way that other business units could understand. Tyler recalled that when he picked up a data-intensive project from his predecessor, after a few missteps, he ‘met with [the business partner] to find out what they really needed . . . what they needed was not the trend but the [reason for the trend].’ Optioning involved not only initial approaches to problems, but also the ability to rethink problems as data access shifted. Describing his work on a special project, Walter explained that ‘[a]s we got more and more data, we realized [our initial way of seeing the problem] was not necessarily true.’ His work was dependent upon the ability to alter initial assumptions and potentially redefine the problem they were addressing.

Optioning did not occur in isolation, but within a context where workers were aware they depended on others for data provision and interpretation. Therefore, individuals conducting analytics needed to communicate about problems in a way that would minimize conflict with others in the organization. Frank explained he ‘had to do the problem defining work with the experts,’ because they might not know him, might not think he knows what he is doing, and might not understand his part of the business. He had to work with them on making sense of the problem to dig deeper and to develop a shared understanding of the questions to be answered.

In summary, analytics work at FSC required more than just getting access to data and analyzing it to come up with straightforward answers to business questions. Participants described doing this sort of more straightforward work, but they also described needing to define the problem, identify the relevant data, make the case for using it, and finally to perform the appropriate analyses. Optioning was critical in analytics work because it allowed individuals to adjust the conceptualization of problems based on the accessibility and usability of data. They used the analysis of data and interactions with experts to redefine and evolve problems as needed. The responses also indicate that obstacles existed that could limit optioning, and the ability of individuals to engage in optioning was influenced by relationships with others.

H1 and H2: Optioning and communicative relationships with organizational experts

H1 held that access to, trust in, and connection with organizational experts would be positively related to optioning, and H1 was supported. Optioning was positively related to expert access (r = 0.31, p = 0.02), connection (r = 0.41, p < 0.01) and trust (r = 0.31, p = 0.03). These relationships held in the final conditional process model, which controlled for data accessibility, task interdependence, communication frequency and time scarcity (see Table 2). H2 further posited that connection and trust would moderate the relationship between access and optioning. H2 received mixed support. We tested each two-way interaction discretely and in combination. For optioning, we found support for the interaction between expert access and connection (ΔR² = 0.13, F[1,41]=14.193, p < 0.01); however, the moderating effect was not as expected. We expected the quality of connection to compound the positive effects of access. In fact, higher levels of connection produced more optioning when access was lower, but higher levels of access and connection combined to reduce optioning.

OPEN IN VIEWER

Table 2.

Conditional process modeling of project autonomy and optioning.

	Project autonomy				Optioning
	b	SE	LLCI	ULCI	b	SE	LLCI	ULCI
Constant	−2.43	2.12	−6.72	1.84	−7.34*	2.03	−11.45	−3.24
Project autonomy	−	−	−	−	0.50*	0.15	0.21	0.80
Expert access (EC)	1.49*	0.44	0.91	2.37	1.68*	0.47	0.72	2.63
Expert connection (EC)	1.73*	0.41	0.91	2.56	1.52*	0.46	0.59	2.45
Expert trust	−0.27	0.15	−0.57	0.03	0.30*	0.15	0.01	0.59
Interdependence	−0.01	0.09	−0.18	0.17	0.16	0.08	−0.01	0.33
Data accessibility	−0.01	0.08	−0.18	0.16	0.01	0.08	−0.15	0.17
Time scarcity	−0.04	0.06	−0.16	0.09	0.01	0.06	−0.11	0.13
Comm. frequency	0.001	0.001	−0.003	0.002	−0.001	0.001	−0.003	0.002
EA × EC	−0.27*	0.09	−0.45	−0.09	−0.35*	0.09	−0.54	−0.16

SE = standard error; LLCI = lower level of 95% confidence interval; ULCI = upper level of 95% confidence interval; Comm. frequency = communication frequency. R²_autonomy = 0.50, F(8, 42) = 5.210, p < 0.01. R²_optioning = 0.63, F(9, 41) = 7.712, p < 0.01. Significant coefficients marked with an asterisk (p < 0.05). Index of moderated mediation, coeff = −0.14, SE = 0.09, 95% CI = [–0.38, –0.01].

RQ2: Marshalling relationships to address data problems

The interview data may help explain this unexpected finding regarding the relationship between expert access and connection. Taken together, the findings may indicate that these processes are not merely multiplicative, but that they involve different mechanisms altogether. Participants’ work placed them in the roles of producer and consumer of analysis. They analyzed data and presented results to more senior FSC executives, and they received information and insights from their own teams and other units within the organization. Participants described strategies associated with how they requested data and how they presented the data to others (e.g. internal partners and peer executives, sales leaders they coached, their leadership, their own teams). At times, participants were drawing on other organizational experts for insight and data to help them conduct and deliver an analysis. At other times, they were asking the other organizational experts to complete and deliver that analysis to them so they could pass it along.

At FSC, if access and connection with other organizational experts operated differently than we expected, they may have also involved different strategies for doing analytics work. This finding would be consistent with the high degree of variability in how frequently participants communicated with other experts (see Table 1). Though they reported needing to get access to data and analysis from others in the organization, in some cases, doing so involved little interaction, and in others it involved a great deal. According to the survey data, under circumstances where participants reported lower levels of access to experts, connection acted as expected. It was related to higher reported optioning. Connection might help participants and the other organizational experts to develop interpretations of data together despite their limited access to them (i.e. increasing optioning). Higher access, but lower connection might have involved repeated requests for data that participants then would analyze (i.e. increasing optioning). Higher connection and access together might have allowed them to more clearly direct the other organizational expert to provide a specific answer to their question (i.e. delegating the analysis, reducing optioning). Understanding the problems participants encountered in analytics work and how they marshalled relationships with other organizational experts to address them can help explain this finding by comparing how participants described analytics problem-solving strategies involving access and connection.

H3 and H4: Project autonomy and optioning

H3 posited that project autonomy would be positively related to optioning. H3 was supported (r = 0.62, p < 0.01). The relationship held controlling for covariates (e.g. communication frequency, time scarcity, interdependence and data accessibility) and independent variables (e.g. access, trust and connection). In the most complex conditional process model (see Table 2), a positive relationship between project autonomy and optioning remained (b = 0.50, SE = 0.15, t = 3.45, p < 0.01, 95% CI = [0.21, 0.80], R² = 0.63).

H4 further posited that the measures of expert relationships would have indirect relationships on optioning mediated by project autonomy. To investigate this hypothesis, we first looked at the degree to which the measures of expert relationships had direct relationships with project autonomy. Project autonomy was positively related to related to access (r = 0.40, p < 0.01) and connection (r = 0.51, p < 0.01), but not trust (r = 0.10, p = 0.48). These relationships held in the final conditional process model (Table 2). We found an interaction between expert access and connection on project autonomy (ΔR² = 0.11, F [1,42] = 9.298, p < 0.01), such that the quality of expert connection moderated the influence of expert access on project autonomy such that the quality of connection mitigated the effect of lower expert access controlling for other factors. To address H4 building on these analyses, we considered the mediation of each indicator and the degree to which the indirect effects of expert access on optioning through autonomy would be moderated by expert connection.

H4 was partially supported. Controlling for the other factors, expert trust did have a direct relationship with optioning, but not an indirect one mediated by project autonomy (coeff = -0.16, SE = 0.11, 95% CI = [-0.49, 0.01], R² = 0.52). The results did indicate the moderated mediation of an indirect effect of expert access and connection on optioning through project autonomy (index of moderated mediation, coeff = -0.14, SE = 0.09, 95% CI = [-0.38, -0.01], R² = 0.63, see Table 2). As access improved, the indirect relationship between connection and optioning through project autonomy diminished (indirect effects of connection on optioning through autonomy by access coeff_lowaccess = 0.34, SE = 0.16, 95% CI = [0.08, 0.79], coeff_mediumaccess = 0.22, SE = 0.13, 95% CI = [0.05, 0.63]), coeff_highaccess = 0.10, SE = 0.11, 95% CI = [-0.07, 0.46]).

This finding is consistent with the analysis above (i.e. H2 and RQ2) that drew a distinction between access for requesting data and connection for collaborating to generate interpretations of data together. Project autonomy might not be as relevant in circumstances marked by higher levels of access, but lower levels of connection. Simply requesting data and analyses would not require project autonomy. Under circumstances marked by higher levels of connection, project autonomy might be useful, because working with other organizational experts would involve their own hierarchies and agendas that might constrain optioning. Also, as above, participants’ accounts of how they managed constraints on their autonomy demonstrated important but differing roles of relationships with other organizational experts.

RQ3: Marshalling relationships for project autonomy

Participants argued that individual and organizational agendas, hierarchical divisions, and legal and regulatory constraints could make their analytics work difficult, and these problems were difficult in part because they put limits on their autonomy to go after questions and data as they saw fit. For example, agendas included preexisting frames for data or preexisting ideas that would influence the results of analysis. Blake commented that his ‘former leadership didn’t want to know’ what the data might indicate. He told a story of data analysis conflicting with an already accepted decision, ‘data changed the outcome . . . but the [region] makes a different decision in the end . . . we didn’t say they were wrong.’ Among the principal stakeholders, one ‘checked out,’ and another agreed to disagree with the outcome. Recognizing and avoiding obstacles posed by existing agendas was part of analytics work.

In other instances, individual agendas could undermine the usefulness of data for broader organizational use. Weber explained how sales associates understood the systems of evaluation, and might be tempted to ‘game the system’ by entering data in ways that favored them. Blake stated this practice more simply, ‘I’m used to people trying to use numbers to lie.’ Agendas could also result in inefficient uses of available data. Charley referred to a ‘culture of ignoring’ prevalent in parts of the organization and noted data were inaccessible at times because leaders refused to let others access them. Another challenge was that over time particular practices built up inertia, such that units develop vested ways of seeing data. Heather remarked that ‘each [region] has its own data,’ and Weber argued that, even in cases where existing data measured the wrong things, ‘changing the number, [how it is measured] is really hard.’ The entrenchment of hierarchies and respective agenda of units or leaders reinforced divisions in the organization, and in doing so created differences in how groups developed and propagated data.

In the interview data, time scarcity also interacted differently with constraints on workers’ autonomy. Blake argued that analytics could be hard and complicated, and so ‘most people will take a shorter route to town.’ Individuals faced pressure to make decisions quickly, yet there was often more data available for them to potentially consider. Charley commented, ‘At [the organization], we have discomfort deciding based on imperfect data yet it’s all imperfect data.’ He also explained that the pace of the work and the reluctance to act on data with any weaknesses, meant the organization would revert to accepted wisdom or convention to make decisions on time rather than making data-driven decisions. Even if individuals had access to data, or developed the relationships with experts to facilitate work with data, time constraints could stifle the ability to exercise desired autonomy when conducting analytics work.

The presence of existing agendas for data and the hierarchical division of authority over data meant that relationships needed to be established across organizational units and across organizational ranks. The strategies mentioned for addressing these difficulties tended to involve richer engagement with other experts that entailed more than making data requests. Navigating these issues meant needing to know not just who to ask but to know how to communicate with the person they were asking. Blake described needing to ‘guess what their issues are going to be using what I know about them.’ He argued that he needed to ‘make them look good . . . I’m not trying to blindside them . . . seeing their mistakes but not doing gotcha!’ Working through informal channels could help as well. Frank argued, for example, rather than sending ‘memos and PowerPoints,’ he wanted to spend ‘a half-hour at someone’s desk’ where the ‘richness’ of ‘being in the room allows for more influence.’ Formal meetings were a space for rationalizing decisions, but informal interactions were central for gathering the data needed to make decisions. Informal communication with individuals early in projects enabled autonomy to the extent that they brought stakeholders on board during preliminary conversations about the work with data and increased the likelihood they would be invested in the project.

Constraints on project autonomy were not only internal to the organization. Participants explained that legal and regulatory frameworks constrained how they could or had to work with data. Charley mentioned that one reason FSC refused to collect data is that to do so might impose a legal risk—real or perceived. He commented that, ‘because of the law, we can’t measure some of these things,’ and there is risk associated with having data, the ‘danger of knowing.’ FSC dealt with a great deal of litigation and had an entire division dedicated to establishing policy for document management, retention and destruction. Confronted by fears that data that might threaten the organization, he explained, ‘that’s usually where we stop,’ and that they were ‘conservative by default.’ Relationships were useful in helping workers identify and understand the actual requirements and address the unease that data might have legal implications.

In summary, participants described working with other organizational experts to make requests for data, delegate and collaborate about pressing business questions. Strategies useful in dealing with the difficulties of autonomy included (a) cultivating an understanding of and respecting the needs of others and (b) having informal conversations to influence how findings would be received and acted upon and to make sense of the constraints on their work with data. The findings indicate that project autonomy was positively related to optioning, and that project autonomy may play a different role when participants were requesting data or analyses than when they were working with organizational experts to make interpretations of data.

Relationships in analytics work

We have conceptualized analytics as a distinct form of multidisciplinary knowledge work that depends on expert relationships. Our findings indicated that in FSC’s efforts to implement analytics, practitioners needed to manage existing relationships or form new relationships with experts who possessed the data they needed, who could help them make sense of them, and who could, at times, collaborate with them to generate interpretations. Optioning and project autonomy were products of a suite of communicative and relational strategies that participants use to overcome challenges related to accessing and integrating data in an environment where existing organizational structures made this difficult.

Related to this first contribution, the findings at FSC demonstrate that analytics did generally involve the relational complications of multidisciplinary knowledge work (e.g. professional identity; impression management; problems of interpretation, translation and meaning making; and organizational politics and struggles for legitimacy). For example, Tyler’s preparations for regular requests were about making the work more efficient and also anticipating the conversations that he would need to have to understand the needs of collaborators and audiences for data (i.e. Barley, 2015). This example, and others like it, demonstrated that to get data, analyze them and communicate recommendations, participants needed strategies for navigating a web of relationships, power structures and existing ideas about what analyses ought to produce.

Building on these ideas, the findings also contribute to research in this domain by highlighting specific communication practices in analytics work that merit attention, and the complex relational dynamics implicated in these practices. For example, practices identified included filtering by audience, holding informal conversations, meeting ahead of meetings and preparing to respond to repeating requests. The findings demonstrated that these communicative practices at FSC depended on a rather nuanced understanding of audiences for data requests and analysis reporting (e.g. knowing who to ask, knowing what to ask for, knowing who is being asked) that went beyond simple maps of who knows what in organizations.

Requesting, collaboration and commissioning

The findings also pointed to relational differences involved in analytics work at FSC that we had not anticipated. We found that participants’ work with organizational experts may be thought of as involving requesting, collaborating and also an idea we are calling commissioning. In some cases, participants requested data, and in others they collaborated with experts to make sense of data. Participants may have also engaged in a relationally rich sort of delegation, commissioning, that would explain the finding that higher levels of access and connection were associated with reduced optioning. In commissioning, participants’ rich engagement with organizational experts (higher access and connection) may have allowed them to have robust conversations that defined the questions at hand, thus enabling the expert to simply go and find the answer. The participants in these cases delegated the analysis to the expert, after richer initial discussions about the work to be done. After receiving the findings they commissioned, participants could then decide what action to take based on the other organizational experts’ conclusions.

This insight is important, because commissioning may depend on the robustness of relationships in ways not consistent with requesting or collaboration. Commissioning would differ from simply making requests for data (i.e. transferring information from one part of the organization to another), because it requires greater depths of access, trust and connection. Commissioning would also differ from collaboration in that it may shortcut or shift processes of analysis and insight generation. Requesting, collaborating and commissioning may involve distinct mechanisms, and future research should consider the extent to which commissioning involves communication strategies and relationships in ways that differ from those involved in requesting or collaborating. For example, commissioning may depend on existing expert relationships developed through previous work to reduce the time needed to address questions at hand (e.g. Majchrzak et al., 2012). Although existing research points to the need for rich relationships in knowledge-intensive work like analytics, in practice, the same relationships may also support what may seem like straightforward exchanges but are not. Distinctions between requesting, collaborating and commissioning may also have implications for autonomy in analytics work.

Autonomy

At the outset, we argued that the implementation of analytics as a new organizational function would require practitioners to have a distinctive expertise and identity, and to be able to exercise their professional craft with discretion, and, as such, optioning would be related to project autonomy. Our findings at FSC confirmed that being able to develop options for interpreting information was related to: having the freedom to take risks, asking what might seem like odd questions about a project, addressing the project as they saw fit, and exercising their own judgment. Project autonomy had a direct effect on optioning, and it mediated the indirect effects of expert relationships on optioning as well. The findings help us understand autonomy in working with data as intertwined with the communicative and relational negotiation of work with other organizational experts.

Requesting, collaborating and commissioning may also differ in the degree to which they allow for and constrain autonomy. When making straightforward requests for data, project autonomy may be less relevant. Participants may not need to rely on organizational experts to provide input on how the data are analyzed and interpreted. In instances of collaboration, autonomy may be important so that analysts can work with other organizational experts without being constrained by collaborators’ agendas and priorities. In commissioning, the need for autonomy may actually shift to the other organizational experts doing the analysis work.

The findings also suggest that the practice of analytics itself requires the possession and application of distinctive expertise and knowledge. This expertise involves not only analytical acumen, but also navigation of complex relationships while maintaining autonomy. This combination of expertise and knowledge is not associated with any specialist function in the organization, but instead requires expertise with interactions across disciplines, as well as expertise in processing, interpreting and utilizing diverse sets of data (Treem and Barley, 2016).

Navigating the practical difficulties of analytics work

The findings also make contributions to the practice of analytics by providing examples and insights related to challenges faced during implementation. For example, implementing analytics will require leaders to improve their understanding of organizational data constraints, quality, ownership, accessibility and interpretability. The increasing complexity and volume of data available may complicate these problems by making the process of developing shared understandings of data more fragile (Tanweer et al., 2016). This study suggests that in implementing analytics, organizations have to cope with relational, communicative, as well as technical problems. The findings indicate organizations may be able to address these problems by helping practitioners (a) know who to ask for data but also know more about the person they are asking, (b) consider audiences in filtering and focusing work with data, (c) prepare in advance for recurring requests and problems, and (d) work through informal relationships to define problems and identify workarounds.

Failing to recognize and account for these dimensions of analytics work may short circuit the sort of exploration required by the practice of analytics (Marchland and Peppard, 2013). Although research suggests that processes like optioning can improve the outcomes of knowledge-intensive work (Cross and Sproull, 2004; Mom et al., 2015; Tsoukas, 2009), for participants at FSC, optioning was not the end unto itself. They did not simply want options; they wanted the answer. At FSC, when a simpler solution was available, it was hard to justify taking time to explore other options. Practitioners of analytics should also be ready to make the case for ways of looking at data that seem unusual or unorthodox. Informal relationships, meetings ahead of meetings and anticipating the needs of partners may be useful in doing so.

The findings also rebuke an outdated vision of analysts and technical professionals as set apart from and misunderstood by organizational leaders. Instead, they point to fundamental difficulties of differences in expertise (Kuhn and Jackson, 2008). Participants’ accounts of their strategies positioned them as data-savvy problem solvers but also political operators who communicated informally with others prior to planned meetings, who carefully selected the audiences in their work with data, and who made use of visuals and stories to help audiences make sense of analytics in the context of what they knew and believed (Barley, 2015).

Accounting for the complicated functioning of analytics and expertise in organizations suggests the need to suffuse not just data acumen but autonomy. To the extent that organizations treat the possession of information and data as key, business units that own data may, for example, hoard it to maintain their power and importance in the organization. By restricting access, business units could retain the right to analyze the data themselves. Such secrecy may actually create an illusion of expertise: Data hoarders may seem to be the ones who produce valuable, data-driven insights simply because others cannot access the data to discover for themselves. As such, the push for analytics in organizations may actually exacerbate such problems, because units see analytics as a way to set themselves apart. Research needs to consider how the development of analytics capabilities may disrupt existing power structures. Negative, unanticipated consequences like data hoarding may be avoided by fostering the relational and communicative capabilities documented here.