[Paper Review] Unlocking the power of AI in CRM: A comprehensive multidimensional exploration

• Meta
• TL;DR
• Introduction
• Key Contributions
• Related Work / Background
  • Dynamic capabilities (DC) and their microfoundations
  • The prior AI-CRM literature
  • Two threads that matter most
• Method / Architecture
  • The six-stage scoping review
  • The thematic-coding procedure
  • Sample limits and trustworthiness
  • Interview demographics
• Objective / Loss Function
• Training Data and Pipeline
  • Literature data
  • Interview data
• Results
  • Result 1 — Deriving the three-dimension, eight-sub-dimension framework
    • Data management capability
    • Multichannel integration capability
    • Service offerings capability
  • Result 2 — The audit trail in Table 5
  • Result 3 — Candidate measurement scales (Table 6)
• Analysis / Ablation
• Limitations and Critical Read
• Takeaways
• Setup and Use (if code is released)
• References and Resources
• Further Reading

Meta

TL;DR

• This paper assembles AI-powered CRM capabilities not from a quantitative model but from a qualitative synthesis. It combines a six-stage scoping review (Arksey & O’Malley, 2005; Fowler & Thomas, 2023) with thematic analysis (Braun & Clarke, 2006), drawing on 64 articles plus 24 interviews with practitioners who have been using AI-CRM systems for at least two years.
• The result is a structure of three dimensions and eight sub-dimensions. (1) Data management: data governance, data analytics, privacy & security. (2) Multichannel integration: content consistency, process consistency. (3) Service offerings: personalisation, automation, meaningfulness & novelty.
• The theoretical anchor is Teece’s dynamic capabilities (DC) and, more importantly, their microfoundations — the individual- and group-level mechanisms by which behaviours, decisions, and routines aggregate into firm-level sensing/seizing/transforming. The paper closes a real gap: microfoundations have been underexplored in the AI-CRM literature.
• There is no new model or algorithm. What you get instead is a conceptual framework plus a candidate measurement-scale pool (Table 6) for each sub-dimension, intended as a launching pad for follow-up quantitative work.
• Limitations are clear-eyed: a small interview sample (n=24) skewed to Australia (50%) and the UAE (25%), thematic coding by a single research team, mixed B2C/B2B industries, and no causal link to financial outcomes. Even so, this is one of the most disciplined attempts so far to break “AI-CRM capability” — a phrase that usually does too much work — into measurable units.

Introduction

AI-powered CRM is no longer a side project for the marketing team. The global CRM market is projected to grow from USD 18.1B in 2021 to USD 25.7B by 2032 at a 3.2% CAGR, and every major vendor — Salesforce Einstein, Microsoft Dynamics 365, SAP, Oracle, IBM Watson — now ships AI modules as core functionality. And yet, both academia and practice keep reporting the same kind of frustration: we deployed it, but it didn’t work. CRM project failure rates land between 18% and 69%, with some executive interviews citing as high as 90% (Edinger, 2018). The surface causes — data quality, user resistance, integration failure — are all real, but underneath them sits a deeper question: what exactly is the “capability” that an AI-powered CRM is supposed to create, and which micro-level behaviours and processes give rise to it?

This paper takes that question head-on. It does so not with the usual SEM-with-200-survey-respondents approach, but by acknowledging the conceptual fragmentation of the field. Prior studies have looked at AI-CRM through RBV, TOE, TAM, value-sensitive design, brand personality, and several other lenses. As a result, papers about data management, chatbot personification, and organisational absorptive capacity all sit under the same umbrella but speak in mutually unintelligible vocabularies. This paper uses thematic analysis to flatten those vocabularies into a shared schema.

For a reader closer to the ML side of the world, the value here is twofold. First, the paper bridges the gap between people who build models and organisations that deploy them, using microfoundations as a translation layer. Second, the candidate measurement scales (Table 6) are immediately useful: each sub-dimension comes with five to fifteen draft items that you can adapt directly into an internal survey for evaluating your own AI-CRM rollout.

Key Contributions

• An integrated taxonomy of AI-CRM capabilities — the three-dimension, eight-sub-dimension hierarchy in Fig. 2. It pulls apart what the prior literature had been treating as a single undifferentiated “AI-CRM” lump.
• Theoretical anchoring in Teece’s (2007) microfoundations of dynamic capabilities. The paper does not just ask what AI-CRM does; it asks which micro-level behaviours and routines produce that capability. This fills a recognised gap (Magistretti et al., 2021; Chatterjee et al., 2022b) where the AI-CRM literature has stayed at the firm level.
• Methodological triangulation — six-stage PRISMA-based scoping review + 24 in-depth interviews + inductive thematic coding by two academic coders for inter-rater reliability. Each data source independently challenges and validates the others.
• Three testable propositions (P1/P2/P3), one per dimension. Combined with Table 6’s candidate measurement items, they give follow-up quantitative researchers a clear specification to instrument.
• Practitioner-grade recommendations: from data governance items (user involvement, access-rights monitoring) to multichannel image consistency, to perceived-reliability items for automation. These are scale candidates today, but they translate cleanly into operational KPIs.

Related Work / Background

Dynamic capabilities (DC) and their microfoundations

Dynamic capabilities (Teece et al., 1997; Teece, 2007) are an organisation’s meta-ability to reconfigure its resources in changing environments, typically decomposed into sensing, seizing, and transforming. But DC itself is a firm-level abstraction — ask which person, with which tool, making which decision produces sensing, and the answer is empty.

That gap is exactly what microfoundations (Felin et al., 2012; Bojesson & Fundin, 2021; Hutton et al., 2021) tries to fill. It looks at individual cognition and motivation, team-level interactions, and procedural routines as the micro-level mechanisms that aggregate into firm-level capability. Reading AI-CRM through microfoundations, this paper argues that “AI-CRM capability” is not a vague organisational asset but the sum of how individuals access, interpret, and trust data + how the organisation standardises and reconfigures those routines.

The prior AI-CRM literature

Table 1 summarises the prior work the authors lean on (Baabdullah et al., 2021; Chatterjee et al., 2021b/2022b/2022c; Ling et al., 2021; Li & Xu, 2022; Monod et al., 2023; Suoniemi et al., 2021; Youn & Jin, 2021; Zhang et al., 2020). Two patterns stand out. First, most studies use a single theoretical lens — usually RBV or TOE/TAM. Second, the outcome variable is almost always adoption intention or firm performance. So the field has answered “what predicts adoption” reasonably well, but “what kind of capability does the deployed system actually create” remains open.

Two threads that matter most

The authors flag two specific lineages that this paper extends. (1) Akter et al. (2022) on dynamic data analytics capability — a self-citation by the second author, with a “data quality, data analysis, privacy” three-pillar structure that maps almost one-to-one onto this paper’s data management dimension. (2) Hossain et al. (2019) on multichannel integration quality — a systematic review built around content and process consistency that this paper imports more or less directly into the multichannel integration dimension. Separately, Huang & Rust’s (2018) mechanical / analytical / intuitive / empathetic intelligence taxonomy informs the automation-to-meaningfulness gradient inside service offerings.

Method / Architecture

There is no model architecture here. The research design itself is the largest section: a hybrid of a six-stage scoping review (Arksey & O’Malley, 2005, updated by Fowler & Thomas, 2023) and inductive thematic analysis.

The six-stage scoping review

1. Stage 1 — Define the research question: “What are the dimensions and sub-dimensions of AI-powered CRM capabilities, and how do they inform and extend existing theoretical frameworks?”
2. Stage 2 — Identify relevant studies: searching Business Source Complete, ProQuest, Scopus, and Web of Science with combinations of “AI” AND “CRM”, “artificial intelligence” AND “customer relationship management”, “automation and CRM”, “technological innovation and CRM”, and “CRM analytics”. English-language only, through the end of 2023. Initial result: 1,055 records.
3. Stage 3 — Study selection: 1,055 → 724 duplicates removed → 331 screened → 39 retained (after stripping out unrelated meanings of “CRM” such as coefficient of residual mass) → 25 added via cross-referencing → 64 included. The PRISMA flow (Fig. 1) makes the funnel explicit.
4. Stage 4 — Charting: each of the 64 papers is coded for citations, theoretical framing, methods, and key findings.
5. Stage 5 — Qualitative interviews (n=24): two channels. (a) Zoom 1:1 in-depth interviews of 45–60 minutes, semi-structured. (b) Phone or written PDF responses for a small minority. The interview guide is published in Appendix A. The participants — managers, marketers, data scientists, CTOs — had been using AI-CRM systems (Salesforce, Microsoft Dynamics 365, SAP, IBM Watson, Adobe, Oracle and others) for at least two years. Demographics: 50% Australia, 25% UAE, 71% male, average tenure 5–10 years.
6. Stage 6 — Collate and report: inductive thematic analysis (Braun & Clarke, 2006), with two academic coders working toward consensus to maintain inter-rater reliability.

The thematic-coding procedure

To reduce 64 papers and 24 interviews into the same coding scheme, the authors apply inductive thematic analysis. (1) Open coding line-by-line on interview transcripts and paper excerpts. (2) Cluster similar codes into sub-themes. (3) Cluster sub-themes into dimensions. The product is Table 4 (literature pattern) and Table 5 (interview quote → sub-dimension → dimension).

Sample limits and trustworthiness

The recruitment funnel: 239 LinkedIn outreaches → 14 purposive consents → 13 snowball additions → 27 responses → 3 dropped for incompleteness → 24 final. Saturation is declared once new transcripts stop adding new codes (Saunders et al., 2018). Trustworthiness rests on Denzin & Lincoln’s (2008) four pillars — triangulation, peer review, member-checking, audit trail.

The authors openly flag the 50% Australia + 25% UAE bias (McKnight, 2007). So the framework is, for now, a generalisation grounded in English-speaking and Middle Eastern markets and tech-forward enterprises.

Interview demographics

Objective / Loss Function

There is no learning objective or loss function — this is not a quantitative model. If we had to express the analytic objective of thematic analysis as a loss, it would look something like:

A good thematic schema is one that simultaneously covers enough quotes, doesn’t fragment into too many themes, and shows low disagreement between coders. The paper doesn’t put numbers on $\lambda$ or $\mu$, but the two-coder consensus procedure plus Cole’s (2024) inter-rater reliability checks are how those last two terms are kept in check.

Training Data and Pipeline

Literature data

Interview data

Results

“Results” here means the dimensions and sub-dimensions themselves and how they were validated, not model performance. The story unfolds in three stages.

Result 1 — Deriving the three-dimension, eight-sub-dimension framework

Fig. 2 is the diagram that anchors the paper. It deserves to be cited from anywhere in the manuscript.

Data management capability

• Intelligent data governance: who owns, who accesses, who is accountable for what data. Whether the organisation has a “single source of truth” (Interviewee #3).
• Data analytics: descriptive (what happened?) → predictive (what will happen?) → prescriptive (what should we do?), the three-stage scheme of Fantini & Narayandas (2023). Most interviewees flag predictive as the most transformative.
• Data privacy and security: GDPR compliance, breach response, encryption, anonymisation. From the transcripts: “Data privacy and data breaches are very common issue in the AI industry” (Interviewee #3, Male 30).

Multichannel integration capability

• Content consistency: whether the same message, promotion, and service performance are uniform across channels. The paper cites Simpson (2019, via Techipedia) on the ~20% drop in perceived brand quality when consistency breaks.
• Process consistency: regardless of channel, whether the underlying process (refunds, KYC, escalations) stays uniform. Drawn straight from Hossain et al. (2019/2020a).

Service offerings capability

• Personalisation: customer-segmentation-driven 1:1 responses. “The more personalised the content you can create, the higher the chances that your customers will be happy” (Interviewee #17, Male 34).
• Automation: removing humans from repetitive work. Most interviewees treat the perceived reliability of automation (“the system rarely makes mistakes”) as the headline KPI.
• Meaningfulness and novelty: beyond speed and accuracy — is the experience meaningful and novel to the customer? Pairs cleanly with Akter et al.’s (2023b) service-innovation scales.

Result 2 — The audit trail in Table 5

Table 5 is what makes the qualitative analysis defensible rather than ad hoc. Every interview quote is mapped to a specific code, sub-dimension, and dimension on the same row, so a reader can see exactly how the coding choices were made.

Result 3 — Candidate measurement scales (Table 6)

This is the most operationally useful part. Each sub-dimension comes with five to fifteen candidate items, each tied to a published reference (Akter et al. 2022, Tallon et al. 2013, Hossain et al. 2020b, Lee et al. 2019, Sun et al. 2020, etc.).

A representative sample:

• Intelligent data governance: “Ability to access very large, unstructured or fast-moving data”, “Ability to integrate data from multiple sources”, “Ability to keep customer data sufficiently private”, “User involvement in policy setting and evaluation”, “Plan and provide storage capacity”.
• Data privacy and security: “Personal information that is transmitted across various channels (websites, mobile apps and physical branches) is protected”, “Customers feel secure about using multiple channels”.
• Content consistency: “The system provides consistent information across all channels”, “Customers receive consistent responses through different channels”, “The service performance is consistent across different channels”.
• Process consistency: “The system images are consistent across all channels”, “The feelings about the service are consistent across all channels”.
• Personalisation: “The system can deliver immediate one-to-one responses based upon customers’ precise demands”, “Recommendations are offered based on personal information across different channels”.
• Automation: “Automated systems have 100% perfect performance”, “The level of automation is reliable / dependable / consistent / accessible”. Notably, the table also includes negative-worded items — “Automated systems rarely make mistakes”, “If an automated system makes a mistake, then it is completely useless” — which is a deliberate attempt to capture the two poles of perceived reliability.
• Meaningfulness & novelty: “Is relevant to the customers’ needs and expectations”, “Is really ‘out of the ordinary’”, “Demonstrates an unconventional way of solving problems”.

These scales are candidate items, not validated instruments. The paper does not attempt scale development itself; it offers a curated pool that subsequent quantitative work can refine.

Analysis / Ablation

There is no quantitative ablation. Instead, the paper uses interview quotes to make a qualitative case for each dimension’s necessity.

• If only data management is missing: “If the underlying data isn’t right and it isn’t filled in correctly, then the whole thing falls over” (Interviewee #19). However good your multichannel and service offerings layers are, if the data layer collapses, the whole system loses meaning. The authors imply a hierarchical dependency.
• If only multichannel integration is missing: content can be excellent, but message conflicts across channels translate into roughly a 20% drop in perceived brand quality (Simpson, 2019).
• If only service offerings are missing: data and channels may be tidy, but weak personalisation leaves customers unable to articulate why the company even has their data. The Loebbecke et al. (2020) substitution-of-empathy critique gets cited here too — automating personalisation can sometimes drain the genuine human empathy the customer was looking for in the first place.

The authors translate that trade-off into the paper’s propositional structure: P1 (data management), P2 (multichannel), P3 (service offerings) each substantially influences AI-CRM capability. Confirming those is a job for follow-up SEM, but the qualitative evidence in the transcripts is rich enough to support the claim that any one of the three going to zero collapses the whole thing.

Limitations and Critical Read

What the authors acknowledge:

• The review is bounded by the current AI-CRM literature, which limits external validity (Arksey & O’Malley, 2005).
• It relies on self-reported data, with the usual response-bias risk (Arnold & Feldman, 1981; Tourangeau & Yan, 2007). Future work should triangulate with observational data and objective KPIs.
• The study does not link AI-CRM capabilities to direct business outcomes — customer equity, operational efficiency, financial performance — and so cannot make causal claims about whether the framework produces value.

What the reviewer would add:

• Sample bias: 24 interviewees, half from Australia and a quarter from the UAE, 71% male. The markets where AI-CRM adoption is most active — the US, continental Europe, East Asia — are barely represented, which narrows generalisability.
• B2C/B2B mix is not separated: banking, hospitality, pharmaceuticals, and construction all sit in the same sample, but AI-CRM use cases differ a lot across industries (B2B leans on lead scoring, B2C on personalisation). The current framework absorbs that variance silently rather than modelling it.
• Microfoundations is invoked, not fully delivered: the authors use microfoundations as a theoretical umbrella, but Fig. 2 is a structure of capabilities, not a diagram of how individual behaviours and routines aggregate up. The aggregation story remains implicit.
• Reverse-keyed items in Table 6: the automation scale mixes positive and negative wording (“Automated systems have 100% perfect performance” with “Automated systems rarely make mistakes” and “If an automated system makes a mistake, then it is completely useless”) on a single ladder. That’s a defensible attempt at controlling for response sets, but follow-up factor analysis may struggle to extract a clean single factor.
• Limited comparison to neighbouring microfoundations work: Akter et al. (2022) on humanitarian analytics empowerment and Magistretti et al. (2021) on AI-related dynamic capabilities are adjacent and should have been benchmarked against. The question of why these eight sub-dimensions and not someone else’s seven or nine is left open.

Takeaways

• “Does it have AI?” is the wrong question to ask of a CRM. The right question is which of the eight sub-dimensions does it actually support, and at what level? Data governance + analytics + privacy / content consistency + process consistency / personalisation + automation + meaningfulness is the checklist.
• Perceived reliability of automation is a leading indicator of adoption. A single visible automation failure decays trust for a long time, which is why “rarely makes mistakes” deserves to be an operational KPI, not a soft claim.
• Multichannel integration has two distinct legs: content consistency and process consistency. Sending the same email across channels does not guarantee that the refund flow is uniform. They have to be measured separately.
• The microfoundations framing has a lesson for ML engineers too: model performance is not business capability. A 0.99 AUC model embedded in an organisation that lacks data governance, content consistency, and personalisation discipline will translate to roughly zero business impact.
• The paper is most useful as the front end of a follow-up programme: scale development → SEM → longitudinal validation. If you want to evaluate your own AI-CRM maturity right now, Table 6 is the seed of an internal survey you can run next week.

Setup and Use (if code is released)

There’s no code or dataset release. What you can copy is the research design itself, as a recipe:

1) Lift the candidate items from Table 6 verbatim.
2) For each of the eight sub-dimensions, pick 5–10 items and turn them into a 5-point Likert scale.
3) Sample frame: people who have used the AI-CRM system for 2+ years (sales reps, CRM admins, data stewards).
4) Pilot with n=20–30 → exploratory factor analysis → drop noisy items.
5) Main study with n=200–300 → confirmatory factor analysis → test the eight-factor structure.
6) Use SEM with business outcomes (NPS, customer equity, operational efficiency) as dependent variables.

This recipe is exactly the future-research direction the limitations section flags as the natural next step.

References and Resources

• Paper: Journal of Innovation & Knowledge 10 (2025) 100731
• DOI: 10.1016/j.jik.2025.100731
• Authors: kkska205@uowmail.edu.au · sakter@uow.edu.au · venkaty@uow.edu.au
• License: CC BY 4.0 (open access)

Further Reading

• Artificial intelligence in customer relationship management: A systematic framework for a successful integration (Ledro et al., 2025) — A sister paper from the same year in Journal of Business Research. Where this paper treats the capabilities that AI-CRM produces as the dependent variable, Ledro et al. focus on the integration process itself. Reviewed here as paper 0005 — together the two give complementary perspectives on the same domain.
• Explicating dynamic capabilities: the nature and microfoundations of (sustainable) enterprise performance (Teece, 2007) — The original source for dynamic capabilities and microfoundations. The sensing/seizing/transforming triad and the language this paper imports come from here.
• Artificial Intelligence in Service (Huang & Rust, 2018) — Splits AI intelligence into mechanical / analytical / intuitive / empathetic. Read as the theoretical gradient behind this paper’s automation-to-meaningfulness axis.
• Building dynamic service analytics capabilities for the digital marketplace (Akter et al., 2020) — A self-citation by the second author. Its data analytics capability model (selection / collection / interpretation / dissemination) is the direct ancestor of this paper’s data management dimension.
• Theorising the microfoundations of analytics empowerment capability for humanitarian service systems (Akter et al., 2022) — Same author group applying microfoundations to humanitarian analytics. The methodological backbone of that paper is essentially what gets reused here.
• Adoption of AI-integrated CRM system by Indian industry: from security and privacy perspective (Chatterjee et al., 2021) — Empirical evidence from India that security and privacy are dominant adoption drivers. Reinforces this paper’s data-privacy-and-security sub-dimension.
• Artificial intelligence (AI)-enabled CRM capability in healthcare: The impact on service innovation (Kumar et al., 2023) — Healthcare-specific AI-CRM splits capabilities into clinical / service / AI-engagement. A useful contrast that highlights how industry context changes the right slicing.
• Multichannel integration quality: A systematic review and agenda for future research (Hossain et al., 2019) — The original source of the content/process consistency split. This paper’s multichannel integration dimension borrows the schema almost wholesale.