Agentic AI
Accenture
Accenture built the largest AI practice among the traditional systems integrators, and its financial results back…
Area
Customer analytics is the practice of modelling who buys what and why, then measuring whether an intervention changed it. In enterprise use it covers segmentation, propensity, personalisation and marketing measurement.
Agentic AI
Accenture built the largest AI practice among the traditional systems integrators, and its financial results back…
AI Consulting
Bain and OpenAI have run a joint services alliance since 2022, and Bain became an OpenAI…
Agentic AI
Tredence’s generative AI practice centres on building and deploying agentic systems that automate multistep workflows for…
Agentic AI
Claude is a family of large language models made by Anthropic, available as a web and…
Customer analytics is the practice of modelling who buys what and why, then measuring whether an intervention changed it. In enterprise use it covers segmentation, propensity, personalisation and marketing measurement.
Area reviewed 14 Aug 2026
What the work involves
This work predates the current AI cycle by two decades, which makes it the most mature area in the directory and the one where vendor claims are easiest to check. Propensity models, segmentation and lifetime value are well-understood problems with established methods. The AI layer mostly changes how personalisation is delivered, not how it is measured.
The measurement problem
A personalisation model will almost always appear to work, because the customers it targets were more likely to buy anyway. Without a holdout group you are measuring selection, not impact. This is the single most common flaw we see, and it is why we weight measurement rigour above model sophistication when assessing firms here.
Firms in this area
See all 27 firms in this area
Research notes
We asked every firm one question first: describe the holdout design on your most recent personalisation engagement. Seven of 27 could describe a clean one without prompting. Eleven described something partial. Nine could not, or described a pre-post comparison with no control, which is not measurement.
| Holdout design described | Firms | Our read |
|---|---|---|
| Clean control group, unprompted | 7 | Assessable, shortlist candidates |
| Partial or inconsistent | 11 | Ask for the test design in writing |
| Pre-post with no control | 9 | Treat reported uplift as unverified |
Make holdout design a shortlist criterion rather than a technical detail settled later. A firm that cannot describe one on a first call will not build one for you, and every accuracy number in the engagement afterwards will be unfalsifiable.