Irrespective of the prediction, trend or new amazing technology, its impact is only realised when successfully deployed by an organisation’s expert team(s). But with so many potential opportunities for change (see the rest of our EMEA predictions series), on top of increasingly complex estates and unforgiving demand, technical leaders are pivotal in setting a clear direction for the organisations they serve.
To help, I’ve captured just a few of my thoughts for how leaders can differentiate some broad demands to make the best impact in the year to come. This blog is by no means a leadership manual (gods’ forbid) or exhaustive in any way, but an attempt to help stimulate thoughts and hopefully to prioritise some of your own unique approaches. Here are the four topics I’ll cover:
- Combine functional expertise
- Prioritise robust data foundations
- Understand & control analytics models
- Focus on the right technical delivery ‘core’
Let’s get to it.
Combine functional expertise
Significant incremental impact from your data and analytics (D&A) comes from its expansive use in specific context: from various business processes to operational outputs to customer interactions and beyond. It’s incredibly difficult to know where this impact will blossom and so ‘democratising data’ and ‘empowering everyone’ have become go-tos in the vision and strategy decks, but often forgotten and under-resourced are the expertise and capabilities needed to immediately support, differentiate and effectively scale.
Operational teams that deal with D&A day-in-day-out (i.e. security, IT, service, engineering, insight & reporting etc.) are ready-built pools of this resource with the required functional expertise. Current organisational structures, strategies and tooling often keep these teams separated and buried under business-as-usual pressures.
Be prescriptive about localised vs centralised capacity
So what to actually do? Leaders need to set an empowered and overarching strategy to combine the expertise in these teams with principles and mandates that make sense to your organisation, including:
- consistent leadership and definitional approach to incidents
- not building separate capacity in transformation or ‘special project’ teams
- allowing extra capacity to experiment, continuously retrain, and absorb the unknown
- embracing ideas and insight from everybody but trialling and implementation from a few
Combining expertise does not just need to mean throwing everyone into one blob; it means being prescriptive about localised vs centralised capacity. Replication of D&A capabilities locally should only be allowed when there is a clear differentiated need and centralised capabilities should be broadly accessible and transparent. This must be supported by structures and an ethos that promote the active sharing of expertise and experience amongst critical functions. In the midst of many DevWhateverOps frameworks this prescriptive consistency is not easy, but can provide valuable capacity (with direct operational context and insight) to generate business changing impact (that scales).
Prioritise robust data foundations
The trends over the last few years have thankfully moved, from just trying to put data in one location, to a focus on accessibility and visibility. These data foundations need to be a priority for all leaders this year, but with further constrained economics and budgets, a simple low-level blanket coverage won’t do much beyond the data swamps we currently have to deal with.
What’s needed is robust foundations that build in timeliness, intent and context based on the purpose of the D&A capability, with the flexibility to adapt for various actual outcomes. The data pillar(s) in organisational strategy gives technical leaders a real opportunity to build robust data foundations.
Not all data needs to be real-time. Defining a tiering approach on a spectrum from real-time to periodic can help
So what to actually do? You should challenge your teams who are building, designing or contributing to your D&A foundations with the following principles:
- Right-time: not all data needs to be real-time. Defining a tiering approach on a spectrum from real-time to periodic can help. By implementing a mechanism to challenge the need to move ‘up’ the scale towards real-time, you’ll be able to both optimise for cost and ensure the quality of real-time data coverage by focusing resources.
- Intent: data should be broadly separated by the initial intent for its use. I prefer to use categories such as “optimise, understand and challenge” but choose your own to make them relevant to outcomes of your business strategy.
- Sensitivity: be explicit and judicious in classification of data. Most data does not need to be restricted beyond default organisational controls. When it does, it's helpful to define explicit ways to make it more broadly shareable (masking, abstraction, grouping, etc.). The utility of D&A capabilities increases significantly with a data sharing ethos, which is undermined without these kinds of controls.
- Flexibility: ideally build for reuse and with the understanding that the actual outcome may require change, optimisation and enrichment.
Understand & control analytics models
Analytics and especially AI (we can leave the debate of defining that for now) will be part of everybody's business strategy again this year - but with more emphasis on how and where it can start delivering broader impact.
As leaders this year, we need to take more of a role in understanding and controlling for the broad effect of advanced analytics and ensuring that we have the monitoring and processes in place to drive the outcomes we expect and avoid potentially damaging (or disastrous) missteps. Generative AI and extremely large models are currently all over the news (see Predictions: AI and Automation for more details), but reports of their implementation remind me that we still have no prevalent systematic way to address problems caused by their intrinsic non-deterministic nature. Addressing these models’ capability overhang (unplanned and unexpected capabilities) or unintended drift with current post-processing systems or guardrails is ineffective - just look how unsuccessful prompt engineering has been working with LLM-powered web search.
Focusing on upstream data can still have the biggest impact for most organisations. Being more data-centric than model-centric means controlling and optimising for the data more than the model. This approach often encourages more control, drives better results and helps to directly build stronger data foundations.
Think glassbox not blackbox
So what to actually do? It’s not sexy, but the ability to audit and monitor for data quality is critical, so ensure you have it in-place or invest in augmenting existing capabilities. Something as simple as automating checks for basic distributional statistics can be a step-change in avoiding problems down-stream. And then, when it comes to deploying analytics in 2023, these are some principles I’d be using:
- Use statistical analytics first. It will provide a good baseline for improvements (if needed) and can generally scale well.
- Use ‘proven’ deterministic models next. This will allow for more focus on data quality, testing, optimisation and controls operational deployment.
- Document, and ideally standardise, training data sets.
- Define performance, operational and ethical benchmarks for all systems (ideally with strong guardrails, oversight responsibilities and defined thresholds).
- Experiment with generative or large separately trained models (e.g. LLM, GAN, transformers etc.) but don’t use anywhere near operational systems.
- Monitor, monitor, and monitor - upstream collection and context, pre-process, key inference points, postprocess, outputs and ideally down-stream effects. The more unfamiliar, large or non-deterministic the model, the more monitoring you need. Think glassbox not blackbox.
Focus on the right technical delivery ‘core’
Prioritisation is not new, but it will be a pervasive mantra for successful leadership teams through 2023
Your ‘core’ (apple analogy optional) is the area of your organisation that will remain static to enable the rest of the transformation and improvement to be successful. This can be for a variety of reasons from complexity, cost, regulation or external dependencies.
Prioritisation is not new, but it will be a pervasive mantra for successful leadership teams through 2023. Large transformation programmes that cover a broad spectrum of the business needs have been all the vogue for years, but will be hard to sustain throughout 2023.
We need to stop just talking about prioritisation and really drive it top-down. This means not producing strategies with lists of 10 or more ‘equally important’ big initiatives or overarching programmes, but defining the most important changes needed and creating the space for success.
Choosing a ‘core’ can be a really useful method for looking at things top-down, especially when providing technical feedback to boards and senior management initiatives on strategy for the year ahead.
Separate organisational and technical areas into stabilise and optimise (the ‘core’) vs improve and transform
So what to actually do? Spend time deciding on your core for the next planning cycle. Generally, I prefer to separate organisational and technical areas into stabilise and optimise (the ‘core’) vs improve and transform. A ‘core’ should be selected based on the organisation’s ongoing requirements to remain viable and the top priority for change. This is a balancing act (and trade-off) but needs to start explicitly top-down to allow each area, unit and team to do their own balancing and trade-off.
Major transformations are often described as “changing all the components in an aircraft whilst it’s still flying”, but just like not having enough of a ‘core’, it results in just one thing: you crash and burn.
That’s more than enough for now (I always like to end on a high note). If you’d like to know more, share some of your experiences or just point out where I’m wrong, please get in touch. To read more on other topics, please check out our EMEA Predictions 2023 series here.
Until next time…