Generative AI and the ‘Self-Healing’ Company

A chance conversation this week reminded me that it is 30 years ago this month that I first began my career as a management accountant. This prompted some reflection that for all the various technologies, tools and frameworks that I have used over the years the central professional curiosity is still the same:  to what extent can we design and implement a comprehensive performance management and measurement system which contributes to the reliable planning and execution of objectives in a sustainable way; that integrates people and systems; spans deterministic and non-deterministic processes; delivers more value than cost; controls operations and delivers conformance to quality standards without stifling innovation and entrepreneurial spirit; is flexible and resilient and can evolve to deal with changing circumstances and can focus management attention on strategic matters rather than operational firefighting. In short…can be self-healing.

The mental model I originally used for this was the autopilot of a large ‘fly by wire’ airliner which can adjust not just to the unchanging demands of physics but also the less predictable elements of weather.  Over time that mental model has evolved to the ‘self-healing’ – able to resist the business equivalents of both bird and baggage handler strikes and perhaps even the eruption of Icelandic volcanoes.  So the problem has remained fairly constant.  Let me take the occurrence of this 30^th anniversary and the emergence of Generative AI to consider whether the solution is any more likely

The Levers of Control

Let us place this whole discussion within the context of a Business Performance Management and Measurement System (PMS) or perhaps a Management Control System (MCS).  The literature would probably distinguish between the two, assigning a wider scope to the former, but for my current purpose I am going to make them broadly synonymous.  The literature is indeed vast.  For a bite-size start you could do worse than: Andy Neely (ed):  Business Performance Management which is a few years old but does lay the foundations.  Neely paraphrases Michael Lebas and Ken Uske’s chapter: business performance ‘should be equated with purposeful action taken today designed to produce meaningful results tomorrow.’  This is pithy and foundational enough for me to have memorised and I can quote it as freely as I can Edward Fox’s speech in ‘A Bridge Too Far’.

Another nugget from the book comes in the Marshall Meyer chapter ‘Finding Performance: new discipline in management.’  Meyer writes:

“An optimal PMS would have measures that:

• Would be few in number – perhaps 3 financial, 3-non financial
• Leading non-financial would predict lagging financial measures
• Pervaded all levels and all divisions
• Were stable and evolved slowly
• Would drive compensation.”

My next reference point is Robert Simon’s influential ‘Levers of Control’ theory.  Simon hypothesises four ‘levers’ that can be used to steer the organisation: Belief Systems, Boundary Systems, Diagnostic Control and Interactive Control.   On reflection, most of the initiatives I have been involved in have focused on only one ‘diagnostic’ hammer and have gone looking for amenable business process nails.  Invariably real world thumbs get in the way.  The result was poor fit, lack of adoption the need for continued ad-hoc, ‘interactive’ management firefighting.

So here we get to our first meaningful challenge.  Whereas I have always seen the relevance of this research to be perhaps the incremental creeping along the diagnostic/interactive continuum with ever-more complicated if/then/else statements, does Generative AI open up the potential for a more material leap:  an order of magnitude improvement or even paradigm shift?

Lets temper this optimism with some notes of caution, though.  First would be to point you at the field of System Dynamics which identifies complex systems as being the complex interaction of stocks and flows linked through feedback loops and time delays resulting in non-linear and emergent behaviour.  System Dynamics and agent-based modelling provides us with some very useful and powerful simulation tools, but they are not for the feint-hearted.  Secondly W.Ross Ashby’s work on cybernetics has throws up the ‘law of requisite variety’: that a control system must be at least as sophisticated as the reality it seeks to control, or perhaps more so, since adding measurement and control elements to a system makes it more complicated.  This places a computational and cognitive limit on our efforts.  So, we are aware of these crocodiles as we enter the swamp.

Knowledge-Intensive Business Processes

In my career I have focused on two flavours of process management design.  One has been the desire to achieve automation and commoditisation across industrial-scale banking activities such as ‘straight through processing’ or large-scale technology infrastructure optimisation.  The other has been more human and less deterministic approaches of consultancy projects.  In hindsight I must admit that the former has informed the latter, with the desire to corral highly variable, non-deterministic activities into repeatable patterns, identifying ‘happy paths’ and ‘exception handling’ and wanting to limit the damage that pesky humans could do.

Latterly though, I have come to embrace the squishy bits but have done so within the conceptual bucket of the ‘Knowledge Intensive Business Process.’  Wikipedia tells me the term was coined in 1990s but of course Peter Drucker referred to the knowledge worker in 1959 and if you read Adam Smith’s ‘Wealth of Nations’ after a couple of glasses of wine then you could easily find some resonant concepts.  While you are in the stacks you could also look for Dorothy Leonard’s ‘Wellsprings of Knowledge’ that talks about mining knowledge of your experts and Ikujiro Nonaka and the ‘Knowledge Creating Company’ which explores turning your ‘tacit’ knowledge into the ‘explicit’ kind.

It seems self-evident to me that this area will be the absolute nexus of the growing debate on how AI will impact knowledge work at the individual, task, process and corporate level. Lets dig into the process level first.

Di Ciccio, Marella & Russo [2015] identify the characteristics of Knowledge Intensive Processes:

1. Knowledge-driven (as opposed to merely data-driven)
2. Collaboration oriented
3. Unpredictable
4. Emergent
5. Goal-oriented
6. Constraint- and rule-driven
7. Non-repeatable.

All of these characteristics give us useful distinctions but I was particularly taken with their use of ‘emergent’ and ‘goal-oriented’ and the idea that objectives might be nested within one another.  An overall goal to solve a problem could have a standardized, repeatable, high-level approach even if the particular problem instance is itself unique and as-yet unsolved.  This simple insight delivers a double-whammy of benefit:  client processes can be represented as multi-faceted and multi-layered and modelled and decomposed accordingly but the analyst’s approach to solve for them might in itself be multi-layered.  At a stroke this resolves what for me has always been a tension at the heart of ‘Scrum’ which on the one hand wants to champion the creativity of developers but on the other wants to them to commoditise and commit to their sprint delivery.

The Anatomy of a Task

Once again coming at this from what I will call my legacy mindset I would typically have tried to decompose any process down to a series of ‘atomic’ tasks…a task that can reasonably be executed in one processing chunk by a single resource be they human:

• write a letter
• address the envelope
• affix a stamp
• seal the envelope
• post the letter

..or a machine:

• source the data
• clean the data
• summarise the data
• report the data

With less tractable tasks you might decompose to ‘next step’ + a blob of ‘the rest’.  Similarly group tasks are viewed with suspicion that joint responsibility equals none-delivery so you would typically try to get someone on the hook..and don’t get me started on ‘definition of done’.

David Brown’s paper Causal Mechanisms of Technological Fit in a Knowledge Work Environment is among other things an interesting literature review of ‘task/technology fit’ theory that provides a very rich set of distinctions for our analysis.  Brevity forbids a detailed exposition but let me try to summarise for my purposes:

• To define a task he starts with a simple ‘inputs to outputs’ and single resource simplification, but suggests that this clarity changes when you add more people.  I would posit that Generative AI profoundly disrupts that single resource paradigm as you could leverage it to play any number of expert roles.
• That opens up the Task Type circumplex of Joseph McGrath in identifying 8 broad task types analysed on a Behavioural/Conceptual and Co-operation/Conflict continuums.  The tasks are: generating ideas, planning, executing, resolving conflicts of power, conflicts of opinion and conflicts of interest, solving problems where there is a likely right answer and where there is likely no right answer.  This gives us a very rich taxonomy to decompose tasks and identify a beneficial role for the new tools.
• Finally he talks about ‘task complexity’ as a major determinant of task/technology fit.  One of the drivers of task complexity is ‘analysability’ and the ability to fit a task into a standard solution pattern or else consign it to the need for unique and creative response or simply ‘too difficult’.  While this is in the hands of an individual constrained by skills, knowledge and experience then of course it is a significant issue.  But surely the situation changes materially when you could input a set of data or problem description into the AI tool and say:  ‘please suggest the most appropriate problem solving approach for this type of raw data.’  Even more exciting would be ‘please parse this natural language description of a computational problem into a wolfram alpha or python script, solve and visualise appropriately.’

Round-trip analysis.  Process…

It is a commonly held problem in technology that requirements analysis and specification too often become detached from solution design and technical implementation.  The Agile manifesto that ‘values working software over documentation’ is a recognition of this.  There have been attempts to integrate design artefacts and physical implementation in the shape of enterprise architecture repositories but round-tripping from whiteboard hypothesis to implemented database schema and back again is at best a fragile ambition which does not survive first contact with the enemy of time.

That is not to say that there is not great value in something a little less ambitious: round-tripping from analysis artefacts to working prototype software which allows you validate your hypotheses, simulate your process and experiment with UX designs.  Let me give you some examples:

Mapping business processes is a pretty routinely-conducted business planning and analysis activity and a swimlane diagram is a good way to capture current process and explore target process in software implementation.  It also supports less technical needs like on-the-job training and general operational presentations.  Informal notations are fine, but BPMN is a rich and intuitive model which allows you to visualise business process and better capture nuanced real-world behaviours.  It is great for business storytelling and can de-mystify complex operations for managers who don’t do much walking about.   However, BPMN can also an executable specification.  In state-of-the-art tools like Trisotech the model and the executable workflow are one and the same thing.  You can go seamlessly from ‘whiteboard’ doodling to process deployment.  You can also abstract up to a less deterministic and more knowledge-centric CMMN ‘case management’ approach and reify down to a more granular DMN to better represent algorithmic, data and rule driven decisions.  Trisotech is an expensive and specialist tool but I am very surprised that more widely available Robotic Process Automation (RPA) tools have not embraced such a visual approach.  To my mind they haven’t yet distinguished between the case, process and task and algorithm level and are visual mish-mash of which repels all but the most tech savvy of business SMEs.

..and Data

So that is process modelling…what about data?  I have long thought that a similar seamless transition or even round-tripping should be available to take the audience from conceptual to logical to physical data model.  We have waited in vain for Visio or Miro or something to approach this from left to right or perhaps Microsoft Power Platform Dataverse to approach it from right to left.  We have waited for two years for Microsoft Loop to deliver structured data artefacts that can on the one hand be manipulated casually by groups of users but on the other hand represent structured and version controlled artefacts which can be the basis for building prototype applications.  The jury is still out.

Perhaps the answer lies in unstructured and linked data and the semantic web.  Why not use AI to generate your own corporate knowledge graph capturing the current state of your internal and external thinking on a particular domain and use it to automatically interrogate inbound news items to see how new or important it is?  You could literally come into the office and find the AI with its feet on the desk reading the Financial Times on your behalf.  Never again would you have FOMO over TLDR.

Other potential areas for ‘Self-Healing’

DevOps, XOps and Hypothesis-Driven Design

I have been influenced by Jez Humble’s very accessible introduction to DevOps: Accelerate. The major relevant points for this discussion are as follows:

• Integration of processes and measurement systems
• The need to enable all manner of infrastructure automation to facilitate not only production deployment but all research and experimentation as well.
• Hypothesis-driven development.

Self-Healing Infrastructure

While we are at the business end of deployment…its not an area deep of expertise for us but Infrastructure as a Service (IaaS) is clearly suggestive of self-healing infrastructure and the use of AI and ML.

Self-Healing Testing

As anyone who has ever tried it will know, web automation for simulation and testing is notoriously flakey.  We are now seeing the mention of ‘self-healing testing’ and the use of multiple protocols to try and re-try the job.  Presumably the back-stop would be a natural language description of overall objective translated into some technical remedy.  This could be a very big deal to put automated UAT and regression testing into the hands of non-technical audiences.

Causal Inference

Everyone is aware of the statistician’s mantra that ‘correlation is not causality’.  In order to establish causality you would need to run an experiment with all of the incumbent cost and lack of expertise.  Well, perhaps not any more.  Causal discovery (reverse engineering the possible causal relationships between variables) and Causal Inference (evaluating the strength of those relationships) is but a Chat GPT-authored python script away.  Business process cause and effect can be brought out from the c-suite gut-feel shadows into the objectively testable light. 

The Swiss Army Knife of Technology?

I hope these reflections have given you some ideas as to how Generative AI and Large Language Models combined with one or two useful methodologies and frameworks, might be used in small-scale, knowledge intensive operations.  From helping me brainstorm blue-sky ideas on user adoption approaches and proof-read my cv through to write python scripts to carry out causal analysis or enable risk-driven escalations which were previously the domain of the bounded rationality of flustered managers, Generative AI is becoming for me a daily or even hourly interaction.  As I write my inbox pings with the news that Add-Ins are coming to ChatGPT….what about multi-modal interaction that takes voice instructions and draws geometrically accurate pictures.

In short, I can see this richer and more granular decomposition of what makes up a knowledge intensive task combined with the domain specific fine-tuning and prompt engineering delivering a multi-faceted set of super-powers to knowledge workers but at the same time offering previously unimaginable opportunities for process automation.  It has become a common refrain that ‘AI won’t take your job but someone who knows about AI will.’  Who knows, it could be both….