In Case You Missed It
The Org Chart You Didn’t Design. Conway’s Law has gone into reverse, as AI agents reshape organisations
Executive Summary
Workflows and methods ‘mutate’ as front-line workers find better ways to complete their work, creating numerous variants.
If harnessed, this variation is valuable as it can deliver performance improvement but also increase resilience to shocks.
Selection identifies which variants to scale, retain or retire. It must govern what enters and remains in active use, so that experimentation does not create unmanageable complexity.
AI tools are turbo-charging mutation whilst making selection more challenging, creating opportunities but also risks.
AI therefore presents a paradox. It raises volatility in the environment, gives workers the tools to respond to that volatility, but hands managers new ways to constrain how workers respond.
Management systems (e.g., budgeting) built around standardisation and central approval are poorly suited to overseeing a mutating portfolio of workflows. Managers sometimes respond by trying to eradicate variants.
To unleash the power of front-line-led innovation whilst containing risk, managers need to shift from being ‘selectors of variants’ to ‘builders of selection machines’.
In 1943, Salvador Luria, a research scientist, was trying to unwind at a faculty dance in Bloomington, Indiana. His research into the behaviour of bacteria attacked by a virus was at an impasse. Most of the bacterial cultures grew into a handful of resistant colonies, but a few resulted in hundreds. During the evening, Luria watched a colleague win three dollars on a slot machine and immediately understood what his experiments were telling him. If exposure to the virus had induced resistance, the numbers of resistant colonies should have varied much less between cultures. The large differences instead suggested that resistant strains of bacteria were already present. Luria later shared the 1969 Nobel Prize in Physiology or Medicine with Max Delbrück and Alfred Hershey.
A business can become more inventive while becoming less capable of adapting. AI makes it easier to build new workflows, but those methods can enter daily use, spread and acquire dependencies before their performance and risk have been established. The challenge for leaders is not just evaluating an increasing number of ideas. It is governing an ever-expanding population of methods which are already up and running, and deciding which should scale, remain available or disappear.
Mutation Helps Business
Companies mutate and produce variants too. Organisations can be thought of as a collection of workflows (e.g., order to pay, hiring, software development). Workflow mutations arise over time, as employees avoid bottlenecks (e.g., ineffective IT systems) or find faster ways to complete a task. At Heathrow Airport, security officer Samit Saini used a coding tool to build an app that explains security rules to passengers in different languages. It was subsequently adopted across the airport’s security operations. It’s not uncommon to find dozens or even hundreds of variants of one workflow. When Accenture examined its purchase requisition process across 50 countries, it identified 14,000 different routes through the system, even though a single global instance of SAP was in place. Standardising the system doesn’t always standardise the work.
Businesses that can harness the power of these variants will benefit in two ways. Firstly, variants can deliver performance improvement. Built and tested by those who will live with them day in, day out, they are more likely to reflect the realities of the front line, automate low-value tasks and leave human intervention where it matters. Xerox built a system called Eureka that allowed local technicians to submit repair tips. Once validated by experienced peers, they were made available to all colleagues. In one case, a technician in Montreal uncovered a quick fix that avoided the need to replace a $40,000 copier in Brazil.
Secondly, retained variants can deliver increased resilience to shocks such as cyber attacks or supply chain disruptions. In March 2019, ransomware disrupted IT systems at Norsk Hydro. The company was able to run some operations whilst the computers were recovered, using old paper-based processes and retired employees who volunteered to help.
But not all variants are better than their predecessors. Firms must decide which to scale, retain or retire. Variants should be scaled if they improve end-to-end performance at an acceptable cost and risk. They should be retained if they preserve a usable alternative method (for example, they can function without using an AI model). And they should be retired if their value (either performance or resilience) no longer justifies the costs and risks of keeping them running. Novelty alone is not grounds to retain a variant.
The unit of selection here is a method, not a person. A company evolves only when better ways of working are identified, retained and passed on.
Balancing Mutation and Selection
In business, variation becomes a liability when workflows enter active use faster than the organisation can evaluate, coordinate and maintain them. The numbers won’t add up as financial models use different assumptions. Departments can’t talk to each other as data models diverge. Technical debt spirals as new code is added but not maintained. In 1997, Nestlé USA discovered its brands were paying 29 different prices for vanilla from the same supplier. Each factory negotiated its own deal and used its own ingredient coding, making purchases harder to compare across sites. The same review uncovered nine general ledger systems and 28 different points of customer entry.
Businesses try to constrain mutation through standard operating procedures and ERP systems that hard-code processes. As Accenture found, this is harder than it looks. Selection among existing variants occurs through a different mechanism. As employees migrate towards variants that help them meet their targets, some spread virally whilst others fall out of use. These mechanisms can limit operational complexity, but they do not guarantee that selection keeps pace with mutation.
Drawing analogies between biological populations and work is not a new concept. Since Nelson and Winter’s work in 1982, economists have described companies as generating, selecting and inheriting routines. But AI changes the relative economics of those activities. Creating a new method has become much cheaper and faster. But the cost or time needed to establish whether that method has value is not falling at the same rate.
AI is Accelerating Mutation
AI tools are massively accelerating workflow mutation. Front-line employees can spin up their own processes, data and technology using off-the-shelf LLMs. What used to take months now takes minutes. In 2025, Spanish banking giant BBVA reported that its employees had built more than 20,000 custom AI tools to support day-to-day operations.
And AI doesn’t just increase the number of variants. It also increases the granularity of what is being varied. Workflows tend to sub-divide when the benefits of task specialisation outweigh the incremental coordination costs. With AI, more granular tasks become economic. For example, AI has made it easier to divide the end-to-end hiring workflow into more granular units such as advert placement, application ranking and interviewing. Each of these units can mutate, and the variants then recombine into hiring workflows.
These two effects multiply when workflow fragments can be freely recombined. Illustratively, with 10 variants at each stage of a three-step hiring process, there are 1,000 possible end-to-end workflows.
The Fitness Challenge
This explosion in variant combinations also makes it challenging to identify the high-performing variants. The fitness of any one variant depends on those variants that surround it. So in our hiring example, an advert placement variant may work best with one candidate ranking variant and poorly with another. As a result, the answer to ‘which is the fittest variant?’ may be ‘it depends’.
The fitness of a variant can be hard to measure. A workflow using an LLM may produce different results from the same inputs. Model upgrades, changing data sets and shifting user behaviour can alter its results over time. Evaluation must distinguish a repeatable improvement from a one-off apparent success or change in conditions. Some evidence simply takes time to arrive. A new hiring workflow can be built in an afternoon, whilst evidence about the quality of its hires may take months to collect.
Fitness also depends on whose outcome is being measured. Suppose a sales team automates quotations, halving the time spent on each one. If this new workflow creates problems for the fulfilment team, the net effect may be only to transfer effort from one team to another. A variant should not count as fit merely because its creator benefits. Selection must account for downstream effects, as local optimisation can make the organisation less effective overall.
Most fundamentally, fitness can be hard to define. An outbound sales email’s open rate yields a clear KPI, whilst the quality of a hiring decision is harder to quantify.
Selection Becomes the Bottleneck
Selection becomes a bottleneck when the flow of variants needing evaluation exceeds the firm’s capacity to assess them. This can happen even if AI makes evaluation faster. AI can compress the time needed to create an experiment without necessarily compressing the period needed to learn from that experiment.
A related pattern is emerging in software development. Studying more than 500,000 GitHub developers across three generations of AI coding tools, Demirer, Musolff and Yang found that the gains shrank as the work progressed towards shipping. Commits (recorded code changes) rose by around 240 per cent cumulatively, projects by roughly 80 per cent, but releases only by 30 per cent. AI boosts creation but that doesn’t all flow into delivered output, as the constraint moves downstream.
Risks Spiral as Mutation Outruns Selection
Any individual new variant may create risk if critical control steps are omitted, cyber vulnerabilities are exposed or looping AI agents ramp up cost. Collectively the risks are more fundamental. The organisation can ‘splinter’ into fragmented units (potentially one per employee), each with its own workflow variants. Processes become opaque to managers, end-to-end processes break at the joins and single points of failure may be created (e.g., dependence on a specific LLM). Agents, collaborating with other agents, can paper over the cracks, obscuring the issues without solving them.
Reaping the benefits of workflow mutation whilst avoiding the risks therefore requires a management system that scales methods that improve performance, retains alternatives that preserve resilience and retires variants that no longer justify their costs.
But this reveals a deeper issue.
Today’s Management Systems Are Weak Selectors
Management systems built around central approval, periodic review and standardised execution are poorly suited to a rapidly mutating portfolio of workflows. Xerox’s Eureka process showed that it is possible to support variation, selection and scaling. The problem is not with management itself, but with management systems which are designed primarily to minimise variation. These systems face four mismatches:
Firstly, mismatched objective. These systems try to minimise variation. Managing for evolvability requires variants to not only be tolerated but encouraged.
Secondly, mismatched velocity. These systems operate on quarterly cycles and annual budgets. AI allows workflows to mutate by the hour, minute or even continuously if agents are asked to self-optimise.
Thirdly, mismatched direction. These systems assume that change originates centrally and flows down to the front line (albeit with input from workers). Workflow variations are built bottom up by workers for workers.
Finally, mismatched scale. These systems treat processes as monolithic and enterprise-wide. Workflow variations are fundamentally local, although they have the potential to be deployed widely.
Selection still happens even if managers fail to govern the process. Methods tend to spread by word of mouth, often because they are convenient, visible or sponsored by an influential colleague. Those pressures are likely to favour what is easy to adopt over what works well for the business. Managers therefore cannot simply allow a selection system to emerge; they must proactively design one.
But it gets worse.
Managers Can Suppress Mutation
Overwhelmed with higher-velocity mutation and a resulting explosion in variants, managers can respond by strengthening controls. Decisions are centralised, tech permissions revoked, SOPs reinstated.
AI provides managers with new methods to exert control too. Pest-control firm Aptive deployed AI to standardise the customer cancellation process. The system prompted staff to ask particular questions in a prescribed sequence. Adherence to the prescribed discovery behaviours (establishing why customers wanted to cancel) increased from 28% to 59% in four months.
So, the great paradox of AI is that it a) increases the volatility of the business environment and the need for firms to improve resilience to shocks; b) gives workers the tools to respond quickly to changes in the environment; and c) gives managers the tools to constrain how workers respond.
The reflex to exert controls is understandable but the evidence suggests it can be expensive. A study of US and OECD firms found that companies which had devolved authority to local managers before the financial crisis outperformed centralised firms in the worst-hit sectors. Turbulence raises the value of local information and the ability to act on it quickly. Centralisation can therefore be a costly response to a shock, just when local information and rapid adaptation become more valuable.
There is a difference between standardising a workflow and standardising the conditions under which workflows are evaluated. Variants can be compared with each other if their inputs are transparent, their outcomes are verifiable and ‘connectors’ to neighbouring processes are standardised. Such standards make variation transparent and testable. Management can specify how success is defined and proven, whilst the front line remains free to discover better ways of doing the work.
Selecting Selection Machines
How do we solve this paradox? The answer is to change the manager’s primary role from selecting individual variants to designing how selection happens. Managers define what counts as fit, what evidence is needed, and which decisions require human judgement. The selection machinery then converts those conditions into operational change by scaling, retaining or retiring variants.
Not all selection machines need to be AI-driven. Selection machines must gather evidence, take decisions and then take action (e.g., by replicating a ‘fit’ variant). Some tests can be automated but others will require human review, especially where decisions are irreversible. An AI’s review of a method it helped create can’t be treated as independent validation. Nor should a second model automatically count as independent. Different models can make correlated errors. Where errors could cause meaningful damage or risk, validation must be independent of the process that produced the variant.
To be effective, a selection machine must gather reliable evidence at low cost and remove avoidable delays. Where outcomes take time to emerge, it must oversee limited trials and defer wider deployment until sufficient evidence is collected. Popularity should not be mistaken for fitness. Workers’ willingness to use a variant and evidence that it is ‘fit’ are not the same thing, and selection machinery must distinguish between them, rather than automatically promoting whatever spreads fastest. And it must translate evidence of fitness into differential replication, by providing winning variants with more resources, more usage, more distribution or more copies.
For example, suppose that three sales quotation methods are already in use in the organisation. The most popular saves time in sales but creates more work in fulfilment, making it the most expensive overall. A second method lowers total cost without reducing quality. A third method is slower but cheap to maintain and can produce acceptable quotes without using the core AI model. Selection would ideally retire the first, scale the second and keep the third running at a limited scale.
The selection machinery would track each quote, link it to the method that produced it, and evaluate the end-to-end success of the method, including corrections, delays and costs downstream. Agreed thresholds would trigger scaling of the method or a human review. The third (backup) variant would retain an owner and be tested periodically with the core AI model unavailable. Management would review the selection machine’s dashboard and tweak the parameters rather than approve every variation individually.
Selection must result in changes to the business, not just produce a ranking of variants. Winning variants need an owner, support and a route to wider deployment. Retained alternatives also need owners and periodic checks that they remain usable. Variants that no longer justify their costs must be eased out of use so that complexity can be constrained. A selection log and audit trail should be maintained so that continuous learning on selection effectiveness is possible. Testing, promotion, retention and retirement must be parts of an integrated system. In time that system may acquire variants too.
Evolvable
Organisations need to continually evolve. There is little benefit in selecting a single best workflow and freezing it into place as a one-off event. A variant that is inferior under today’s conditions may prove valuable when demand changes or a critical system fails. Not every variant is worth keeping. But there is value in maintaining a portfolio of variants that offer useful responses to different conditions. The capability being built is evolvability. This is the ability to keep producing responses that work even as conditions change. Selection machinery must reduce needless complexity without selecting away the capacity to adapt.
Watching the slot machine, Luria realised that introducing the virus had not created resistance. It had revealed which bacteria already possessed it. For firms, the lesson is to keep useful alternatives workable before a shock reveals which ones they need. Firms that fail to build selection machines to harness front-line-driven mutation are gambling with their futures.
Sources
Luria, S. E., and Delbrück, M. (1943). “Mutations of Bacteria from Virus Sensitivity to Virus Resistance.” Genetics, 28(6), 491–511. The original paper establishing that bacterial resistance could arise before exposure to the virus.
Murray, A. (2016). “Salvador Luria and Max Delbrück on Random Mutation and Fluctuation Tests.” Genetics, 202(2), 367–368. An account of the experiment and Luria’s insight after watching a colleague win a three-dollar slot-machine jackpot.
NobelPrize.org. “The Nobel Prize in Physiology or Medicine 1969.” Official record of the prize shared by Max Delbrück, Alfred D. Hershey and Salvador E. Luria.
Microsoft (2019). “Heathrow Airport inspires employee engagement with Microsoft Power Apps.” Microsoft Customer Stories, 7 November. Supplier-published case study describing Samit Saini’s language-book app and the development of employee-created applications at Heathrow.
Accenture (undated). “Turning process friction into flow.” Company case study describing the identification of 14,000 requisition-processing paths across 50 countries despite the use of common enterprise systems.
Bobrow, D. G., and Whalen, J. (2002). “Community Knowledge Sharing in Practice: The Eureka Story.” Reflections: The SoL Journal, 4(2), 47–59. An account of Xerox’s technician-led knowledge-sharing system and how front-line practical knowledge was shared across the organisation.
Norsk Hydro (2021). “2019: Cyber-attack on Hydro.” Updated 11 June. Company account of the March 2019 attack, its uneven operational impact and the use of manual procedures to maintain production.
Briggs, B. (2019). “Hackers hit Norsk Hydro with ransomware. The company responded with transparency.” Microsoft Source, 16 December. Includes the account of retired employees returning to help operate older paper-based processes.
Worthen, B. (2002). “Nestlé’s Enterprise Resource Planning (ERP) Odyssey.” CIO, 15 May. Original reporting on Nestlé USA’s fragmented purchasing and information systems, including the 29 vanilla prices, nine general ledgers and 28 points of customer entry.
Nelson, R. R., and Winter, S. G. (1982). An Evolutionary Theory of Economic Change. Cambridge, Massachusetts: Belknap Press of Harvard University Press. Foundational treatment of organisational routines, search and evolutionary economic change.
OpenAI (2025). “How BBVA is scaling AI from pilot to practice across the org.” 6 November. Supplier-published case study reporting more than 20,000 custom GPTs created across BBVA, with approximately 4,000 used frequently.
Demirer, M., Musolff, L., and Yang, L. (2026). “Writing Code vs. Shipping Code: Productivity Effects Across Generations of AI Coding Tools.” NBER Working Paper No. 35275, issued May; revised September. The revised study covers more than 500,000 GitHub developers and reports estimated cumulative increases of approximately 240% in commits, 80% in projects and 30% in releases.
Cresta (undated). “How Aptive Drove $2.37M in Retention with Cresta AI.” Supplier-published case study of AI-assisted customer retention, including an increase in adherence to discovery behaviours from 28% to 59% over four months.
Aghion, P., Bloom, N., Lucking, B., Sadun, R., and Van Reenen, J. (2021). “Turbulence, Firm Decentralization, and Growth in Bad Times.” American Economic Journal: Applied Economics, 13(1), 133–169. Research on pre-crisis delegation to local plant managers and subsequent performance in sectors hardest hit by the Great Recession.
Kim, E. M., Garg, A., Peng, K., and Garg, N. (2025). “Correlated Errors in Large Language Models.” Proceedings of the 42nd International Conference on Machine Learning, Proceedings of Machine Learning Research, 267, 30038–30066. Evidence that different models can make correlated errors, including implications for model-based evaluation.