AI Task Automation Won’t Save Software Product Management

The forecasted economic impact of Artificial Intelligence is expected to have a profound effect on the future of work. Understandably while the AI build out is still in its infancy, forecasts of the size of the economic impact, and the effect on the global labour force vary significantly. Estimates centre around job role exposure, task automation, and job displacement.   

We have seen many technological revolutions throughout history. Could this just be the next iteration of computing with Natural Language as the abstraction layer? Probably not.

Transformational claims of achieving super intelligence are rife and indeed. Many examples of new value creation exist in the form of newly created markets:

But there’s a catch. The fundamental principles of building digital products are changing.

Traditional software paradigms such as Software as a Service (SaaS) have the benefit of having near-zero marginal cost for the acquisition of new customers. Generative AI on the other hand incurs heavy compute costs, known as inference cost of goods sold (COGS), which must be paid on each invocation in order to continuously serve users.

Contrary to infrastructure required to build out previous technology revolutions (e.g. the internet) AI compute assets depreciate rapidly. Server silicon requires replacement every 3 to 5 years. The CapEx investment is not a one-time sunk cost. Maintaining the fleet requires massive recurring reinvestment. Furthermore, Generative AI solutions are non-deterministic, which introduces additional complexity and overhead in terms of quality assurance and lifecycle management in comparison to traditional software.

Perhaps the most significant paradigm shift, with the return of software agents, is the ability for systems to take autonomous action rather than being restricted to the automation of repetitive actions, e.g. via Robot Process Automation (RPA), or the commands of a human operator. Escalating token costs and cyber security risks have become a feature of the AI rollout. It seems unclear when, if at all, systems will move past “human-in-the-loop” / “human-on-the-loop”; however, clearly this is not out of the realms of possibility. Examples of “out-of-the-loop” systems, such as high frequency trading have existed for some time (with their own guardrails and circuit breakers).

Regarding economic impact, claims that the proliferation of AI technology will lead to unprecedented material abundance, e.g. The Techno-Optimist Manifesto, still seem a far cry from the realities of what co-pilots, chat assistants, and “agentic” systems are able to deliver today. Companies in 2026 are still struggling to realise ROI from their AI initiatives [1,2,3], despite automation of tasks such as content generation and code writing leading to the cost of prototyping / Minimum Viable Product (MVP) development drastically reducing.

But tasks such as code writing were never the primary cost driver in total cost of ownership (TCO) calculations. Looking at task automation in isolation is misleading when looking at new value creation. Tasks must be looked at in aggregate and how the potential efficiency gains translate into upside generation (often positioned as freeing up end users to focus on other value generating tasks) rather than just cost savings. As Mik Kersen notes in Project to Product, treating IT development as a cost centre leads to initiatives measured on project management metrics (scope delivered, delivery time, budget) rather than outcomes (higher adoption, fewer defects, increased revenue).  

As AI automation makes idea generation, artifact generation, and hypothesis testing cheaper, care needs to be taken that this does not merely shift cost elsewhere, as seen in [4]’s analysis of code generation for example.

So where does all this leave your average Product Manager today? How can they adapt to the coming wave of AI, so vividly outlined in Mustafa Sulyman’s book?

1. Paying the AI Infrastructure Bill

In their benchmark research on AI economic growth, Goldman Sachs economists projected that “AI adoption could eventually drive a 7% or almost $7 trillion increase in annual global GDP over a 10-year period” [5]. Similarly, in the same year Mckinsey Global Research Institute estimated a similarly large economic impact of $2.6 trillion to $4.4 trillion in potential economic value added annually through AI adoption across 63 enterprise use cases [6].

In contrast to the consultancies and investment bank’s projections, in [7] Daron Acemoglu estimated a more modest uplift of +1.1% to +1.8% in 10-year cumulative GDP. Actual productivity numbers in recent analyses published by central bank economists would tend to agree with Acemoglu’s more modest outlook: [8] reported +4.0% labor productivity gain (revenue/value-added per employee) at adopting firms with no net labour displacement in the short term; while [9] found micro-level experiments show workers completing individual tasks faster, but taken in aggregate, localized task gains have not yet produced a measurable change in macroeconomic output.

While there clearly is evidence to back the positive business case for AI adoption, the question persists if the limited upside reported thus far can justify the huge capital outlays required to fuel the roll out. Goldman Sachs Global Institute [10], estimates roughly $7.6 trillion of cumulative global AI CapEx spend between 2026 and 2031.cumulative global. To sustain this compute footprint, the technology sector could potentially need to renew a significant percentage of its global hardware fleet every year.

With such a large investment in assets that depreciate within 48 months, the bill cannot be paid simply by helping knowledge workers draft content and software engineers generate code faster. If the value delivered is merely fragmented, task-level time savings that fail to expand enterprise margins, the return on invested capital (ROIC) collapses.

Ensuring that return requires AI value creation to outpace physical asset depreciation in the long run. Roughly speaking there are a few scenarios that will help enable this for companies adopting AI:

  1. Transitioning from task augmentation to end-to-end process automation: whether the process automation ultimately results in labour displacement or replacement is dependent upon a number of factors including the complexity of the process in question and the underlying business model. Even when an AI product creates a new market such as autonomous mobility, unit economics still matter. A one hour trip takes one hour regardless if the driver is a human or an AI. The gig economy was successful because it allowed technology companies to avoid holding depreciating physical assets. For end-to-end process automation to even be a viable option, the capital cost of the system needs to drop below the fully loaded cost of employing someone to carry out the same work.

  2. Collapse infrastructure overheads and inference COGS: utilisation of proprietary Large Language Model (LLM) APIs takes away the need to manage model hosting infrastructure, but inference token costs can scale infinitely. Open Source  LLMs address the high inference costs but require managing hosting infrastructure. Further engineering and architecture innovations, such as Mixture of Experts (MOE), will be required to bring down costs.

  3. Expand the value ceiling (profound physical & scientific breakthroughs for humankind): These are the moonshoots. New drug discovery, democratisation of access to education etc. and enable companies to transform their business models.

2. The Cost Efficiency Fallacy

For the average AI Product Manager, a strategy geared towards end-to-end process automation would seem the more feasible path to a positive return on an AI product development initiative. This would also seem most closely aligned with the modeling of AI value creation in the studies discussed previously, where projected value created by AI is overwhelmingly predicated on task-efficiency, not labour expansion. 

For example, across 8 of the 9 macroeconomic scenarios modeled in [5], direct worker augmentation, rather than labor displacement and replacement, serves as the primary driver of projected economic growth, accounting for approximately 80% of the annual labor productivity boost in their baseline case.

But does modern enterprise AI deployment risk falling into the trap of automating organizational waste? Below are three examples of how this may be manifesting itself in common AI automation use cases today.

2.1 Customer Service: Subsidizing Failure Demand

In Systems Thinking [11], customer interactions are divided into two distinct categories:

  • Value Demand: A customer reaching out to expand their relationship, purchase a product, or unlock a new capability.

  • Failure Demand: Demand caused by an organization’s failure to deliver a working product or service (“Where is my confirmation?”, “Why did my authentication fail?”, “How do I export this CSV?”).

Deploying an AI customer service bot is a standard AI use case that serves failure demand to achieve labour savings, while simultaneously leaving the root cause unaddressed. More crucially, it is difficult to realise value on something that is a) an expected part of the service and b) easily replicable by competitors.

In contrast, AI deployed toward driving user engagement and increased conversion serves value demand and creates differentiation, but is more challenging to implement.

2.2 Agile Backlog Management: Polishing Obsolete Inventory

In lean product development, a large Jira backlog is unrealized inventory. Backlogs bloated with hundreds of feature ideas, edge-case tickets, and speculative improvements that sit idle for months, even years, should simply be deleted.

Instead, the temptation is to use generative AI to update existing user stories with standard acceptance criteria, cluster them into themes, and build roadmaps is another appealing and standard AI use case that is often used to demonstrate progress in organisational transformation programs. But this amounts to polishing obsolete inventory and merely mechanizing agile ceremony, while only delivering incremental value. 

2.3 Meeting Summarisation and Note Taking: Amplifying Communication Overhead

Automated meeting summaries and note taking is another common AI use case that creates the illusion of efficiency gains without creating value. Removing meeting friction can lead to more meetings that generate more content, which can often be summarised into notes, expanded into summary decks, and later contracted again. Generating written content that costs zero marginal human energy to free up worker capacity without a clear strategic purpose can lead to teams filling the vacuum by focusing on work that serves more failure demand and ultimately creates more communication overhead.

3. The Modern Product Manager’s Challenge: Commoditised Intelligence and Disciplined Push Back 

Generative AI has fundamentally commoditized task-level intelligence: handling routine customer interactions, generating Product Requirement Documents (PRDs), summarising workshop notes, generating code; commoditising routine, labour intensive tasks previously requiring human skills.

Redirecting freed up worker’s capacity to value generating tasks and end-to-end process automation requires AI Product Managers to focus on diagnosing and killing failure demand, choosing the high-leverage problems worth solving, evaluating the economic trade-offs, and defining go-to-market strategy. As historian Yuval Noah Harari has observed: “Intelligence is the ability to solve problems; wisdom is the ability to know which problems to solve.” [12]

Delivering on the AI promise also requires disciplined push back against organisational pressure and market hype to deliver efficiency gains. Standard AI use causes that only deliver incremental value, ultimately won’t pay the infrastructure bill.

The AI revolution is undeniably here to stay and will ultimately have a profound and transformational impact on the global economy.

But the AI build out cannot be underpinned by the automation of workplace bureaucracy alone. Transforming the coming wave into durable economic value requires AI Product Managers that can look past the temptation of cheap task efficiency and focus on serving value demand.

References

  1. Look Past Productivity to Get Real Value from AI
  2. The state of AI in 2026: On the road to ROI
  3. AI ROI Fails to Outpace Spend for 57% of Enterprises
  4. DORA 2025 State of AI-assisted Software Development Report
  5. The Potentially Large Effects of Artificial Intelligence on Economic Growth
  6. The Economic Potential of Generative AI: The Next Productivity Frontier
  7. The Simple Macroeconomics of AI.” MIT and NBER Working Paper No. 32487
  8. AI adoption, productivity and employment: evidence from European firms
  9. The AI Buildout and the Economy: Publicly Available Data to Assess AI’s Impact
  10. Tracking Trillions: The Assumptions Shaping the Scale of the AI Build-Out
  11. Systems Thinking in the Public Sector: The Failure of the Reform Regime… and a Manifesto for a Better Way
  12. Nexus: A Brief History of Information Networks from the Stone Age to AI. Random House / Documentary and Lecture Series
  13. Image attribution: A vibrant and high-contrast blog thumbnail designed for maximum readability at small sizes, with a split composition.On the left, a dark teal and cyan background filled with glowing digital circuitry and a large, complex mechanical gear. Obstructing the gear is a giant, bold red ‘X’ with a heavy black outline, symbolizing a stop or failure. On the right, a warm orange and yellow gradient background featuring two simple, large icons in dark orange: a human profile with a glowing brain/cog (representing human product strategy) and a product roadmap/planning map with a start, path, and checkpoint flag.Spanning the bottom and right-center is huge, impact-style text with thick dark drop shadows, designed to be read at tiny resolutions. The text ‘AI’ is in fiery red with an orange outline, ‘WON’T’ in stark white, and ‘SAVE PM’ in bright yellow. The image has thick outlines, highly saturated colors, and no fine details to ensure it remains sharp and impactful even when scaled down to 140×90 pixels. Inspired by modern digital media and corporate tech aesthetics.16:9 –style raw prompt,  Nano Banana 2.1, 27 Oct. 2026, https://gemini.google.com/