10 beliefs on how to get value from GenAI
Those who seek simple checklists in applying GenAI will fail.
Thinking in systems
My favorite coffee shop has an enterprise technology vibe. You hear people discussing SQL query syntax on video conferences, as if text-to-SQL weren’t a thing. One of these days, I will launch a podcast sitcom (in the style of an old-fashioned radio comedy series) called the coffee shoppers. In the pilot episode, the regulars will conspire to rid their favorite establishment of a matcha machine that attracts the wrong element, like hipsters.
One of the regulars (also a devoted reader) asked me to start including a book recommendation related to the issue’s theme. Since I prattled on about Military Power: Explaining Victory in Defeat in Modern Battle last week, let me suggest Thinking in Systems by Donella Meadows.
What drives bad enterprise technology? Short term and siloed thinking. When you implement a security control without understanding the user experience, that’s siloed thinking. When you build a use case without worrying about what the underlying platform might be, that’s the same mistake. It leads to high cost, technical debt and frustrated users. You see a lot of it as companies wrestle with deploying GenAI.
Donella Meadows defined systems thinking as the discipline of understanding how the behavior of a whole emerges from the interactions of its parts over time. It focuses on the relationships, feedback loops and delays across a system that spans the human and technological domains. You might say systems thinking implies seeing the world as a graph — and using that insight to optimize the entirety of a business domain rather than individual pieces of it. The ten beliefs that follow are my attempt to apply that insight to AI strategy.
A colleague asked if I could get a perspective on GenAI strategy on a couple of pages. That’s tough — specificity, nuance and general applicability are uneasy bedfellows. How do you navigate between the Scylla of platitude and the Charybdis of over-simplification? You have to stop thinking about a collection of technologies and a collection of organizations, but of a business-technology system where each part influences every other part in some way.
10 beliefs on how to get value from GenAI
1. The business world is entropic. Generative AI represents a technology discontinuity because it can process entropy rather than ignore or externalize it, e.g.
Complex and interconnected, incompletely documented business rules (e.g. like the ones in 1000-page contracts)
Unstructured or uncorrelated data (e.g. doctors’ notes in eHR)
Unstructured user instructions (natural language input)
Latent information (e.g. status of B2B commercial proposal)
2. GenAI represents a strategic discontinuity because it allows granular analysis and digitization of business domains where you previously needed humans to manage complex and ambiguous information
The largest opportunities will be in the most entropic domains because companies successfully digitized less entropic ones (like consumer auto insurance) years ago
Returns to investment will be lower for B2C businesses than in B2B businesses
They will be lower in operations for process manufacturing than in service operations
They will be high in domains like R&D and product development where digitization has been limited
3. Speed and quality will be as important, if not more important, than efficiency in achieving competitive advantage from GenAI
Returns from technology investment vary massively by period, sector and company — and companies compete away the value of productivity improvements as consumer surplus.
Updating a contract in minutes rather than days may be essential when negotiating a transaction
Reduced revenue leakage may swamp operational cost reduction in setting up the billing for an asset servicing or group health insurance arrangement
You can also use GenAI to create less frustrating user experiences, even if the impact is harder to quantify — even more importantly you can use it to reshape markets by reducing commercial friction and transforming unit costs.
4. The Jevons paradox applies to busywork as well as knowledge work, so you want cyborgs not androids
GenAI can rip the toil out of knowledge work and improve the rigor of your employees’ thinking — you need knowledge workers who can apply judgement and reason across multiple domains, rather than just provide specialized expertise
A lower unit cost of business insight will create demand for business analysis that had previously been uneconomic — if you are willing to rethink how you manage knowledge and even how you produce documents
But when you put GenAI on the desktop, employees will often devote tokens to writing and summarizing emails or performing more complex web searches — while GenAI can help your employees write more effectively, AI-generated prose is vague and unconvincing
5. Even though token prices have been declining roughly tenfold per year, the cost of inferencing will finally require enterprises to treat IT spend as COGS rather than G&A
Run costs will increase in relation to build costs
Companies will need to ensure both that they devote GenAI resources to valuable projects and that they design agentic systems for token efficiency
But inferencing latency may big a bigger problem than token costs
6. Every business system processes entropy somewhere — in product managers who formalize ambiguous workflows, account managers filling in CRM fields, free-text notes nobody reads. Agentic capabilities give you choices on where to put that work, each with a different cost, risk, and speed profile:
Agentic software engineering — automate the business logic that reflects organized complexity; apply your existing QA mechanisms to the output just as you would hand-written code
User interface — deploy chat or agentic interfaces where users complain about screen after screen of fields
Data ingestion — eighty percent of corporate data is unstructured; agents can ingest contracts, service notes, and RFQs that deterministic systems ignore entirely
Run-time decisions — use agentic patterns for heterogeneous decisions (e.g. architecture review); hybrid patterns (deterministic core, agents for edge cases) will become the dominant architecture
7. Context is existential for GenAI, and requires different mechanisms for managing it
Relational databases excel at storing transactions
Knowledge graphs allow you to store and analyze the relationships among entities (e.g. customers, products, process steps, service offerings) that GenAI discovers
Building an ontology and semantic layer (and encoding it as a knowledge graph) both reduces the likelihood of hallucination by over 80 percent and depicts your business as a system, allowing you to make better business decisions
8. Agentic technology will not make enterprise technology functions disappear. You must build robust common platforms in order to apply GenAI at scale, with efficiency, with security and with resiliency
Agentic software engineering replaces procedural programming with declarative programming
Doubling throughput and productivity will only be the start here for those companies that reimagine software engineering processes, rather than just providing engineers with copilot-type tools
Reimagining software engineering can transform the ROI of tech-for-tech investment and the “IT doom loop” and double or triple the EBITDA lift from enterprise technology by freeing up budgetary capacity for new investment and reducing value leakage and engineering deadweight loss
The challenge of Agentic cyberattacks will require enterprises to rethink risk management processes and increase the automation of their technology environments in order to remediate vulnerabilities discovered by frontier models like Mythos at speed
Agentic software engineering will change buy/build decisions and challenge “one size fits none” SaaS models. More enterprises will connect disparate software tools or will contract more niche vendors whose products more closely support their business processes
All of this requires treating enterprise technology as an integrated system and support from across the management team for the required investment and changes in the business-technology operating model
9. Agentic technology will change the relationship between users and the technology function
Users can achieve wonders by acting more like software engineers for “edge” applications
They will increasingly look like “Strats” who sit on a trading floor and update models intraday — the ED charge nurse who reweights a sepsis-triage rule when a new strain appears, the plant engineer who tunes a predictive-maintenance threshold when a new failure mode shows up, the claims supervisor who stands up a fraud agent overnight
10. Advantage will accrue to institutions that metabolize GenAI and scale its adoption
Capturing value will disrupt existing assumptions about organizational structure, ways of working — and especially the role of governance functions like HR and procurement
You need to treat the technology and the operations around it as a single system and evolve them together
You want to change everything someplace rather than something everyplace. Transforming a business domain (rather than funding disconnected use cases) creates local critical mass and provides visible results
Executive support in the face of resistance requires you start with hard problems (rather than easy ones) and a measurable target
GenAI is new and it is different. It does things that we could not get systems to do in any scalable way before. It applies to a far broader set of domains than traditional enterprise technology. Just as the railroad reshaped the geostrategic environment (empowering continental powers relative to maritime ones), GenAI will reshape the competitive dynamics in many sectors.
Those who seek simple checklists in applying GenAI will fail. You might ask why it is different — how it can allow us to automate new types of tasks. You must understand how it will interact with your strategic context and your organizational culture and capabilities. And then you can determine how you might use GenAI as a weapon. None of this is easy. But Edward Luttwak reminds us that strategic advantage derives from doing hard things, rather than easy ones.