ICYMI
Double (AI) Agents – What intelligence services can teach companies about AI agents
TL;DR
· AI models create no business value in isolation. Business value depends on deployment: workflow redesign, systems integration, training and governance.
· Most firms overfund access and underfund deployment. The Deployment Ratio is the dollars spent on deployment for every dollar of core AI cost (model access, tokens, licences and compute).
· History suggests that Deployment Ratios for new technologies can reach 10:1.
· AI may bend the pattern by self-deploying: agents mapping processes, engineering context, building integrations, designing workflows, generating training and enforcing controls.
· Firms delivering AI should track the ratio, fund deployment adequately, and progressively expand agent authority.
On the morning of 26 April 1956, a crane at Port Newark lifted 58 metal boxes onto a converted wartime tanker called the Ideal X, bound for Houston. Loading the Ideal X cost about 16 cents a ton, against $5.83 a ton for loose cargo loaded by hand, a 36× reduction. Malcom McLean’s shipping container technology was revolutionary. But capturing its benefits took decades of further investment. Standards had to be agreed. Ports had to be rebuilt. New labour deals had to be struck. The technology became significant only once the system around it changed.
AI follows a similar pattern: deployment costs can dwarf the visible cost of model access.
Core AI costs are transparent: model access, token fees, licences and compute. Deployment costs are scattered across the organisation and harder to see: workflow redesign, data readiness, systems integration, security, governance, evaluation, training and change management.
The Deployment Ratio is the dollars spent on deployment for every dollar spent on core AI costs. Underestimating deployment leads to inadequate budgets and, ultimately, to AI investments that fail to move the bottom line. MIT NANDA’s 2025 research reported that roughly 95 per cent of enterprise GenAI pilots delivered little to no measurable P&L impact.
Estimating the Deployment Ratio
The pattern is not unique to AI or shipping. When electricity first reached factories, a large electric motor simply drove the same line shafts that steam had turned. The gains came only when factories were redesigned around the workflow, with small motors at each machine. Computers followed the same pattern: Brynjolfsson, Hitt and Yang found roughly nine dollars of complementary organisational capital for each dollar of installed computer capital. ERP implementations run at 1.5–3× the licence fee on mid-market projects and 2–8× on complex SAP programmes.
Hard data on AI deployment is scarce, and the estimates are not apples-to-apples, but they point to similar ratios. An enterprise executive quoted by a16z puts LLMs at roughly a quarter of the cost of building use cases; BCG’s 10-20-70 rule puts algorithms at about a tenth of transformation effort.
We can triangulate with an indicative planning model. Depending on maturity, complexity and regulatory exposure, we might put data preparation at 0.5–4× core AI cost, integration and engineering at 1–3×, specialist talent at 1.5–4×, evaluation and monitoring at 0.3–1×, governance and security at 0.3–2×, and change management and process redesign at 1–3×. If added mechanically, those components imply roughly 5–17×. Because they overlap and share fixed costs, I would use 5–15× as a planning envelope, with 8–10× a defensible working assumption for a typical large enterprise.
The ratio varies by maturity. Pilots may need one dollar of deployment work for every dollar of core AI spend. Enterprise scaling may need eight to ten, as data, integration, training and control costs arrive. Later the ratio falls, as templates appear, skills spread and vendors bundle complementary capabilities with core AI.
Could AI Self-Deploy?
AI might bend the pattern if it can assist with its own deployment. This is not an agent waking up one morning and rewiring the company. Agents reduce the ratio when they automate specific parts of the deployment loop, provided they have context, bounded authority, audit trails, escalation routes and accountable humans who can accept or reject the new design. Without those conditions, agents raise the ratio, requiring more verification and control than non-agentic AI. Self-deployment can take six forms:
The first mechanism is process discovery. Agents read procedures, transcripts and logs to map how work is actually done, including exceptions.
The second is context engineering. Many deployment failures are context failures, so agents build the context layer, classifying documents, reconciling schemas and turning tacit knowledge into reusable assets.
The third is systems integration. Agents map APIs, write adapters and generate test cases.
The fourth is workflow design. Agents decide which steps to automate, where human approval stays mandatory, and which metrics define success. A useful agent does not merely automate the old process; it helps design a better one.
The fifth is training. For each new workflow, agents generate role-specific training and just-in-time coaching, driving down change-management costs.
The sixth is control. Self-deploying agents need AgentOps: machine identities, human owners, permissions, evaluations, budgets, audit logs, kill switches and escalation paths, with authority expanding as measured reliability justifies it.
Halving the Ratio
The evidence for self-deployment is thin. AWS’s Thomson Reuters case study reports a 4× velocity gain from AWS Transform, with modernisation moving from months to a two-week sprint. Moderna provides an operating-model anecdote, with more than 3,000 custom GPTs and a people-and-technology function merged around the redesign of work. Nubank uses a published framework that links context engineering, human-in-the-loop prompt iteration and evaluation to measured gains.
As a planning assumption, a mature enterprise applying all six mechanisms could plausibly halve the long-run deployment ratio, from 8–10× to around 4–5×. Early deployments may sit near today’s 8-10x range while the shared context, integration, training and AgentOps assets are built. But those assets are reusable, so each deployment should be cheaper than the last. Highly standardised workflows might reach 2–3×, while regulated or legacy-heavy environments stay closer to 5–8×. The target is not zero deployment cost, but a halving of the ratio over time.
Actions for Executives
· Track the Deployment Ratio and how it evolves. Record the full cost of deployment alongside core AI costs. A rising ratio can be healthy if the organisation is moving from pilots to production. A falling ratio can be healthy if agent-assisted deployment is reducing marginal cost.
· Fund deployment in the same budget cycle as core AI. Do not approve model access now and hope to find the change budget later.
· Make context and systems agent-readable. Create a single context layer across customer, product, finance, risk, policy and workflow data, with clean ownership, metadata, lineage and access controls. Expose high-value systems through APIs rather than relying on agents clicking through brittle user interfaces.
· Use agents for system integration and modernisation. Let agents read legacy code, API documentation, database schemas, logs and tickets. Ask them to propose mappings, generate adapters, create test cases, draft migration plans and open pull requests.
· Use AI for workflow design before software is built. Feed process maps, SOPs, screen recordings, call transcripts and ticket histories into workflow-design agents. Ask for the target operating model, exception paths, control points, roles, service levels and KPIs before approving a new build.
· Make training a by-product of deployment. Each new AI-enabled workflow should automatically produce role-specific micro-courses, simulations, checklists and just-in-time help.
· Govern agent authority in stages. Agree early which processes agents may inspect, which recommendations they may make, which actions they may take and where human approval is required. Treat agents as workers, with human owners, machine identities, permissions, budgets, audit logs, evaluations and rollback plans. Increase agentic authority (recommend, then draft, then execute) based on measured reliability, not enthusiasm.
Questions for the Board
· Do we track our Deployment Ratio and how it is evolving?
· Is our budget sufficient to effectively deploy the AI we buy?
· Are we enlisting AI to help reduce the costs of deployment?
· What would have to be true for the next AI deployment to cost half as much as the last one?
Metal shipping containers did little to transform trade until the system around the boxes changed. AI may eventually help build more of that system itself. But until it does, the winners will not be the companies that buy the most intelligence. They will be the companies that fund the work needed to deploy it.
Sources & Notes
Source caveat. Several figures below should be read as directional rather than universal. Historical deployment ratios are rarely clean budget lines; they often combine physical infrastructure, labour agreements, organisational redesign and intangible capital.
Containerisation. Marc Levinson, The Box: How the Shipping Container Made the World Smaller and the World Economy Bigger, Princeton University Press, 2006. Source for the Ideal X voyage, the 58 containers, the Port Newark-to-Houston route and the loading-cost comparison.
Factory electrification. Paul A. David, “The Dynamo and the Computer,” American Economic Review, 1990; Warren D. Devine Jr., “From Shafts to Wires: Historical Perspective on Electrification,” Journal of Economic History, 1983. Source for the idea that electrification paid off after factories reorganised around unit drive rather than simply replacing steam with a central electric motor.
Computer capital and organisational capital. Erik Brynjolfsson, Lorin M. Hitt and Shinkyu Yang, “Intangible Assets: Computers and Organizational Capital,” Brookings Papers on Economic Activity, 2002. Market-valuation study of 1,216 US firms, 1987–97, finding roughly $9 of complementary organisational capital associated with each $1 of computer capital. An econometric inference, not a literal budget line.
ERP deployment costs. Implementation-to-licence ratios from ERP Research and industry benchmarks: software licences are typically only 20–30 per cent of first-year cost, while implementation (integration, configuration, data migration, training and change management) runs 1.5–3× the licence fee, and 2–8× on complex SAP S/4HANA programmes. Benchmark and outcome data: Panorama Consulting Group, annual ERP Report, drawn from thousands of implementations, with an average implementation of about $450,000 in 2025. For large SAP R/3 installations, less than 20 per cent of the typical $20 million cost was hardware and software; the rest was consultants, process redesign and training (Brynjolfsson et al.).
Pilot failure. MIT NANDA, “The GenAI Divide: State of AI in Business 2025.” Source of the 95 per cent figure; based on interviews, a worker survey and deployment analysis, so treated here as directional.
External deployment-ratio estimates. a16z, “16 Changes to the Way Enterprises Are Building and Buying Generative AI” (Sarah Wang and Shangda Xu, 21 March 2024), is based on interviews with dozens of Fortune 500 and top enterprise leaders and a survey of 70 more; it says that model access alone is not enough, that implementation and scaling require specialised technical talent, that implementation was one of the largest AI-spend areas in 2023, and quotes an enterprise executive saying LLMs are roughly a quarter of use-case build cost. BCG’s 10-20-70 rule treats algorithms as about 10 per cent of transformation effort, with the rest in data, technology, people, process and change. No published study computes a clean enterprise foundation-model deployment ratio; the 5–15× band is a synthesis of these sources and historical precedent.
Component ranges. The category-level ranges are an author synthesis for planning purposes, not directly observed benchmark ratios. Directional support comes from a16z on implementation and specialised talent; D. Sculley et al., “Hidden Technical Debt in Machine Learning Systems,” NIPS 2015, on production ML maintenance costs, glue code, configuration, monitoring and pipeline jungles; Eric Daimler and Ryan Wisnesky, “Informal Data Transformation Considered Harmful,” arXiv, 2020, on enterprise data integrity and data-cleaning burdens; NIST AI 600-1, “Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile,” July 2024, on governance, measurement, security, privacy, monitoring and value-chain/component-integration risks; Deloitte CFO Signals Q3 2023 and “What Does Generative AI-Ready Look Like for Finance?” on talent, data readiness and workflow integration; and MIT NANDA’s GenAI Divide report on workflow adaptation and integration. The ranges should be read as overlapping, not additive audited budget lines; “talent” means incremental specialist capability not already counted elsewhere.
Self-deployment ratio estimate. The estimate that a mature self-deployment stack could lower repeatable-use-case ratios to 4–5×, and in narrow standardised workflows perhaps 2–3×, is an author synthesis rather than an observed benchmark. It extends the planning ranges above by assuming the organisation has reusable context, integration, training, workflow-design and AgentOps assets. Fragmented, regulated or legacy-heavy environments may remain above 5×.
Agentic modernisation anecdotes. AWS Transform product materials describe agentic modernisation of mainframe, VMware, Windows and legacy code, including automation of assessments, code analysis, refactoring, dependency mapping and transformation planning. AWS customer stories used directionally here include Thomson Reuters (4× velocity improvement, 1.5 million lines of code modernised per month, transformation time moving from months to a two-week sprint), ADP (thousands of business rules extracted in hours, more than 90 per cent manual-effort reduction and 80 per cent faster rule extraction), IDEMIA (4× faster application transformation and 30 per cent TCO reduction) and CSL (10× faster initial wave planning, 12× faster application discovery and 10.5 weeks saved across 1,072 applications). These are vendor-published case studies, so they should be treated as useful anecdotes rather than independent benchmarks.
Moderna. Wall Street Journal reporting, May 2025, on the merger of Moderna’s HR and technology functions under Tracey Franklin as Chief People and Digital Technology Officer; Moderna–OpenAI partnership materials for the 3,000+ custom GPTs embedded across legal, research, manufacturing and HR. Company-published figures, so directional.
Production support-agent deployment. Nubank, “Building Customer Support AI Agents at 100M-User Scale: An Evaluation-Driven Framework,” arXiv, June 2026, is used for the importance of context engineering, human-in-the-loop prompt iteration, LLM-judge evaluation and ideation-to-production validation. The reported card-delivery deployment showed a 37 percentage-point improvement in AI transactional NPS and a 29 percentage-point gain in self-service rate versus prior agent variants; treated as one firm’s published account, not a universal result.