The VRIO Test for AI Use Cases: Why Deep and Narrow Beats a Hundred Pilots
Share
Only 4% of companies take a focused approach to AI. Those companies get twice the ROI of everyone else.
That number comes from a 2024 BCG survey of 1,000 CXOs and senior executives, cited in the Harvard Business Review article "Stop Running So Many AI Pilots" by Goutam Challagalla, Mahwesh Khan, and Fabrice Beaulieu. Read it slowly. Ninety-six percent of large enterprises are doing AI the way that produces half the return.
If you are a COO or transformation lead counting your active AI pilots right now and quietly wondering why the P&L hasn't moved, this post is for you. The problem is not your models, your vendors, or your data lake. It is the shape of your deployment.
Why AI pilot sprawl produces so little
The HBR authors call the common pattern "shallow and broad": many unrelated, one-off use cases scattered across the company. A few automated processes in accounting. A few in supply chain. Some in marketing, finance, HR. Each one justified by a clean, immediate ROI case.
The logic feels responsible. Deploy widely, see what works, fund the winners.
It fails for four reasons the article makes uncomfortably concrete.
The advantage is copyable. Shallow use cases sit on top of commodity capability. If your competitor can buy the same tool next quarter, you have bought a temporary efficiency, not a position.
Nothing gets rethought. When gen AI touches only 5% to 10% of the tasks inside a function, no executive has any reason to redesign that function. The old process survives, now with a chatbot attached to it.
Initiative fatigue sets in. Use cases running without a larger objective wear people down. When staff see that the tenth pilot changed nothing about their week, skepticism replaces curiosity — and you have spent your change budget buying resistance.
ROI becomes the wrong scoreboard. Immediate-return screening systematically selects against the deep work, because deep work accumulates return later and larger.
Reckitt hit this fork in late 2023. The proposed use cases spanned the business — drafting presentations, customer support, procurement contract optimization — and most guaranteed time savings. Management approved none of them as a strategy. Their conclusion, in the article's framing: pursuing unrelated use cases would not transform strategy or create meaningful advantage.
What deep and narrow actually means
Deep and narrow means concentrating AI intensively in one area where the tasks are interconnected, then fundamentally rethinking how the work is done.
Here is the part most readers miss, and it matters more than anything else in the article. Depth can be scoped two ways:
- A single function. Reckitt chose marketing. L'Oréal chose the consumer journey.
- A single end-to-end process. A large regional lender — which HBR anonymizes under the pseudonym "Acme Bank" — chose its mortgage process.
The bank example is the one to study, because a process crosses the silos a function cannot. The lender's original mortgage flow ran adviser → loan officer → underwriter, with manual data entry into the origination system and separate document verification at each pass. The bank rebuilt it: automated processing of all mortgage paperwork, plus a conversational interface that lets underwriters interrogate predictive models and generate approval or rejection explanations. Decision quality, speed, and customer responsiveness all improved.
Note what came first. The team mapped the essential steps, evaluated where it made sense to work differently, and eliminated redundancies. In the article's own words: "The team developed a more-effective process, and gen AI made it even faster and easier."
Process redesign was the intervention. AI was the accelerant.
Reckitt's results after less than two years of deliberate adoption: product concepts generated up to 60% faster, and brand and marketing communication processes 30% or more efficient depending on the process. The article is explicit that these gains would not have been possible without concentrating investment on interconnected tasks inside one domain.
How to go deep and narrow in four steps
The article gives a four-step method. Each step, in practice, is a process question.
Step 1 — Find the most promising domain
AI opportunity sits on a spectrum: office productivity (meeting summaries, decks), domain reinvention (end-to-end rethinking of a process or function), and value generation (new business models, AI infused into products). Office productivity will not buy you an advantage. New business models are closed to many firms — you cannot infuse AI into toothpaste.
Domain reinvention is open to nearly everyone. That is where you go.
The lender chose mortgages because they carried large margins, sat at the center of the customer relationship, and drove cross-sell. Strategic weight, not technical convenience, set the target.
Step 2 — Anchor it to a lasting advantage
Deep deployment should protect an existing strength or create one that is hard to copy. Combine AI with distinctive expertise, data, capability, scale, or customer relationships.
L'Oréal could build its Beauty Genius diagnostic assistant because it had been compounding augmented-reality capability since 2015 and could draw on advanced skin-biology research. Competitors can clone the chatbot. They cannot clone the research. IKEA is compressing interior design projects toward ten minutes of designer time — against consultants charging around $99 per room over several days — on the back of the largest interior-designer community and project archive in the category.
Ask the honest version of this question: which of our strengths does this redesign compound?
This is also where the article stops short. It tells you to anchor to a lasting advantage but gives you no test for whether you have one. There is a forty-year-old test that fits perfectly.
The VRIO test for AI use cases
VRIO screens a resource on four questions: is it Valuable, is it Rare, is it Inimitable, and is the firm Organized to capture its value? Map the article onto it and the fit is almost word for word. Strategic weight is V. Reckitt's customer data and L'Oréal's skin-biology research are R. "Easily copied by competitors" is a failed I. And "wholesale changes and a reimagining of how work is done" is O.
Now the uncomfortable part.
Foundation models are the least rare resource in enterprise history. Your competitor gets the same model next quarter at the same price. So value, rarity, and inimitability cannot come from the AI. They come from the proprietary data and domain expertise you point it at — and from the process architecture that makes its output usable.
Which means O is the binding constraint. It is also the one letter almost nobody tests.
Most enterprises already hold valuable, rare, hard-to-copy resources that sit unexploited, because the process, decision rights, handoffs, and standards around them were never redesigned. VRIO is blunt about that outcome: an unorganized firm captures competitive parity at best, and value it cannot hold.
Use the four tests as a gate before funding anything:
| Test | Question | Fail signal | Verdict |
|---|---|---|---|
| Valuable | Does it improve a process that moves margin, risk, or customer outcome? | It saves time on internal admin | Office productivity, not advantage |
| Rare | Does it draw on data, expertise, or scale competitors lack? | It runs on generic inputs any rival has | Parity at best |
| Inimitable | Could a competitor replicate it in two quarters with the same vendor? | Yes | Temporary gain — cap the investment |
| Organized | Are the process, roles, decision rights, and standards redesigned to exploit it? | Old process, AI bolted on | Unrealized advantage — this is where value leaks |
Only use cases passing all four earn deep-and-narrow investment. Everything else is an experiment, which HBR says is perfectly acceptable — as long as you stop calling it transformation.
Run this gate across your current portfolio and pilot sprawl stops looking like enthusiasm. A hundred pilots is a hundred use cases screened on V for efficiency, and never screened on R, I, or O at all.
Step 3 — Sequence efficiency before growth
Start with cost efficiency inside the chosen domain. Cost comes down faster than revenue goes up, and early wins buy political permission.
Reckitt ran exactly this sequence. Five initial pilots proved immediate marketing efficiencies — one showed media campaign analyses moving from days to hours. Executives were convinced. Only then did Reckitt go deep across marketing and turn the capability toward growth.
So pilots are not the enemy. Pilots as a permanent operating state are.
Reckitt then did the unglamorous work: it surveyed and interviewed 2,000-plus global marketers, catalogued 300 distinct tasks, estimated time spent on each, and plotted every task on two axes — degree of possible automation and opportunity size. That analysis narrowed the field to roughly 100 tasks worth targeting.
Read that again. Before the transformation came a full task-level map of the domain.
Step 4 — Watch the competitive gap
Your competitors are running the same play. Ask whether your top rival could replicate a valuable strength of yours using AI. They do not need to match you — a good-enough, simpler, cheaper alternative is sufficient to erode your position. Deploy in ways that widen the gap.
The step nobody budgets for
Change management is 70% of the challenge. Perfecting data is 20%. Using gen AI effectively is 10%.
That breakdown, from the article, is the price tag on the letter O — and it should reorganize your programme plan. Almost every enterprise AI budget is inverted against it — heavy on tooling, thin on the redesign, governance, documentation, and adoption work that determines whether anything sticks.
This is also where deep and narrow gets hard. It demands wholesale change: new process designs, reallocated roles, rewritten procedures, retrained staff, evidence that the new way is being followed. Most organisations do not have a system for that. They have consultants, slide decks, and a Visio file somebody owns.
Which brings us to the practical question the article leaves open. If depth is the answer, what do you actually run it on?
ESSAM: the operating layer for deep and narrow
ESSAM is an E2E Agentic Process OS. It exists because O is where deep-and-narrow programmes die.
Valuable, rare, and inimitable are analytical judgments — a good strategist can reach them in a workshop. Organized is engineering work: redesigned value flows, rewritten policies and procedures, RACI and OLA, approval controls, deployment to the people who run the process, and evidence that the new way is actually being followed. That is not a workshop output. It is the 70%, and it needs a system of record rather than a project.
ESSAM maps the process conversationally, classifies every step as value-adding or non-value-adding, produces process statistics and diagnosis, then generates an optimized redesign with rationale you can challenge. It assesses degree of change, runs FMEA on implementation risk, and generates the artifacts that make the redesign real — SOPs, OLAs, RACI, policies and procedures, board-ready and PDCA A3 exports — with routing, approval, versioning, and audit trail built in.
Then it deploys. Governed processes reach front-line staff through Teams or WhatsApp instead of a portal nobody opens, collects their feedback, and feeds it back into continuous improvement. It also surfaces the AI use cases specific to each process — so your use-case pipeline comes from process evidence rather than from whoever pitched loudest.
Map that against the four steps:
| HBR step | What ESSAM does |
|---|---|
| Find the promising domain | Baselines processes and quantifies waste, cycle time, and non-value-added effort so target selection rests on evidence — the V test, measured |
| Anchor to lasting advantage | Surfaces AI use cases from process evidence rather than from whoever pitched loudest, then makes the value flow explicit so redesign compounds a real strength — and delivers O through generated SOPs, policies, RACI, OLA, approvals, and front-line deployment |
| Sequence efficiency then growth | Waste and efficiency scoring gives you the fast cost case first, with the redesign already documented for the next stage |
| Watch the competitive gap | Sprint 26's bidirectional BPM integration pushes governed designs into your runtime and returns execution telemetry, exceptions, SLA breaches, and cycle times for continuous re-optimization |
That last row is the piece HBR does not cover, and it closes the VRIO loop. The article tells you to go deep. It does not tell you how to know the redesign held six months later. Telemetry flowing back into the process design is O-verification: proof the organisation actually changed, not merely that a design was approved. It is also what turns a one-off deep bet into a capability that keeps compounding — the kind rivals genuinely cannot copy, because they would have to reproduce your entire improvement loop, not your tool stack.
ESSAM is GDPR, ISO/IEC 27001:2022, and SOC 2 Type II certified, and is in active use with enterprise process owners in banking today.
Pick one process. Go deep.
You do not need a hundred pilots. You need one process that matters, mapped honestly, redesigned properly, deployed to the people who run it, and measured continuously.
Bring us the process you would most like to fix — mortgage origination, onboarding, trade finance, claims, procurement, whatever is quietly costing you the most. We will baseline it with you live and show you the waste, the redesign, and the AI use cases hiding inside it.
Which process in your organisation would you go deep on first — and what has been stopping you?