DrBee Insights · AI in Business

The Invisible75%

When UPS added up what it cost to make its routing AI actually work, the algorithm turned out to be the small line item. Morgan Stanley and DBS Bank, in completely different businesses, discovered the same thing. What all three spent the rest on is the most useful thing a business owner can learn about AI.

By DrBee · 10 August 2026 · 12-minute read

Everyone judges AI at the same moment: the demonstration.

A chatbot summarizes a document in seconds. A forecasting model produces a dashboard. An optimization system recommends a more efficient schedule. The demo works, the project gets a green light, and the business waits for results.

I have spent twenty-five years in the IT industry, and for most of them I have watched businesses being urged onto one AI journey or another: neural networks, then machine learning, now large language models. Very few truly succeeded. For a long time the explanation was access. In the IBM Watson era, serious AI arrived as an enterprise programme with an enterprise price, far beyond the reach of a small business. Today the equivalent capability sits on any desktop, some of it free.

Access has been solved; adoption has not. Official US measurement found fewer than one in five businesses using AI in any business function between December 2025 and May 2026, with the smallest firms adopting at around half the rate of the largest.^[11] The pattern is even sharper at home: Singapore's official measure put SME adoption at 14.5% in 2024, up from 4.2% a year earlier yet still barely a quarter of the 62.5% recorded among larger enterprises.^[13] And where firms do use it, the work stays at the shallow end: sales and marketing is the most common home for AI, and most adopters confine it to three or fewer functions.^[12] Everyone can now afford the model. Something else is still stopping them.

A single number points to the answer. When UPS built ORION, the route-optimization system that plans delivery sequences for its US drivers, testing and rollout represented more than 75% of project cost, and the deployment organization grew far larger than the team that built the algorithm. The clever mathematics that demonstrated so well was, financially, the small part.

That would be a logistics curiosity if it were only UPS. It isn't. Morgan Stanley, adopting generative AI in a regulated wealth-management business, ended up building something similar: not a better model, but a permanent apparatus around the model. DBS Bank, scaling AI across an entire bank, found its results came from infrastructure and governance that no demonstration ever shows. Three different industries, three different kinds of AI (a route optimizer, a retrieval-based assistant, a portfolio of hundreds of models) and the same inverted bill.

So this article asks a discussion question rather than selling an answer: what exactly is in the invisible 75%, and why did three organizations that share almost nothing all end up paying for it? The pattern that emerges is not obvious, and it is a transferable lesson at any scale of business.

The model is what you buy. The system that proves it, constrains it and repeats it is what you build. The building is most of the bill.

75%+of UPS ORION's cost was testing and rollout, not the algorithm

US$50Msaved per year for every 1.6 km cut from average daily routes

15→3months to deliver an AI use case at DBS after platform reuse

<1 in 5US businesses using AI in any function, Dec 2025 to May 2026

Lesson One

The AI was the cheap part

What the money actually bought at UPS

US$50Msaved a year per 1.6 km trimmed from daily routesORION is an example of prescriptive analytics. Instead of predicting what may happen, it recommends what should be done: a delivery sequence that satisfies pickup commitments while cutting unnecessary distance. The business case was clear: UPS estimated that cutting just 1.6 kilometres (one mile) from the average daily distance travelled across its US routes could save approximately US$50 million a year.

The algorithm cleared its benchmarks early. Then the real spending started, because the laboratory routes failed the road. Initial sequences were inconsistent and difficult for drivers to follow; the optimization did not reflect enough reality, and commercial map data turned out to be too inaccurate to trust. What followed was roughly three years of intensive field testing: one operating location, then two more, then another eight, the question shifting from “does it work?” to “can we reproduce it?” This ran alongside continuous validation of map and operational data, and “ORION rides” that let decision-makers experience prototype routes instead of judging a presentation.^[1][2]

It worked, and endured. By December 2015, more than 35,000 US drivers were using ORION; INFORMS reported savings above US$320 million by then, against an estimated full-deployment cost of US$250 million. These are historical programme figures, not a current audited profit number. But UPS was still reporting in 2025 that AI and machine-learning route planning (the capability that grew out of ORION) saved approximately 16 to 23 kilometres (10 to 14 miles) per driver per day.^[3]

The same inversion at Morgan Stanley

Morgan Stanley did not build its model at all: its internal Assistant is GPT-4-based, retrieving from controlled firm content, later joined by Debrief, which turns consented meeting recordings into notes and draft follow-ups that advisers review. The model was, in effect, bought.

What the firm built was an evaluation factory. Subject-matter experts and prompt engineers graded answers for accuracy and coherence. Test sets covered summarization, translation and different meeting types. Morgan Stanley refined retrieval methods as the knowledge collection expanded. A daily regression suite tested sample questions for newly introduced weaknesses. Generative AI can answer the same question differently, and a minor model or prompt change can improve one task while degrading another. A zero-data-retention arrangement kept proprietary data out of public model training.^[4][5]

None of that machinery is the AI. All of it is why the AI could be used.

And at DBS

DBS began strengthening its data and AI capability in 2014, and its headline numbers are large: more than 2,000 models across over 430 use cases reported in 2025, generating approximately SGD 1 billion in economic value from data analytics, artificial intelligence and machine learning.^[6] These are company-reported figures. Public disclosure does not allow the SGD 1 billion calculation to be independently reproduced or AI to be separated from data analytics. It indicates a mature internal measurement programme, not an audited causal result for every model.

But look at where the capability came from. A central data and AI platform, reusable components, templates and automation cut typical use-case delivery from 12 to 15 months down to two to three months, by the bank's own account.^[7] A 700-person “Data Chapter” kept central expertise while embedding data professionals in business squads; since 2021, more than 9,000 employees have taken data and AI courses.^[8] Platforms, templates, training, people: plumbing. The plumbing is what made the four-hundredth use case cheap.

What this means for the rest of us

The uncomfortable takeaway is that most AI budgets are built upside down. The visible licence is usually the smallest cost; real adoption includes data preparation, integration, testing, security, training, exception handling, workflow changes and maintenance. Even when the model itself is cheap, the operating change around it is not. A business that budgets only for the software has funded the demonstration, not the adoption.

At a ten-person firm the invisible 75% does not disappear. It changes currency: from dollars to owner-hours. The small-business equivalent of UPS's three years of field testing is a few weeks of running the tool against real jobs (kilometres per completed job, punctuality, overtime) with the people doing the work empowered to flag what the software cannot see. Unglamorous, and decisive.

Lesson Two

Trust was manufactured, not hoped for

Here is the second thing the demo never shows: the moment a real person has to decide whether to believe the machine.

UPS hit it immediately. Experienced drivers did not automatically accept that a counter-intuitive, computer-generated route was better than their own judgement. Telling them to trust it was never going to work. So UPS engineered reasons to believe: frontline drivers and site managers helped distinguish a theoretically efficient route from a drivable one, and the “ORION rides” existed precisely so that sceptics could experience the route rather than take it on faith. Crucially, UPS also tracked the percentage of time drivers followed the ORION route as a leading indicator (trust, measured weekly) and linked it to the chain of longer-term results: continued use after the deployment team left, lower average distance driven, reduced fuel and costs.

98%adoption among adviser teamsMorgan Stanley's version was starker, because a fluent, unsupported answer in financial services could mislead an adviser, violate policy or damage client trust. Its response was to make reliability itself the product: answers grounded only in approved firm content, graded evaluations, a daily regression suite, and a specific division of responsibility. The AI prepared and retrieved; qualified advisers checked and decided. The result was near-universal use: Morgan Stanley reported 98% adoption among financial-adviser teams, and its vendor reported that accessible document coverage increased from 20% to 80%, with follow-ups that took days happening within hours.

Morgan Stanley and its vendor report these results. No independently verified public figure shows the profit created. High usage suggests utility, but is not the same as return on investment.

DBS built trust into its decision rules. Its PURE principles require data use to be Purposeful, Unsurprising, Respectful and Explainable. The question is not only “can we use it?” but “should we?” This sits alongside materiality assessments, a model register, mandatory controls and senior accountability, with early generative-AI deployments kept internal under substantial human oversight.^[9] Its customer-facing corporate assistant, DBS Joy, came later: used by more than 20,000 corporate and SME customers and, DBS reports, contributing to a 23% increase in customer-satisfaction scores since its phased launch in 2025.

The pattern across all three: trust is a built artifact. It has components (evidence the user can inspect, boundaries on what the system may touch, a named human who remains accountable) and those components cost money and design attention, which is why they live in the invisible 75%. The demo answers “can it do the task?” The trust system answers a different question: “under what conditions can its answer be believed?” Not every use needs the same amount of it:

Low consequence

Internal first draft

User review and basic data rules

Material but reversible

Product recommendation or staff scheduling

Tested boundaries, monitoring and escalation

High consequence

Financial, employment, safety or customer-entitlement decision

Validated data, documented evaluations, approval, audit trail and appeal

A small business can build the Morgan Stanley trust machine in miniature, and the recipe is concrete. Start with an internal assistant using approved procedures and product documents. Build 50 to 100 questions from actual work. Record the expected source and acceptable answer. Test again whenever the model, prompt or content changes. Do not let it send consequential advice directly to a client until the failure modes are understood. Writing a question and recording its expected answer takes roughly ten to fifteen minutes, so the full set is two to three working days of senior attention. That time, not any subscription fee, is the real price of being able to believe your own tool.

Lesson Three

AI fails at handover, not at launch

The least-told story in the UPS case is what happened after success. At some locations, performance declined once the implementation team left. The algorithm had not changed. The routes had not changed. What remained unchanged was the management system: local managers were still measuring drivers against past route performance, while ORION required comparison with a newly calculated ideal. The technology had changed; the yardstick had not. The old yardstick quietly pulled behaviour back to the old process.

This is worth sitting with, because it names the silent killer of AI projects. An organization's existing measures are its immune system. Introduce a new way of working without redesigning what managers measure people against, and the immune system rejects the transplant: not at launch, when everyone is watching, but months later, when no one is.

9 OMTsend-to-end processes DBS redesigned in 2025The other two cases show what deliberate handover looks like. Morgan Stanley made evaluation a continuing operating function, not a launch phase: the daily regression suite and accountable-review roles are permanent parts of how the work is done. DBS went furthest: in 2025 it reported nine “Operating Model Transformations” that redesigned processes end-to-end for human-AI collaboration and reskilled the people in them. The same programme is reshaping employment itself. In February 2025, DBS said it expected about 4,000 temporary and contract roles to reduce over three years through natural attrition as AI takes on more work, while creating around 1,000 new AI-related jobs.^[10] Scale cuts in both directions; an honest account of AI value includes the workforce consequences alongside the economic ones.

Handover discipline also answers the question every owner eventually asks: “We bought the tool. Why hasn't revenue moved?” An output creates no financial value until it changes an action or decision:

AI output→ Changed work→ Operational improvement→ Financial or customer outcome

An AI tool may reduce the time needed to draft a quotation. That is useful, but it does not guarantee more revenue. Staff must use the saved time to send quotations sooner or handle more qualified enquiries; customers must then respond positively; and the additional wins must be commercially worthwhile.

If revenue remains unchanged, the business can trace the chain to find out why. If drafts are faster but quotations are not sent sooner, the problem is workflow or adoption. If quotations are sent sooner but the win rate does not improve, speed may not be the real sales constraint. If more work is won but revenue still does not rise, pricing, job value or delivery capacity may be limiting the result. Revenue is therefore an important final outcome, but monitoring it alone cannot explain whether the AI worked or where the expected value was lost.

That is why the project definition matters before any tool is chosen. “Use AI in customer service” is not a project definition. “Reduce the time required to resolve routine delivery-status enquiries from eight minutes to three without increasing repeat contacts” is testable: it names the constraint, the baseline and the evidence in one sentence, and it tells you at handover whether the new system is actually running or quietly reverting.

A robot ship captain in a chart room studies a glowing holographic model of a full iceberg with a route plotted around it through three waypoints

“Which business constraint are we changing, what else must change around the AI, and what evidence will prove that the new system works?”

The demonstration is real. It is also the cheap part.

What to take away

UPS, Morgan Stanley and DBS did not discover a guaranteed formula for AI success. Their results depend on their industries, scale, data and execution, and their published numbers vary in evidential strength: some are operational case findings, others are company or vendor reports. What is transferable is the shape of the investment. Now it has a name. The invisible 75% is three things:

The proving system

Field tests, evaluation sets, data validation: the machinery that turns “it demos well” into “it works here.”

The trust system

Evidence, boundaries, accountable humans: the reasons a person can believe the answer.

The handover system

Redesigned measures, workflows and roles: what stops the old process growing back.

A team that cannot answer all three parts does not yet have an AI adoption plan. It has a technology purchase.