Synergised Consulting
Method

Where AI Automation Actually Pays Off: What Rockwell and Citi Just Proved

Synergised Consulting Ltd

7 min read
Title card: Where AI Automation Actually Pays Off, on Synergised's cream ground with coral rule

Last Updated: 16 November 2026

AI automation pays off where the work is rule-bound, high in volume, low in exceptions and already measured on a number the business tracks. Two large, independently documented programmes this year prove the point from opposite directions: Rockwell Automation's Singapore plant cut defects by 35% only after redesigning how work flows through the site, and Citigroup is reworking roughly 100 manual processes end to end rather than adding tools to unchanged workflows. The pattern on the losing side is just as consistent, and it is the one an owner in their first 100 days needs to recognise early.

What the Rockwell Numbers Actually Measure

The Rockwell Singapore figures describe a redesigned site, not a model deployment: units per person-hour up 43%, defects down 35% and time-to-competency down 67%, measured on operational metrics that existed before the technology arrived. That is what makes them credible and what separates them from a marketing percentage.

According to the World Economic Forum's Global Lighthouse Network case study [1], the Singapore site runs a high-mix, low-volume operation with more than 1,000 stock-keeping units and over 20,000 changeovers a year, and it deployed more than 50 digital and AI solutions across flexible automation, AI-enabled quality control and intelligent maintenance to get there. Singapore's Economic Development Board independently describes the same three results [2]. Read the three figures together and they describe a change in how work flows: throughput, quality and the speed at which a new operator becomes useful. None of them is a model benchmark.

The useful lesson sits in the measurement, not the percentages. Every one of those numbers was tracked the same way before and after the programme, over a long enough period to exclude one good quarter, with the process change described. That is the test a buyer should apply to any case study, including a flattering one.

The three questions that test any vendor case study

When a case study quotes a result, ask three things. What was the denominator, meaning which metric was measured and was it tracked before the programme started. How long is the measurement window, because a percentage measured over a quarter can be a good quarter. And what changed in the process, because if the answer is only "we added the tool", the result describes the site's redesign or its baseline conditions, not the tool. If the write-up cannot answer all three, treat the percentage as marketing and move on. Microsoft's own customer story about the same Rockwell programme, for example, is a vendor describing its own platform's role in the result, and its figures are not treated as independent evidence here [1].

Why the Same Tools Disappoint in an Office

The same technology bolted onto an unchanged process reliably fails, and the failure is now measured at scale: research from MIT's NANDA initiative found 95% of the organisations in its dataset got no measurable return from their generative AI pilots [3]. The gap is process design, not model quality.

According to Fortune's reporting on the MIT study [3], the research rested on 150 interviews with business leaders, a survey of 350 employees and an analysis of 300 public AI deployments, and it attributed the stall to what it calls a learning gap between the tools and the workflows they were dropped into. Generic tools excel at individual tasks because they are flexible. They stall inside a business because the surrounding process was never defined well enough for any tool, human or machine, to execute it consistently.

Citigroup is the counter-example running at scale. According to American Banker [4], Citi is redesigning about 100 of its most manual processes end to end with AI and automation, reviewing progress weekly, while most other banks surveyed are adding AI to existing workflows or making only partial changes, and seeing smaller returns for it. The comparison is the whole lesson: same technology, same year, same industry pressures. The bank that changed the processes expects the returns. The banks that added tools to unchanged processes are still waiting.

The Four Conditions That Decide Whether It Pays

A workflow is a candidate for automation only when it is rule-bound, high in volume, low in exceptions and already measured. If any one of those is missing, the honest first-100-days decision is to document and simplify the workflow first, and automate it later or not at all.

Each condition earns its place. Rule-bound means the steps can be written down so a person who has never done the task could follow them; judgement calls scattered through the work defeat a machine the same way they defeat delegation. High volume means the fixed cost of defining, testing and supervising the automation is spread across enough repetitions to be worth it. Low exceptions matters because supervision is the hidden cost: a workflow where many cases need a human decision needs a human in the loop permanently, and the economics often collapse there. Already measured means a baseline exists, so nobody has to invent one after the fact and argue about it afterwards. MIT found the largest actual returns in back-office automation rather than the sales and marketing uses most budgets chase [3], which is usually where these four conditions are easiest to satisfy.

A worked example of the test

Take a distribution business where customer orders arrive by email and a coordinator retypes each one into the order system, checks stock, and sends a confirmation. The work is rule-bound: every order takes the same steps. It is high volume: it happens dozens of times a day. Exceptions exist but are rare: a part-shipped line, a credit hold. And the cost is already visible: the coordinator's hours are a known quantity, and order errors show up as returns and credit notes the business already counts. That workflow passes all four conditions, and the automation decision is genuinely available. Now take the same business's handling of a large customer's quarterly pricing dispute: low volume, judgement throughout, no defined process, no metric. Automating that would be spending money on work that was never a workflow. The example is illustrative, but the two shapes are common, and most owner-operated businesses hold several of each.

What an Owner Should Do in the First 100 Days

In the first 100 days, measure one already-tracked metric for one candidate workflow, document the process as it actually runs, and only then decide whether to automate it. The register this produces is the deliverable, and it costs almost nothing to keep.

What this looks like in practice, and it is short. Pick the workflow where work waits or repeats. Write down its actual steps, including the informal ones. Record the current number: hours spent per week, error or rework counts, cycle time from request to done. Then decide, on that evidence, whether the workflow meets the four conditions above. Where it does, automate with the baseline in place so the after is comparable to the before. Where it does not, the documentation itself is the improvement, because an undefined process cannot be delegated, sold or audited either. In discovery workshops, owners rarely arrive with a list of what to automate; the list gets built in the conversation, walking a normal week task by task, and the repetitive screen-based bundles separate themselves from work that genuinely needs a person. Citi's weekly review cadence is the same discipline scaled up: a named list, revisited on a schedule, against numbers already tracked [4].

A dated before/after register changes what the owner can stop asserting and start showing: not that the business runs better, but that a named workflow improved by a stated amount, on a metric that existed before the change, with the measurement method written down where a buyer's diligence team can read it.

Sources

  1. [1] World Economic Forum, Global Lighthouse Network case study: Rockwell Automation Singapore, published 22 June 2026.
  2. [2] Singapore Economic Development Board, "From Singapore to the world: How Rockwell Automation's local innovations and digital transformation are being scaled as global best practices", July 2026.
  3. [3] Fortune, "MIT report: 95% of generative AI pilots at companies are failing", 18 August 2025, reporting MIT NANDA's "The GenAI Divide: State of AI in Business 2025".
  4. [4] American Banker, "Exclusive research: Citi is redesigning 100 processes. What's stopping others?", 12 November 2026.

Background reading (qualitative guidance, no figures): The four-condition framing is our own operating synthesis, drawn from the published evidence above and from discovery-workshop practice; it carries no figures of its own.