Unilever’s digital twin rollout
sensors first

At a food factory in Poznan, one filling line stops for minor reasons a fifth less often than it did a year ago. At the same moment, the plant digital twin proposes a correction, and that proposal sits on a screen waiting for a person to approve it. Put those two facts side by side and you have the real content of Unilever’s expansion plan: the model can already work out what to change, and letting it act on that is still gated by a human.

Forty sites in eighteen months, from five plants

Unilever is working with Accenture to put factory digital twins into roughly forty locations over the next eighteen months. The number has a basis: five plants that already run the technology. They are a personal care plant in Raeford, North Carolina; a food factory in Poznan, Poland; a personal care plant in Gandhidham, India; a home care facility in Cu Chi, Vietnam; and one more in the United States.

Four of them have reported results. Raeford cut waste by a fifth and raised capacity by a tenth. Poznan reduced minor stoppages by up to twenty per cent. Gandhidham cut product defects by thirty per cent over four years. Cu Chi saved one to two per cent on premium ingredients. Those figures come from Unilever’s own release, so they are vendor numbers, and each one is a reading from a single plant over a particular period rather than an industry statistic.

Line the four readings up and something easy to miss appears. The first two showed up soon after the work was done. The third took four years. The same technology paid back at rates that differ by an order of magnitude, so any budget that treats the four as equally fast will draw the wrong cash-flow curve.

The spread of the pilot list matters too. The five plants sit in the United States, Poland, India and Vietnam, across personal care, food and home care. The usual pattern is to prove a concept in one region and copy it sideways. Here the pilot set was multi-regional from the start, which surfaces the hard part early: regulation, language, operating habits and supply chains all differ, and one template will not carry across them.

The twin proposes, a person approves

Cuthbert, the Unilever executive responsible for this work, described the current state plainly. What the twins deliver today are suggested corrections that require human approval before anything is done. Agents inside the technology can already act on low-risk processes, which the company frames as paving the way to more independence later. The phrase that carries weight is low-risk, because it is doing all the work in that sentence.

The sentence worth writing down is the next one. As the team learns how humans and agents can coexist, she said, that is adjusting depending on the process area every week. That overturns a widespread assumption. Most governance designs assume a boundary is set once and then executed; here the boundary is a variable renegotiated weekly, and the unit of negotiation is the process area, not the plant and certainly not the group.

Why the smallest unit is a process area can be worked out from mechanics. How tight a boundary should be depends on what it costs to correct the action if it is wrong. A wrong temperature setpoint on a filling line is fixed in minutes. A wrong ingredient ratio can scrap a whole batch. A wrong interlock action stops the line, triggers an investigation, and sometimes an incident report. Inside one plant those correction costs differ by orders of magnitude, so they should not share a single line.

That is also why autonomy is not a switch you flip for a site. It is a table filled cell by cell, and each cell records whether an action can run on its own, who approves it when it cannot, and what the approver looks at. The finer the table, the more cells can be opened. An enterprise holding one blanket agent-permission policy is effectively managing every cell as if it were the most expensive kind of action: it pays in throughput and buys nothing in surprise.

The scarce input is sensor placement, not models

Cuthbert was more transparent than most vendors about how sites get picked. Unilever runs several hundred factories worldwide. A plant qualifies when the technology can significantly improve and standardise a particular operation, with the goal of replicating the reworked process elsewhere. One further criterion is harder: the plants chosen tend to be the more technically advanced ones, because they already have the infrastructure, the cybersecurity and the shop-floor sensors.

Her phrasing was that there is a minimum amount they need in place to collect information at the right parts of the process. That is what makes forty sites a conditional number. The candidate pool is not several hundred factories; it is the subset whose instruments already sit at the right points. Without them the twin has no input, and the project stops being a software project and becomes an equipment project with a different order of cost and schedule.

The common enterprise mistake at this step is to rank the expansion by business appetite first and send engineering to fill the data gaps afterwards. Reversing the order costs far less. List the plants that already capture data at the critical process steps, treat that list as the candidate pool, and pick from it by business value. Instrumentation is the entry ticket, not the first task after kick-off.

One uncomfortable corollary follows. What sets the pace of expansion may be the equipment investment of previous years rather than the size of this year’s AI budget. A group that spent on models but never put sensors on its lines will find fewer plants qualify than it expected, and that gap cannot be bought back quickly; it can only be waited out.

No technical metric appears in the success criteria

Asked how these applications are judged, Cuthbert said the criteria are wired as hard to the business as possible: a specific service impact, an improvement in agility or responsiveness, and the classical productivity and growth measures. Model accuracy is not on that list, and neither is automation rate.

That detail works as a method. When the acceptance standard is written in business language, every step of the project has to translate back into business language, and any step that cannot translate exposes itself in review. Write the standard as how many processes the agent took over instead, and the team optimises for takeover. A rising takeover rate does not mean the business improved; it means people stepped out, not that output went up.

One benefit is easy to overlook. Cuthbert said the capability she is most looking forward to is predicting from sensor data when equipment will need maintenance. The value is not only a stoppage avoided; it is spare parts arriving early, which pulls down inventory and emergency purchasing. That payoff lands on the supply-chain and inventory lines, not on line efficiency. Cost cases that only count line efficiency will understate what these applications actually move.

Then there is the human half. The reporting mentions a gamified training tool Unilever uses to build digital skills across planning and factory operations, and Cuthbert warned against underestimating the organisational and cultural impact, or the need to bring the organisation along. That is the budget line most often cut: training and redefined roles. Cut it, and performance drifts down after go-live, because nobody on site knows what the system is for.

What one replicated plant actually costs

Bring the threads together and replication splits into four blocks. First comes the instrument survey: what the key process steps already capture, and what is missing. Second comes the judgement work of writing the rules that experienced operators carry in their heads into something executable, usually the most labour-intensive block. Third is boundary negotiation, process area by process area. Fourth is training and role adjustment, so the people on site understand the new division of work.

Only the first block can be budgeted in advance. The other three float with conditions on site. That explains a familiar pattern: software licences are a small share of the total cost, and overruns almost always happen in judgement and boundary work. Buyers who estimate total schedule from a software quote will end up optimistic.

Replication carries one more hidden cost. Every plant needs a person with both business authority and process knowledge to negotiate the boundaries. If that person cannot be named, the project parks itself in a state of technical readiness with unfinished process agreements. The better predictor of whether a group can spread a pilot is how many of its plants can produce such a person, not how many licences it has bought.

What this case does not prove

The limits first. Every site reading comes from one company’s release material with no independent audit, so it is not an industry benchmark. Results differ widely by product category, and a ten per cent capacity gain at one plant does not transfer to your line. The weekly adjustment the executive describes is the live state of one company already running digital twins, not an industry average. One of the four readings took four years to appear, so comparisons across them need the time axis attached.

What transfers is the mechanism, not the percentages. Three layers: approval boundaries are drawn per process area and get renegotiated on a schedule; the expansion pool is set by instrument placement rather than appetite; and success criteria are written in business language with technical metrics kept out of acceptance. None of the three requires a digital twin. All three apply wherever an agent is about to touch a physical action.

Our judgement

Our judgement is this: the obstacle between five plants and forty is not the model and not the licence, it is the ability to replicate an approval boundary. Every added plant requires renegotiating, process area by process area, what may run on its own, who signs off, and on what evidence. No software absorbs that work, because what it consumes is the judgement of people who know the process. The scarce resource is governance capacity, not compute.

Three moves follow. First, before kick-off, write down the boundary for every candidate process area: the action, the approver, the evidence, and a realistic estimate of how many days the negotiation takes. Plants that cannot finish that negotiation do not belong on the expansion list yet. Second, treat instrumentation as an entry condition rather than a post-kick-off task: list which data the critical steps already capture, what is missing, and what closing the gap costs in money and time. Third, write acceptance in business metrics and make the replication template a contractual deliverable, so what the buyer receives is a reusable process and its criteria rather than a platform licence.

The question worth leaving with a buyer is this: if you had to replicate tomorrow, could you name who negotiates the approval boundary at the next plant, against which standard, and how many days it takes? If not, what you hold is a working pilot, not yet a repeatable process.

References

Supply Chain Dive (via Yahoo Finance), "Unilever puts business results first in factory digital twin rollout", by Antone Gonsalves — source for the forty-site and eighteen-month plan, the five pilot plants and their readings, the human-approval state and weekly boundary adjustment, the site criteria and instrumentation minimum, and the gamified training; site readings are Unilever’s own release numbers