The Cheap Part

In the summer of 2023, two unions and three organisations began the work of getting money to workers on plantations in Ecuador and Costa Rica. The findings had been documented for some time. The next two years went into letters, a formal filing, eighteen months of waiting before the union was recognised as a party to the proceeding, and negotiation. A mining cooperative that knows how it operates and cannot produce the licence application that would let it sell pays the same cost from the other side. This essay is about that part of the process, and about why it is getting cheaper.

I. What it cost to be heard

The banana that reaches a German supermarket has usually passed through very few hands. A plantation in Ecuador or Costa Rica, a trader, a ripening facility, a shelf. By the standards of this field the chain is short, which is one reason the retailers at the end of it have been able to say, over many years, that they know their supply chains reasonably well.

In the summer of 2023 that claim was put to a test. Oxfam Germany, working with the Ecuadorian agricultural workers’ union ASTAC and the Costa Rican union SITRAP, wrote to four German retailers, Aldi, Edeka, Lidl and Rewe, about conditions on plantations that supplied them.¹ According to Oxfam’s published account, the unions had documented wages below the legal minimum, aerial spraying of pesticides while workers were in the fields, discrimination against union members, and dismissals of older workers shortly before they would have qualified for pension entitlements.² The cases concerned several plantations across both countries and were presented to all four retailers at the same time.

The four companies responded differently. Each side’s public position is set out below, because the difference between them is the subject of this essay.

Two of the four retailers entered into discussions with the unions. Oxfam reports that Aldi engaged with SITRAP over conditions in Costa Rica and that, in 2025, affected workers there received payments from the producer.³ The amounts and the names of the recipients were kept confidential, and the settlement did not include an admission of liability. Lidl, according to its own published account, had already been operating a living-wage scheme for its banana supply since 2022.⁴

The other two retailers took a different view. Edeka stated publicly that it had examined Oxfam’s information, that the allegations could not be confirmed with respect to the farms from which it sourced bananas, and that it had provided Oxfam with a detailed written response and explained its position in person. Edeka rejected the suggestion that it had been unwilling to engage.⁵ Rewe stated that it was conducting a gap analysis of its banana supply chain and had drawn up an action plan for the plantation concerned.⁶ Oxfam’s account differs on one point: it says that the action plan was developed without involving the workers or their union, and that Edeka declined an invitation to meet the union and the workers directly.⁷ Both accounts are on the record, and readers can weigh them.

On 2 November 2023, Oxfam and ASTAC, with legal support from the European Center for Constitutional and Human Rights and funding from the Catholic development agency Misereor, filed a formal complaint against Edeka and Rewe with the Federal Office for Economic Affairs and Export Control, known by its German initials as BAFA.⁸ Some explanation is needed for readers outside Germany. Since January 2023, larger German companies have been subject to the Supply Chain Due Diligence Act, which requires them to identify and address human rights risks in their supply chains and gives affected people, and organisations acting on their behalf, the right to file complaints with the supervising authority. BAFA is that authority. The complaint against the two retailers was among the first filed under the new law.⁹ Twenty pages of documentation were submitted for each case. The complainants were workers in Ecuador, represented by Oxfam and ASTAC, who asked to remain anonymous for fear of reprisals.¹⁰

In late January 2024, BAFA found the complaint sufficiently substantiated to open a formal investigation.¹¹ According to ECCHR, it then took roughly eighteen months before ASTAC was recognised as a full party to the proceeding with access to the file.¹²

One further detail belongs in the record. The Rewe supplier at the centre of the Ecuadorian case had been certified by the Rainforest Alliance until July 2023. Oxfam states that the certification was withdrawn following the union’s complaint rather than through the certifier’s own audit process.¹³ The Costa Rican banana producers’ association, CORBANA, responded to the allegations in November 2023 by describing them as partly false, general or imprecise.¹⁴ That position, too, is part of the record.

Now consider what the process required.

Two unions with years of presence on the plantations, who knew which workers had been in the fields during spraying and which had been dismissed, and could say so with names. An international organisation to carry the case to Germany. A legal organisation to draft a filing that a supervisory authority would accept. A development agency to fund the work. Eighteen months of correspondence before a union was accepted as a party. Two years from the first letter to a payment, for plantations whose conditions had, on the unions’ account, been known locally for some time.

That is the pattern this essay is concerned with, and it recurs across the field. The evidence about a supply chain problem often exists, or could be produced. The scarce part is the capacity to carry a finding from the people who have it to the party that has to act on it, in a form that party will accept, while protecting the people who provided it. Two unions and three organisations for one case is roughly what that capacity currently costs. There are many thousands of plantations.

The same gap runs in the opposite direction. A mining cooperative in Peru knows where it works, who works there, and how much comes out of the ground. It cannot produce the environmental filing, the land agreement and the technical dossier that would turn that knowledge into a licence without years of a support organisation’s time. Workers who want a condition changed and producers who want to exist on paper are stopped at the same point. Both know. Neither can produce the document an institution will accept. Part VI takes up that side.

A second observation from the case is smaller, and I state it without drawing a conclusion. Three of the four retailers approached in 2023 had, by then, made public commitments on living wages in banana production. Aldi and Rewe were members of a working group of German retailers that had committed in late 2021 to a phased introduction of living-wage criteria for own-brand bananas, an arrangement the German competition authority had examined and cleared.¹⁵ Lidl had adopted its own scheme.¹⁶ Edeka, according to Oxfam’s 2026 review, had set no such targets.¹⁷ Companies that had found a way to cooperate on wage standards found themselves responding separately when a complaint arrived about conditions on the same plantations. Part VIII returns to why that might be, and what would change it.

II. Why joint action is less constrained than it looks

Ask why buyers do not act together on a shared supplier, and the first answer is usually competition law. Competitors coordinating toward a common supplier risk being treated as a cartel, the argument runs, which is why industry initiatives route through neutral third parties and why nobody has built a shared complaints mechanism. The concern deserves a proper answer, and the answer is shorter than the concern suggests. What follows is context, not legal advice.

In June 2023 the European Commission adopted revised guidelines on cooperation between competitors that include, for the first time, a chapter on sustainability agreements.¹⁸ The chapter recognises that many such agreements do not affect competition at all, and for those that set a common standard it provides a soft safe harbour: where participation is voluntary, the process is open, no commercially sensitive information changes hands beyond what the standard requires, and consumers do not face higher prices or less choice, the Commission will not treat the agreement as restricting competition.¹⁹ Agreements outside the safe harbour are assessed case by case and can still be exempted where their benefits outweigh their effects.²⁰

The German competition authority, the Bundeskartellamt, has applied exactly this reasoning to bananas. In November 2021 it examined a proposal by a working group of German retailers, coordinated by the development agency GIZ, to phase living-wage criteria into their own-brand banana sourcing. Its case report found no infringement of German or European law, and the reasons it gave are the ones that matter for anything proposed later in this essay: no exchange of information on purchase prices, costs, volumes or margins; no agreement on how any resulting costs would be passed on; voluntary participation; own-brand products only; and a production standard rather than a pricing arrangement.²¹ The same authority, in the same month, blocked a dairy proposal that would have set a uniform surcharge on raw milk across the whole market, and the contrast is instructive: a sustainability objective does not make a price agreement lawful, but it does not make a standards agreement unlawful either.²² The British and Dutch competition authorities published guidance in the same year that points the same way.²³

None of this addresses a shared complaints channel directly, and a lawyer asked about one would want to know what information passes between the buyers. Where a channel delivers pattern findings about labour conditions rather than supplier-level commercial detail, and where it is held by an organisation independent of the buyers, the exchange is limited to what any standard would require. That would need advice in each instance. But the direction of the published guidance is not in doubt.

So the obstacle sits somewhere else. The chair of the Dutch authority has said that companies avoid sustainability cooperation out of concern about infringements that would not arise, and the head of the Commission’s competition directorate observed that nobody had taken up its offer to discuss such agreements in advance.²⁴ Legal departments advise caution, caution costs nothing, and the boundary goes untested. That is a reasonable response to uncertainty. It is also a decision that can be revisited, which a legal barrier would not be. The question for the rest of this essay is therefore not whether buyers may act together, but what a shared arrangement would have to do to be worth building.

III. What the comparative evidence says

The research comparing complaint mechanisms across supply chain initiatives is not large, but it points in one direction.

James Harrison and Mark Wielga examined the complaint mechanisms of multi-stakeholder initiatives covering much of the world’s certified palm oil, sugar, timber and garments.²⁵ Of twenty-six mechanisms they identified, only six had published enough to assess. Across three of them, the palm oil roundtable, the sugar initiative Bonsucro and the Forest Stewardship Council, they found one case in which affected people had received even a potential remedy. Across the other three, the Bangladesh Accord, the Fair Wear Foundation and the Fair Labor Association, some form of remedy had been provided in between 41 and 58 per cent of adjudicated cases.²⁶ The numbers on the first group are small, which the authors note, and the comparison establishes a direction rather than a rate. Other work points the same way: a ten-year study of forty standard-setting initiatives found fewer than one in five with affected people in their governance, and reviews of the OECD’s national contact points, the one intergovernmental complaints system in this field, have found remedy to be rare.²⁷

The distinction Harrison and Wielga draw is not about procedure. A mechanism can meet every effectiveness criterion in the UN Guiding Principles and deliver nothing. They are separated by whether anything happens to the responsible party when a finding goes against them.

Three features mark the mechanisms that deliver. Affected people can trigger them without waiting for a buyer’s audit schedule. Consequences run through binding commercial commitments rather than goodwill. And the party doing the assessing is not paid by the party assessed.

The Accord is the case that shows what those three look like at scale, and it is current. It began in 2013 as a legally binding agreement between garment brands and two global union federations after the Rana Plaza collapse, and in April 2026 it was renewed for six years with around two hundred signatory brands, extended to Pakistan since 2023, and with a worker complaints mechanism that from late 2025 covers issues beyond safety.²⁸ Its enforcement is not a certificate. Brands that fail to require remediation at their suppliers can be taken to binding arbitration, and in 2017 the Permanent Court of Arbitration accepted claims against two brands, which settled the following year.²⁹ The union federations behind it have spent 2025 and 2026 arguing publicly, in Washington and elsewhere, that binding agreements of this kind outperform voluntary schemes, and the comparative research above is their evidence.³⁰

The refinement that matters for the rest of this essay is this. The Accord’s complaints mechanism is, formally, an internal one: complaints go to a body that works alongside the factories, and remedy happens inside them. It does not work because it is internal. It works because the consequence of ignoring a finding is external, and two hundred brands have signed a document saying so.

That is the condition on which everything that follows rests. An internal mechanism works when there is external leverage behind it.

IV. The channel that exists and is not used

There is already an industry that builds grievance channels for supply chains, and its experience is worth examining before proposing anything new.

For roughly a decade, a group of technology providers has offered what the sector calls worker voice tools: a hotline, a mobile application, an SMS survey, a messaging number, deployed at a factory or a farm and paid for either by the brand that sources from it or by the operator itself. Workers are informed that the channel exists. Some use it. The provider aggregates what comes in and reports upward. The reach can be substantial. One of the larger programmes in Southeast Asia, operated by the Issara Institute, reports more than half a million worker contacts over three years and remediation affecting over three hundred thousand people, on its own published figures.³¹

The channels exist, then, and the question is who uses them. On that, the research is not encouraging.

A study of digital worker feedback systems published in the Economic and Labour Relations Review, and an earlier assessment commissioned by the Open Society Foundations, reach similar conclusions from different starting points.³² When a channel is paid for by the employer or the brand, workers do not know who holds what they say. They do not know whether a complaint about a supervisor reaches that supervisor. They do not know whether the number on the poster in the canteen is a route to help or a route to a list. Both studies describe the trust problem as structural rather than incidental: not that particular tools were badly designed, but that a channel whose custody is unclear is a channel that workers have reason to avoid.

The programme that reports the largest reach is instructive on this point. The Issara Institute operates independently of the brands and the operators whose supply chains it works in, and says so on every channel it runs. Its own explanation for why workers call is not the technology but the fact that workers trust the organisation on the other end.³³

The same programme surfaced a second finding, smaller and more precise. When it examined who was using its channels, it found that women from one migrant group were underrepresented, and that the reason was not reluctance but access. Phone ownership among those women was low.³⁴

After a decade and considerable investment, the worker voice industry has demonstrated that a grievance channel cannot simply be installed. It is a claim about who holds what a worker says and what they will do with it, and workers assess that claim before they pick up the phone. The technology is comparatively easy. Custody is the difficult part, and trust in custody is local. It attaches to an organisation people already know, and it does not transfer to a brand’s logo or a provider’s application.

This gives the essay its constraint, and it applies on both sides of the cost described in Part I. A grievance channel cannot be scaled in the way a piece of software can. Neither can formalisation support. The organisations that walk a mining cooperative through licensing spend years at a site, and what they build in that time is the same thing the worker voice programmes depend on, an organisation that people already have reason to trust. That cannot be purchased or deployed, and it does not transfer.

The infrastructure beneath it can be scaled. That means the languages a model handles, the custody arrangements that survive a change of funder, the classification that keeps a case out of the wrong hands, and the templates that turn a spoken description of an operation into the filing an authority requires. If that layer is built once, local channels and local formalisation become affordable for the organisations that already hold the trust. If the tool is built first, it becomes another number on another poster, or another app that nobody at the site has a phone for.

V. Three complaints, three destinations

Most grievance mechanisms I have encountered treat complaints as a single category with one intake and one custody arrangement. Some of the ones I helped design did the same, and in retrospect the consequence was predictable. A mechanism built to handle wage disputes received, from time to time, an account that belonged somewhere else entirely, and because nothing in the design distinguished the two, the account went where wage disputes went. Nobody was harmed as far as I know. But that is a statement about luck rather than design. In practice complaints are three categories, distinguished not by what happened but by who is affected and who has the authority to remedy it, and the distinction matters because getting it wrong is the error that exposes people.

Path one: the operation can and should resolve it.

A supervisor who abuses his authority over a shift. A wage deduction the payroll office cannot explain. A machine running without its guard. Unpaid overtime. A harassment complaint between colleagues. In each of these the operator has a legal duty to act, is the fastest route to remedy, and may reasonably hold the record because it has to do something with it. An internal channel is not merely permitted here; in most jurisdictions it is required.

Part III established the condition, and it applies. Path one works when someone outside the operation can make a finding cost something. The Accord’s complaints mechanism is a path-one channel with two hundred brands behind it. A path-one channel with nothing behind it is a suggestion box.

Path two: the operation is the counterparty.

This path covers forced labour, recruitment fees paid to an agent in another country before the worker arrives, confiscated passports, discrimination against union members, and underpayment that is a policy rather than an error. It also covers retaliation against someone who used path one. In all of these an internal channel is not merely weak but structurally useless, because the recipient is the accused. Testimony has to sit outside the operator’s reach, and remedy runs through buyers, regulators or courts.

This is the path the banana case travelled. It is also the path taken by the workers at Top Glove, and their route illustrates how far the channel can be from the harm.

Top Glove is the world’s largest manufacturer of rubber gloves, based in Malaysia, with a workforce drawn substantially from Nepal, Bangladesh and Myanmar. Many of those workers had paid recruitment agents in their home countries to secure the job, in amounts equivalent to months of wages, and had arrived already in debt. Debt of that kind is among the indicators of forced labour identified by the International Labour Organization. The situation had been documented by NGOs and journalists over several years.³⁵

On 15 July 2020, US Customs and Border Protection issued a withhold release order on the company’s gloves, on the basis of information indicating that they were produced with forced labour, and detained shipments at the border. In March 2021 the agency escalated to a formal finding. The company engaged a British consultancy, Impactt, to verify remediation and to operate a confidential helpline for workers. By the third verification, in April 2021, Impactt reported that 36.3 million dollars in recruitment fees had been repaid to 12,676 current and former workers. On 9 September 2021 the finding was lifted. Malaysia’s trade ministry later estimated the cost to the company of the border action at 3.6 billion ringgit.³⁶

The workers’ effective channel was a customs authority on another continent. Not a hotline, not an ombudsman, not a union, but a border with the power to stop goods. And the fourteen months the process took ran from the moment an authority with that power acted, not from the moment the conditions were first documented, which was years earlier.

Path three: nobody inside the operation is the injured party.

A community’s spring changed three summers ago, or the village below a tailings dam does not trust it, or clearing has gone beyond a permit boundary, or dust and noise are crossing a fence line. The person reporting may not be a worker at all. Neither the operator nor the buyer is the natural addressee; the permitting authority is, and the buyer’s role, where it has one, is leverage rather than remedy.

Yanacocha in northern Peru is the reference case for what this path produces when it works as designed. On 2 June 2000, a contractor to the mine spilled 151 kilograms of elemental mercury along 41 kilometres of road passing through the village of Choropampa. Residents who did not know what the substance was collected it, and many were poisoned.³⁷ The mine was part-owned by the International Finance Corporation, the World Bank’s private lending arm, which has an independent accountability mechanism, the Compliance Advisor Ombudsman. That office oversaw an investigation in July 2000 which found significant discrepancies in the company’s waste management and emergency procedures.³⁸ In December 2000 a local defence committee lodged the first formal complaint. A dialogue roundtable involving more than fifty institutions followed, along with an independent water study and a monitoring programme, and in 2006 the complaint was closed.³⁹ Ten further complaints reached the ombudsman over the following two decades. The last, in December 2017, was supported by around a thousand documents: hair and blood tests, medical reports, photographs. The ombudsman’s review of that complaint, completed in 2020, found no grounds to open a compliance investigation. In the same month the complaint was filed, the IFC sold its shareholding in the mine.⁴⁰

Twenty years. Eleven complaints. A dialogue table that still exists. One early investigation that identified real failings, and no subsequent finding that the lender’s standards had been breached. The outcome was largely process, and process is not nothing: the water study was real, the monitoring was real, and the community had not had either before. But it is not remedy, and the distinction is one this essay returns to.

One intake, three custodies.

The three paths require the same intake and entirely different custody. A worker in a processing park who calls to describe a wage deduction and, in the course of describing it, mentions the fee she paid to an agent in Kathmandu has moved from path one to path two in a single sentence, probably without knowing it. The decision about which path applies is made at the moment someone begins to speak, often before they know themselves where their case belongs.

That is a design problem more than a governance one. A system that takes a spoken account and routes it correctly, rather than directing everything into one channel, is a bounded and testable piece of work. The failure that matters most is misclassification downward: a path-two case handled as path one delivers the complainant’s testimony to the party it accuses. Any design has to be built around avoiding that error, which means accepting some loss of speed in exchange for routing uncertain cases to external custody.

Companies gain from this arrangement as well, and it is worth saying so. Path one was always theirs to handle. Path three gives them early warning of a permit problem before it becomes a regulatory one. Path two is the category they cannot own, and a serious operator will acknowledge that, because a channel that claims to handle path-two cases internally is one that workers have reason to distrust.

The reason this classification has never been built is not that nobody thought of it. It is that sorting a spoken account into three destinations, in the speaker’s own language, at the moment it arrives, was work that only a trained person could do, and organisations with a trained person available for that could count them on one hand. That is the cost that has moved, and the next part is about what moves with it.

VI. What cheap analysis changes for the people on the ground

Suppose that analysis were free tomorrow.

The plantations in Part I would still have four customers, and the process would still have depended on which of them chose to engage. Yanacocha would still have a dialogue table. Top Glove’s workers would still have needed a customs authority on another continent. Analysis was not the binding constraint in any of those cases; enforcement and custody were. Falling analysis costs remove something narrower than it sounds. They lower the cost of being on record, and the cost of reading what the record says.

Three things have changed in roughly the last eighteen months that did not hold before, and they explain why this essay is written now rather than in 2023. The first is that models can now work with speech, not only text, in a widening set of languages that have no commercial market and therefore never had a product built for them. Coverage is uneven and largely unmeasured, which Part VII returns to, but it exists in a way it did not. The second is that models can now produce structured output reliably enough to fill a form, in the sense that a spoken account can become a set of fields an authority recognises, rather than a transcript somebody still has to work through. The third is cost. The analysis behind a single case, the reading and cross-referencing that used to take a trained person days, has fallen in price by more than an order of magnitude, and a channel that could afford to process one account a week can now process one an hour. None of these is a new capability in the sense the technology press uses the phrase. Together they change which organisations can afford to run a channel at all.

The change matters in three places, which share one infrastructure. I take them in the order of how contested they are, beginning with the one that has no opposing side.

Getting legal

A small-scale mining cooperative, or an association of smallholder farmers, faces a considerable body of documentation before it can sell into a formal supply chain. It needs licence applications, land records, a cooperative registration and environmental filings, then the traceability records a buyer requires before it will accept a shipment, and after that continuous reporting for as long as it remains inside the chain. Almost all of it is in a language and a register that nobody at the site writes fluently, and a large share of a support organisation’s time goes into producing it. This is among the main reasons formalisation programmes take years.

The figures are specific. In Peru, the environmental management instrument a small operation must file under the national formalisation regime costs, by one consultancy’s published estimate, between 3,000 and 12,000 soles, before the tax registration, the land-use agreement, favourable opinions from three separate agencies and a technical dossier are added.⁴¹ In Ghana, a country with over a million people engaged in artisanal mining, researchers found that active small-scale mining licences rarely numbered more than a thousand across two decades.⁴² On the agricultural side, the geolocation requirement under the European Union’s deforestation regulation costs a producer organisation between 5,000 and 15,000 euros for the mapping alone, according to Fairtrade International’s estimate.⁴³ A 2026 analysis of the Colombian coffee sector put compliance costs for a cooperative of five hundred farmers at between 105,000 and 215,000 dollars and estimated that up to 280,000 coffee-growing families could be excluded from the European market.⁴⁴ Not because they had deforested, but because they could not produce the documentation.

Under the arrangement this section describes, a miner or a cooperative secretary describes the operation by voice, in their own language, and a model produces the filing in the form the authority requires. The support organisation reviews and submits it. The same channel carries the ongoing compliance reporting: the daily record of who worked, what was extracted, what was sold and to whom, spoken at the end of a shift rather than written up at the end of a month. No party benefits from this process remaining slow, which distinguishes it from everything else in this essay.

Two qualifications are needed, because omitting them would be the kind of overstatement this essay is trying to avoid.

The first is that records do not make anyone honest. They make misrepresentation expensive and verification cheap, and they reward the cooperative that reports a rockfall over the one that does not. A buyer reading a year of daily records is not looking for the absence of problems. They are looking for the density of reporting, because an operation that records nothing is one that nobody can vouch for. Some categories, child labour among them, do not belong in a self-reported ledger at all, and a design that assumes otherwise has misunderstood what the ledger is for.

The second is that this does not yet exist. Voice-based advisory services exist and have been evaluated: an agricultural assistant deployed across four countries has an independent evaluation supporting its usefulness.⁴⁵ Voice input into a form field exists; one Indian state registers births by speaking the name into a phone.⁴⁶ A laboratory prototype routes spoken civic complaints with 92 per cent accuracy on a thousand-sample dataset.⁴⁷ But no deployed, independently evaluated system currently converts a spoken account into a licence application in a low-resource language. The components are real. The chain has not been assembled. This section describes what would have to be built, and the figures above suggest why it might be worth building.

Being heard

The three paths from Part V, with a model working after a human intake. The cost curve does not add the channel itself. It changes three things around it.

Language and literacy. Complaints in most mechanisms still travel by form, in writing, in a second language. Spoken intake in the language people actually use removes three of the six access barriers named in the official commentary to the UN Guiding Principles, namely language, literacy, and knowledge of how the mechanism works.⁴⁸

Pattern over case. When each case cost weeks to prepare, an organisation could bring one. When it costs days, it can bring the pattern. A single account of a recruitment fee is one worker’s word. Forty accounts naming the same agent in the same city are a different kind of object, and one that a regulator or a buyer’s contract clause is equipped to act on. Enforcement, in this field, has followed aggregation more reliably than it has followed severity.

Aggregation as protection. This is the point I have not seen argued elsewhere, and it may be the most consequential. With one recipient, a single account can be traced to the person who gave it. With several buyers receiving a pattern finding, it cannot. The Business & Human Rights Resource Centre has recorded more than 6,400 attacks on human rights defenders in a business context since 2015.⁴⁹ In that environment, the ability to deliver a finding without delivering a person is not an efficiency. It is what makes speaking survivable.

A finding from Part IV belongs here. When the Issara Institute examined who had used its channel, the quietest speakers were those who had dropped out of the pattern first: not because they had nothing to say, but because the channel, through phone ownership and language, had been shaped so that they could not enter it. A pattern built only from the accounts that arrived is a pattern of who could speak, not of what happened. Any system that claims to show what is occurring at a site has to be designed with the people who did not call in mind, and that is a harder problem than anything in the model.

The audit that hears the workers

Social audits fail in ways that are widely acknowledged within the field. The visit is announced. The operator selects the workers the auditor interviews. The interviews take place on site, during working hours, within earshot of management. One worker voice provider found serious issues at 35 per cent of sites that had passed an audit, once workers were asked about conditions away from the audit setting.⁵⁰

The Accord addressed this by making the channel a condition of the agreement. A signatory brand’s supplier factories are covered by the complaints mechanism whether the factory likes it or not, and the brand has signed a binding document saying so. The digital version applies the same principle at a different scale. The buyer writes into its sourcing standard that a channel must exist, that an organisation independent of the operator must hold it, and that the buyer receives pattern findings and owns nothing else. The audit becomes the occasion on which the channel is introduced, not the period during which it exists. Workers are told, in their own languages, that it is there, who holds it, and that it remains open after the auditors leave. Patterns flow into the audit report. Everything else stays with the holding organisation.

The weaker version should be described so that it can be recognised and declined. A QR code distributed by the auditor, open only during the audit week, feeding individual accounts into a report the operator paid for. Such a channel is a snapshot, announced in advance, held by someone with no established trust, and delivered to the accused party. It would be worse than no channel at all, because it would teach workers that the channel is a trap.

The warning from Part IV applies at the smallest scale. The code on the canteen wall has to belong to the organisation, not to the company.

What this section does not claim

The finding about phone ownership is the one to end on, because it disciplines everything above. Women from one migrant group were absent from a channel before anyone spoke, and the reason was a device. So none of this should be imagined as an application. It is a voice note, a phone call, a feature phone, a number that works on the cheapest handset in the market. Access is the bet, and it is a bet on the least connected person at the site rather than the median one.

Nothing in this section scales trust. The layer beneath the channels scales. That layer consists of the languages a model handles, the custody that survives a change of funder, and the classification that keeps a path-two case out of the operator’s hands. If that is built once, the organisations that already hold the trust can build the channel. The proposition is deliberately modest.

VII. What would have to be built

This part is closer to a recipe with trade-offs than to a specification. Each item is a decision someone will have to make and be prepared to defend.

Custody first, before anything is collected. Not the buyer, not the operator, not the model provider. The question is not where the server is located but who can seize it, who switches it off when funding ends, and what happens to the testimony if the holding organisation ceases to exist. Three questions, answered in writing, before the first account is taken. An organisation unable to answer them should not open a channel, because a channel with unclear custody is a list of people who complained.

Three requirements, as one field organisation described them. Independent servers, meaning infrastructure that the organisation or a data steward administers rather than a commercial platform whose terms of service and exposure to legal process sit outside their control. Written protocols, meaning rules for who sees what and when a case moves. And models with limits: a model used to structure testimony must not be usable to identify the people who provided it. That means declining the cross-referencing that would de-anonymise, declining to reconstruct identity from content, and doing so as a property of the deployment rather than a policy the operator can switch off. This is a product requirement, and it is addressed to whoever builds the model.⁵¹

Intake stays human. The evidence on automated questioning of people describing harm is not encouraging. The established interviewing standards, from the World Health Organization’s recommendations through the Istanbul Protocol to the Murad Code, exist for reasons no current model satisfies: safety before information, referral pathways agreed before anyone is asked anything, and the capacity to recognise when a conversation should stop.⁵² Peer-reviewed evaluations of conversational AI in crisis and gender-based violence contexts have found systems failing at precisely those points.⁵³ A model works after the account exists, by transcribing, translating, structuring and matching. A model that conducts interviews at scale is not the goal, and on the current evidence would not be an improvement.

Classification errs upward. Part V identified the failure that matters, a path-two case routed to the operator. The rule that follows is asymmetric. When uncertain, route to external custody and accept the delay. Do not route to the operator on a guess. In practice this means a classifier tuned to over-flag, a person reviewing everything flagged, and the operator never seeing what was routed away from it. Whether a model can identify indicators of forced labour in an account that does not name them, debt, confiscated documents, fees paid in another country, is a testable question, and to my knowledge nobody has published the test. The relevant evaluation is not overall accuracy but the relative cost of the two errors, weighted so that one path-two case routed inward counts for more than many path-one cases routed out.

Separate by cause, not by wording. Forty accounts that mention the same agent are one pattern. Forty accounts that use the same phrase because the same translator produced them are not. A system that clusters testimony has to distinguish shared cause from shared language, and the distinction is architectural: get it wrong and the pattern is an artefact of the pipeline.

Path three needs a before. Yanacocha’s thousand documents arrived seventeen years after the spill. A community able to show, from a sealed record, what its water had been like in 1999 would have had a different case. A baseline could be built from satellite archive, permit conditions and community testimony, collected while nothing is wrong and sealed until there is a dispute. A baseline has one property that nothing else in this essay has. It benefits, asymmetrically, whichever party is telling the truth.

No channel without a competent recipient. A channel that collects testimony where nobody can act exposes people and returns nothing. Before a channel opens, someone has to be able to name the regulator, the contract clause, the court or the buyer mechanism that a finding will reach, and to say how. Where nobody can be named, that gap is itself the finding, and the honest response is to publish the gap rather than open the channel.

Define failure in advance. What would have to be true in eighteen months for this to have worked, and what would signal that it should stop. Written down, and published afterwards. Much of what has been built in this field has no such statement, which is part of why so much of it is still running.

The trade-offs I would name rather than resolve. More questioning yields better usability and less authenticity, and the line between them is a judgement rather than a rule. Independence favours open models, which can run on a server the organisation controls; capability in the languages that matter favours commercial ones, which cannot. Whoever holds the data also loses it if they are raided or defunded, and the only way to hold nothing is to help nobody. None of these has a settled answer. All of them have to be tested in practice rather than resolved at a desk, and the honest position is to say which way one leans and why.

What the model provider owes, made more specific than in the first essay. Per-language accuracy, published, for speech, and measured end to end: not the word error rate on a benchmark but whether a filing produced from a spoken account in that language was accepted by the authority it was sent to. Detection of path-two indicators as a published evaluation. Zero data retention for this class of deployment, as a contractual property rather than a setting. A commitment on model deprecation, because a channel built on a model version that disappears within eighteen months is not a channel for an organisation with one technical volunteer. And eligibility that does not require registered charitable status, since the organisations that hold the trust are frequently not incorporated in any form a verification service recognises. None of this is philanthropy. It is what makes a company’s own path-one channel worth anything, because a path-one channel with no credible external counterpart is one that workers have good reason to distrust.

VIII. Who pays for a channel that belongs to nobody

Return to the plantations, and to the second observation from Part I.

Four retailers, one duty, and a process that depended on which of them chose to engage. The retailer that engaged carried the cost of doing so: the negotiations, the payments, the exposure that comes with being the party named in a resolution. Those that did not engage received the improvement regardless, because a plantation that has paid its workers has paid them for everyone who buys from it. That structure gives every buyer a reason to wait for another to move first. It does not dissolve with better intentions. It dissolves when the channel stops being a case and becomes infrastructure.

Four retailers who each build their own channel to the same plantation pay four times for the same testimony, and each of them ends up holding a list of people who complained. Four retailers who jointly fund an infrastructure that a local organisation holds pay a quarter each, receive the same pattern findings, own nothing, and none of them has gone first. Part II described why the competition authorities in Europe, Germany, Britain and the Netherlands would be unlikely to object to such an arrangement, provided it is designed as a standard rather than a pricing agreement.

Funded, not owned is not only the ethical answer to the paradox. It is the economic one.

Funding should go there, in this order. First, measured language coverage for the languages spoken at production sites, because what is not measured does not improve and an organisation deciding whether to rely on a model has nothing to decide with. Second, custody infrastructure that survives a change of funder, which is the unglamorous layer that nobody funds. Third, the maintenance that turns a pilot into a system. Fourth, and only fourth, the channels themselves, because those have to be built by the organisations that hold the trust, and money cannot create that.

Two things stand between this and the world, and the essay should not claim to have resolved either.

The first is an objection I have not answered. It came from someone who has spent years working at mine sites, and it goes like this. If knowing something creates a duty to act on it, then whoever benefits from not knowing has no reason to fund the finding out. On that reading, the thinness of ground-level evidence is not only a resource constraint. Some of it is a preference, and the field has spent years treating one as the other. I do not know how much of it is which. I can say something narrower. When documentation becomes cheap, the standard of what a company could reasonably have known shifts without any law changing, because the defence that the information was not reachable becomes harder to sustain. That shift takes time, and it cuts both ways, since the same capability documents the defence as readily as the claim. Whether it changes the incentive to fund the finding out is a question I would rather leave open than answer badly.

The second is Yanacocha. The IFC’s ombudsman is among the better-designed external mechanisms in existence, and Yanacocha shows what it delivers when it operates as intended: twenty years of dialogue tables, independent water studies, monitoring programmes, one early investigation that named real failings, and no subsequent finding that the lender’s standards were breached. Process is not nothing. The studies were real, and the community had not had them before. But process is not remedy, and a channel with a competent recipient and no leverage behind that recipient produces a well-documented conversation. This is the version of failure that most resembles success, and any design has to state which side of that line it is on.

One further thing belongs at the end, because it is the largest.

Every channel described in this essay is a sensor. Each one, held locally, produces a record of what is happening at one site. Nobody has yet drawn the map that a network of such sensors would eventually produce, because the sensors do not exist. But the map has been drawn once by hand. A university research team spent six months reading, cross-referencing and tracing supply chains until it could show where forced labour entered the automotive industry, and when the report was published in December 2022 it prompted sourcing reviews across the sector. It was subsequently corrected twice, to remove companies it had named in error, which is what a map drawn by hand looks like.⁵⁴ If we can see it, the lead author said at the time, so can you. That is what a network of channels would eventually make ordinary. It is also a different essay.

Which leaves the closing point, and it is the one the whole argument has been approaching.

The first buyer to fund a channel it does not control gives up the ability to manage what emerges from it. That is the real price. It is the reason this has not happened, and it is the only version that would work.

An invitation

This essay was written to be argued with, and the first one improved considerably because people who work in this field did exactly that. If you have run a grievance channel, or tried to, I would like to know whether an independently held one has ever survived a change of funder in your experience, because I have not seen it and would be glad to be corrected. If you have been on the buying side, I would like to know whether the objection about not knowing rings true, and whether anyone who holds a due diligence obligation has ever funded the evidence that would trigger it. If you have worked with testimony that arrived through translation, I would like to know where the line between shared cause and shared wording actually falls in practice, because the answer decides whether pattern findings mean anything. And if you have any idea what a year of a working channel costs at one site, all in, I would like to hear it, since nobody I have asked has been able to say.

I am reachable by email and on LinkedIn, and I answer.

Sources

1. Oxfam Germany, press release and case documentation, 2 and 3 November 2023; ECCHR case summary, “Edeka und Rewe verstoßen gegen Lieferkettengesetz”. Oxfam’s letters to the four retailers were sent in summer 2023 and concerned four cases on banana and pineapple plantations in Ecuador and Costa Rica.
2. Oxfam Germany, “Edeka und Rewe verstoßen gegen Lieferkettengesetz”, 2 November 2023; ECCHR case summary. Findings are as documented by ASTAC and SITRAP and published by Oxfam; they have not been independently verified for this essay.
3. Oxfam Germany and ECCHR public materials on the 2025 settlement. Terms confidential; no admission of liability.
4. Lidl Germany, “Existenzsichernde Löhne und Einkommen in den Erzeugerländern von Bananen”, company website; Lidl states it moved to living-wage bananas from May 2022.
5. Edeka statement as reported by Lebensmittel Praxis, 3 November 2023, and Logistik Heute, November 2023.
6. Rewe statement as reported by Logistik Heute, November 2023.
7. Oxfam Germany, 2 November 2023.
8. ECCHR press release, 3 November 2023; Oxfam press release, 3 November 2023. The complaint was filed on 2 November 2023.
9. Gesetz über die unternehmerischen Sorgfaltspflichten in Lieferketten (Lieferkettensorgfaltspflichtengesetz, LkSG), in force since 1 January 2023; BAFA is the supervisory authority under § 19 LkSG.
10. ECCHR press release, 3 November 2023.
11. ECCHR case summary: BAFA deemed the complaint well-founded and initiated an investigation in late January 2024.
12. ECCHR, on the recognition of ASTAC as a party to the proceeding. The duration is as characterised by ECCHR.
13. Oxfam Germany, 2 November 2023: the supplier Otisgraf was Rainforest Alliance certified until July 2023.
14. CORBANA statement, November 2023, as reported.
15. Bundeskartellamt, case B2-90/21, case report 8 March 2022; press release 18 January 2022. Working group members as listed in the case report and in the working group’s own publications: ALDI Nord, ALDI SÜD, dm-drogerie markt, Kaufland and REWE Group; Lidl participated in the working group’s early phase.
16. Lidl Germany, as above.
17. Oxfam Germany, “Gerechte Löhne auf der Bananenplantage?”, April 2026.
18. European Commission, Guidelines on the applicability of Article 101 TFEU to horizontal co-operation agreements, adopted 1 June 2023, chapter 9 on sustainability agreements.
19. Ibid., paragraph 549 and the six cumulative conditions of the soft safe harbour, as characterised in commentary by Slaughter and May, Bird & Bird, Herbert Smith Freehills and Allen & Overy (2023).
20. Ibid., section 9.4 on the application of Article 101(3).
21. Bundeskartellamt, Fallbericht B2-90/21, 8 March 2022, and press release of 18 January 2022. Working group members: ALDI Nord, ALDI SÜD, dm-drogerie markt, Kaufland and REWE Group. On the absence of exchange on purchase prices, costs, volumes and margins: the case report and commentary by the Initiative for Sustainable Agricultural Supply Chains (INA), 18 January 2022.
22. Bundeskartellamt press release, 25 January 2022, on the dairy initiative.
23. Competition and Markets Authority, Green Agreements Guidance, 12 October 2023; Autoriteit Consument & Markt, Policy Rule on sustainability agreements, October 2023.
24. Remarks by Martijn Snoep, chair of the ACM, and Olivier Guersent, Director-General for Competition, as reported in competition law commentary, 2023.
25. Harrison, J. and Wielga, M. (2021), “Assessing the effectiveness of non-state-based grievance mechanisms in providing access to remedy for rightsholders”, Business and Human Rights Journal; Harrison, J. and Wielga, M. (2023), Business and Human Rights Journal 8(1).
26. Harrison and Wielga (2023), p. 43. The 41 to 58 per cent figure refers to adjudicated cases at the three mechanisms named, not to all complaints received.
27. MSI Integrity (2020), Not Fit-for-Purpose; OECD Watch, Remedy Remains Rare and subsequent reviews; Owen, J. R. and Kemp, D., on operational-level remedy in the extractive sector.
28. International Accord for Health and Safety in the Textile and Garment Industry: renewal announced April 2026 for a six-year term with around 200 signatory brands (Sourcing Journal, 11 April 2026); Pakistan Accord in effect since 1 January 2023, more than 140 signatories by August 2024 (US Department of Labor, ILAB); expansion of the Bangladesh worker complaints mechanism beyond occupational safety from November 2025 (International Accord, programme updates).
29. Permanent Court of Arbitration, cases brought by IndustriALL Global Union and UNI Global Union against two signatory brands, admitted 2017, settled 2018; IndustriALL, “Bangladesh Accord’s binding arbitration process scores legal win”, October 2017. Brand names confidential by order of the tribunal.
30. IndustriALL and UNI Global Union public statements, 2025 to 2026, on binding agreements versus voluntary schemes, as reported by Sourcing Journal, April 2026.
31. Issara Institute, published programme figures. Contact and remediation counts are programme self-reporting and have not been independently audited.
32. Economic and Labour Relations Review (Cambridge University Press), “Digital worker feedback infrastructures”; Farbenblum, B., Berg, L. and Kintominas, A. (2018), Transformative Technology for Migrant Workers: Opportunities, Challenges, and Risks, Open Society Foundations.
33. Issara Institute, description of its worker voice channels.
34. Issara Institute, finding on phone ownership among migrant women workers, as reported.
35. Reporting by Verité, Transparentem and others on recruitment fees among migrant workers in Malaysian manufacturing, 2018 to 2020; ILO indicators of forced labour.
36. US Customs and Border Protection, withhold release order of 15 July 2020, finding of 29 March 2021, modification of 9 September 2021; US Department of Labor, Bureau of International Labor Affairs, “Example in Action: Top Glove WRO and Subsequent Modification”; Impactt Ltd, “Impactt supports Top Glove with modification of Customs and Border Protection Forced Labour Finding”; Malay Mail, 1 March 2022, reporting the Ministry of International Trade and Industry’s estimate.
37. Compliance Advisor Ombudsman, case Yanacocha-01/Cajamarca; CAO, “Investigation into the Mercury Spill of June 2, 2000 in the Vicinity of San Juan, Choropampa, and Magdalena, Peru”, July 2000.
38. CAO investigation report, July 2000, and case summary Yanacocha-02.
39. CAO case summaries Yanacocha-01 and Yanacocha-02; the Mesa de Diálogo y Consenso was convened in September 2001 and the CAO’s involvement closed in 2006.
40. CAO case summary Yanacocha-11/Cajamarca, complaint December 2017, appraisal and monitoring report December 2020; IFC divestment reported for December 2017.
41. Peru’s formalisation regime under Decreto Legislativo N° 1293 and the REINFO register; cost range for the IGAFOM environmental instrument from a Peruvian consultancy’s published guidance. Consultant estimate, not an official figure.
42. Hilson, G. et al. (2022), “Brick by brick, block by block: Building a sustainable formalization strategy for small-scale gold mining in Ghana”, Environmental Science & Policy.
43. Fairtrade International estimate as cited by ISEAL Alliance, “Navigating EUDR: compliance and beyond”, May 2024.
44. Cardona-Trujillo, Muñoz-Mora and Restrepo Areiza (2026), SSRN working paper on EUDR compliance costs in the Colombian coffee sector.
45. Digital Green, Farmer.Chat, with an independent evaluation by 60 Decibels, 2025.
46. National Informatics Centre, Pehchan civil registration system, Rajasthan, integrating Bhashini voice input.
47. Rajkumar et al. (2025), “A zero-shot LLM framework for multimodal grievance classification, urgency scoring, and abuse detection in civic feedback systems”, Scientific Reports 16:2299. Tested on a synthetic dataset; no field deployment reported.
48. UN Guiding Principles on Business and Human Rights, commentary to Principle 31.
49. Business & Human Rights Resource Centre, database of attacks on human rights defenders, 2015 to 2025.
50. Labor Solutions, published analysis of worker voice findings at previously audited sites.
51. Requirements as described by the Alliance for Responsible Mining in conversation with the author, September 2026.
52. World Health Organization (2001, 2007); OHCHR, Istanbul Protocol, second edition (2022); Institute for International Criminal Investigations, Murad Code (2022).
53. Heston, T. F. et al. (2025), Scientific Reports, on mental health chatbot performance in crisis detection; Izaguirre-Choperena et al. (2025), Frontiers in Political Science, on AI chatbots in gender-based violence support.
54. Murphy, L., Salcito, K., Uluyol, Y. and Rabkin, M. (2022), Driving Force: Automotive Supply Chains and Forced Labor in the Uyghur Region, Sheffield Hallam University, Helena Kennedy Centre for International Justice; amended October 2024 and July 2025. The characterisation of the industry response is the author’s own observation and is not independently documented.

This essay is personal work. It reflects my own analysis and is not connected to my employer or to any organisation I work with. The companies named in Part I appear because the case is a matter of public record. Nothing here draws on any source other than published material, the positions of all parties are reported as they stated them, and no view is offered on any company beyond what those sources support.

Written with AI-assisted research and drafting; the argument and its errors are the author’s.