Your Weight in AI Tokens
Before comparing the bills
Caribbean leaders need to compare the cost and quality of completed work before using AI to redesign jobs.
A salary and an AI subscription appear to be comparable expenses. Put them in neighbouring cells, annualise the subscription, and the difference can look like a saving already waiting to be claimed.
Before treating that difference as a saving, establish whether the proposed system can deliver the employee’s responsibilities to the required standard. Test how work reaches completion, including the exceptions that make a routine process difficult.
New research on the work around AI
A manager deciding whether to cut a role needs to know what happens to the work after that person leaves. A preprint posted on 22 October 2026 makes some of that work visible. Researchers at Microsoft Research Africa interviewed 27 workers in 21 small and medium-sized businesses in Kenya and Nigeria. Participants described checking evidence and correcting outputs that did not fit local requirements before sending AI drafts to clients. The interviews took place in July–October 2023. This is newly published qualitative research, not a test of current models, a causal estimate of productivity or a representative Caribbean sample. Abdulhamid and colleagues, Making Generative AI Workable at Work
Map responsibility before changing capacity
Consider a hypothetical invoice-dispute workflow. An employee retrieves correspondence, checks delivery records, reconciles a mismatch, secures approval for a correction and updates the account. Some steps are suitable for AI assistance. Others depend on access rights, reliable records or a decision the organisation has reserved for an authorised person.
The employee may also compensate for poor system design. They know that a customer uses a different trading name, that a second delivery was recorded elsewhere, or that an apparent duplicate should not be cancelled. If those dependencies are absent from the process map, a staffing reduction can remove a control that management never formally recognised.
Documenting the knowledge may make some work easier to automate. It also reveals where the better investment is a system integration, a clearer policy or a conventional software rule rather than a language model.
The method I apply here is People First, Problem Second, AI Last: establish the human need, define the problem and then choose the technology. In the invoice example, the customer needs a correct account and the clerk needs reliable evidence. The operational problem is a broken handover. A model may help retrieve or explain records, but repairing that handover remains part of the project.
Before changing staffing, map the workflow’s trigger, required evidence, system access, decisions, handovers and accepted endpoint. Ask the people doing the work where the documented procedure differs from reality. Include the colleague who receives a failed handover, not just the person whose task appears to become faster.
Map evidence and responsibility alongside the task. Missing records and approvals remain part of the operating model. Conceptual illustration, not a measured result.
Distinguish capacity claims from employment evidence
Klarna’s disclosures show why precision matters. Its IPO prospectus connected a fall in full-time employment from 5,527 at the end of 2022 to 3,098 in June 2025 with an efficiency strategy including AI. Its separate estimate of work equivalent to more than 700 agents was derived from conversation volumes. These figures measure different things. Klarna prospectus, pages 152 and 247
For 2025, Klarna subsequently reported AI handling 80% of service chats and continued access to human representatives. This describes a combined service model. The company-reported performance does not establish wholesale replacement followed by reversal. Klarna 2025 annual report, page 184
The same report attributes approximately US$59 million in 2025 customer-service savings to the AI assistant. This is the company’s attribution, not an independently measured causal effect. Those reported gains belong in the discussion alongside the human support it retains. Klarna 2025 annual report, page 184
Neither a reported work-equivalent nor another company’s headcount change validates a staffing decision in your organisation. They can inform the questions you ask, but they cannot supply the missing evidence about your service obligations, demand or dependencies.
Compare operating models on equal terms
An AI subscription provides access under a plan; an application programming interface, or API, lets software call a model and can be billed by tokens and other services. Tokens are units of information processed by the model, often words or word fragments for text. They are not a standard unit of completed business work. OpenAI, for example, maintains separate billing for ChatGPT subscriptions and API usage. OpenAI billing documentation
A defensible comparison needs the current operating cost and the proposed operating cost over the same period and workload. Both should include people, software and correction effort. For the proposed model, budget for implementation, integration, data preparation, training, evaluation, security, ongoing oversight and exception handling. State how one-off costs are spread over the evaluation period and who carries them.
Account for risk as well as expenditure. Some errors create a modest rework bill; others expose customer information or permit an action the business has no authority to take. Rules protecting customers and restricting authority should remain pass-or-fail requirements. They should not disappear inside an attractive average return.
Compare accepted outcomes, full costs, failures and time on equal terms. This is an evaluation template, not a forecast of savings.
Apply the test to a Jamaican distributor
For a hypothetical distributor in Kingston, add a supplier invoice in US dollars and a customer account in Jamaican dollars to the dispute example. The clerk must identify the currency of each amount, the exchange rate used for the transaction and the date attached to it. An AI response that treats both dollar signs as the same currency can sound reasonable while leaving the account wrong.
Keep the original currency, conversion record and authorisation beside the proposed correction. Use the business’s documented transaction rate, not a model’s recalled rate or an unlabelled figure from a web search. The Bank of Jamaica’s counter-rate tables identify dated rates for transactions at its own counters; they do not establish the rate this distributor actually paid. A reviewer needs the business record.
This is an application to Jamaican wholesale distribution, not evidence that the African interview findings predict Jamaican performance. That transfer is the weakest part of the argument. The proposed local trial must establish whether the workflow actually improves. Other Caribbean markets need their own tests: the JMD example cannot simply be relabelled for a different currency, contract or approval process.
Risks the trial must contain
In the distributor example, a mistaken currency conversion can leave a customer overcharged or a supplier underpaid. The customer, clerk and business then carry the correction burden. Keep payment execution outside the trial, use approved or de-identified records and require an authorised person to check the source amounts before any correction reaches the live ledger. Record disagreements and unresolved questions before approval. Preserve the previous process so staff can recover if a tool is unavailable. These are proposed controls for a hypothetical failure mode, not a claim that a named business suffered it. They also protect the potential benefit: quicker preparation is useful only if the account is still correct.
Make the pilot answer a staffing question
A pilot that measures output generation alone cannot establish whether staffing can change. Measure completed cases, acceptance quality and the human work required throughout the process. Use representative inputs, including unusual cases, and keep a record of the test conditions. Where practical, compare similar work with and without assistance rather than attributing every change after launch to AI.
Set thresholds before the results arrive. Define what counts as completion, which errors are unacceptable, how exceptions are handled and when the system must fall back to a person. Review the distribution of results: good average performance can conceal an unmanageable queue of difficult cases.
The review function needs its own capacity plan. If AI increases the number of drafts but the same specialist must check each one, the review queue can become the next bottleneck. Measure that person’s time and the downstream workload. Avoid counting a saving in one team while moving an equal or larger burden to another.
Invest the gain in human capital
Evidence for augmentation is more specific than a general promise that AI will make everyone faster. In a published study of 5,172 support agents, AI assistance increased issues resolved per hour by 15% on average. Less experienced workers benefited more; the most experienced workers showed small speed gains and small quality declines. The authors caution against generalising from one firm, occupation and tool. Brynjolfsson, Li and Raymond, 2025
Evaluate how the tool changes your team’s capabilities. Test whether a newer colleague can handle a wider range of cases under appropriate supervision, and whether experienced staff can spend more time on difficult decisions. Knowledge should become easier to retrieve while remaining open to challenge.
Treat recovered time as capacity until there is evidence of how it will be used. It becomes a financial saving only when a cost is genuinely avoided or removed; additional output creates value only if there is demand for it. Do not count the same hours as both a payroll saving and capacity for new work.
Three actions for the operations manager
- Map the invoice-dispute process with the accounts clerk and warehouse contact by 30 October 2026, as a proposed starting schedule. List the evidence and approval needed at each handover. Record the current completion time, correction effort and unresolved cases before introducing assistance.
- Test 20 de-identified historical cases by 4 November 2026, including five with missing or conflicting records. Compare the existing process with an AI-assisted version using matched case difficulty and different reviewers where possible to reduce memory effects. Count accepted resolutions, total staff minutes and every material error. This proposed diagnostic trial can uncover problems. It cannot certify reliability or authorise live payments; production decisions need a larger, representative evaluation matched to the risks.
- Review the results with the team before changing staffing. Stop if the system makes an unauthorised correction or exposes restricted information. Expand only after the business has met its agreed quality threshold and can staff the exception queue. If 20 suitable cases are unavailable, extend collection. Missing evidence cannot justify expansion.
Use the AI Subscription Value Calculator for the subscription component. It takes the fee, measured net hours saved and a chosen hourly value in one currency. As a hypothetical illustration, JMD 3,000 per month, four net hours and JMD 2,000 per hour return JMD 8,000 of time value and JMD 5,000 after the fee. These are assumptions, not vendor prices. The tool excludes other operating costs and risks, and it does not convert currencies. Add the missing costs before making an investment decision.
A staffing proposal should then explain which responsibilities have moved, who owns the exceptions and how knowledge will transfer. If those answers are missing, the trial has identified more work to do before a role can responsibly change.
Questions to settle before procurement
Does occupational exposure mean a job will disappear?
The ILO’s 2025 assessment put about 25% of global employment in occupations with some generative-AI exposure. This measures potential task exposure, not a forecast that those workers will lose their jobs. It is a global estimate, not a Caribbean employment projection. Your workflow still needs to be tested in its actual setting. ILO Working Paper 140
What should a vendor demonstrate beyond a good answer?
Ask for a test using a de-identified case that contains conflicting records. For a Caribbean rollout, include a local currency or policy ambiguity and require the vendor to use the applicable source record. The system should show which source supports each proposed action and stop when the evidence is insufficient. Record that behaviour in your acceptance criteria before you sign.
How should we record the time staff spend reviewing AI?
Give each case a log with preparation, generation wait, verification, correction and escalation time recorded separately. Include the reviewer and receiving team. Keep elapsed customer waiting time separate from staff effort, since both matter but describe different costs.
What if a model changes after the pilot?
Keep the approved test cases and expected outcomes as a repeatable check. Record the model, settings and connected data sources used in the pilot. Rerun the checks after a material change, and retain a way to return to the previous process.
What should happen to knowledge the pilot captures?
Assign an owner to each procedure or exception note and give it a review date. Ask another employee to follow it without the original author present. That checks whether the organisation has transferred usable knowledge rather than simply collected more documents.
More research
- ILO: Generative AI and jobs, 2025 update, for a task-level approach to occupational exposure.
- Generative AI at Work, for measured productivity, heterogeneous effects and limits to generalisation.
Related sites: AI Boss HQ, StarApple AI and StarApple Analytics.
Analysis. The invoice workflow and diagrams are illustrative. Sources checked on 28 October 2026.