Valenta Told Kingston to Optimise Before You Automate. I've Watched Caribbean Companies Skip That Step for Years

Photo via Unsplash

A Toronto-headquartered automation firm walked into a room of Kingston executives this month and told them, in effect, not to buy what it was selling yet. Valenta co-founder and CEO Jayesh Kasim, in the country for sessions that drew representatives from GraceKennedy Limited's Foods, Financial Group, and IT Operations divisions, put it plainly: "For us, we solve business problems using technology. Our methodology is to always optimise before we automate." That is an unusual thing for a vendor to lead with, and it is the correct thing, which is why it is worth taking seriously rather than filing under standard consulting caution.

The numbers behind the advice are specific enough to check. Valenta's Caribbean franchise, run through Aurora Technologies Limited under founder Erica Anderson, reports that finance processes eating up around eight hours a day typically fall to one or two once a company fixes the underlying workflow and only then layers automation on top. Roughly 80 percent of the firm's initial engagements in the region start in finance: accounts receivable, accounts payable, payroll, the repetitive, rules-based work Kasim recommends automating first while keeping actual decision-making under human review. Less complex projects launch within three to five weeks. Every digital assistant goes through rigorous testing before production, according to Valenta's managing partner for Jamaica and the Caribbean, Roger Grant, who says the firm works closely with subject matter experts rather than treating a workflow as a black box to be automated from the outside.

The Order Nobody Wants to Sell You

My training is in physics, and the habit that discipline leaves you with is asking what a system actually does to whatever you feed into it, not what you hope it does. An AI model does not know the difference between a well-run finance process and a badly run one. It learns the pattern in front of it and executes that pattern faster, at scale, with fewer humans able to catch the moment it goes wrong. Feed it eight hours of manual reconciliation built around a workaround nobody has questioned in five years, and you do not get one hour of clean output. You get eight hours of the same workaround, compressed and harder to unwind because now it is embedded in a system instead of a habit. Kasim's line, optimise before you automate, is not caution. It is a description of how the technology actually behaves.

I have watched this sequencing problem play out at board level for years through StarApple AI's own training work across the region. It shows up as governance: in the study we published on our board-level AI programme, organisations that completed the training cut the time to stand up AI and data governance from an 11 to 15 month build down to 6 months, and the number one reason it had dragged that long in the first place was that boards approved AI initiatives before anyone had mapped what governance the initiative actually required. Same failure, different layer. A vendor selling the tool before the process is documented, and a board approving the initiative before the governance is designed, are the same mistake wearing different clothes.

Photo via Unsplash

What Optimise Before You Automate Actually Looks Like

Strip the marketing language away and Valenta's approach is fairly unglamorous, which is exactly why it works. Start with high-volume, rules-based activity where the correct answer is not in dispute. Test the digital assistant rigorously before it touches production, with the subject matter experts in the room, not just the engineers. Keep the decision-making steps, the calls that require judgment rather than pattern matching, under a human's authority. Aurora Technologies' Erica Anderson framed the payoff in terms that had nothing to do with efficiency metrics: "You're giving them time to go home. You're giving them time to be with their kids, and reducing the stress and panic at the end of each reporting cycle." Valenta also holds ISO 9001 and ISO 27001 certification and states plainly that it does not use client data to train its models, a compliance detail that matters more in a region where data protection legislation is still maturing than it might in a market with decades of precedent behind it.

None of that is exotic. It is what a competent systems integrator has always done, AI or no AI. What is notable is that it is being said out loud, in Kingston, by the vendor, in a market where the more common pitch is the opposite: buy the platform, and the process problems will sort themselves out once the data starts flowing. They do not sort themselves out. They get faster.

The Data Backing the Discipline

The same week Valenta was in Kingston, PwC published its 2026 AI Jobs Barometer, and the Jamaica Gleaner carried the local read on July 8. The headline numbers are global, but they describe exactly the gap between tool-buying and capability-building that Kasim's pitch is aimed at. Jobs requiring AI skills are growing nearly eight times faster than the overall job market. Workers with AI expertise command a 62 percent wage premium. Organisations most exposed to AI that are actually leading in its use recorded 163 percent labour productivity growth, not the modest single-digit gains a chatbot bolted onto an unfixed process tends to produce. Entry-level roles in AI-exposed occupations have grown 35 percent since 2019, which cuts against the standard fear that AI simply deletes junior jobs.

PwC Jamaica's Hugh Thompson put the local framing on it well: AI is increasing the importance of uniquely human capabilities, and the workers who benefit are combining AI literacy with critical thinking, communication, and leadership, not just prompt fluency. His colleague Jossett Francis Wint made the same point from a different angle: the bigger story is how jobs are changing, not whether they disappear. Read against Valenta's process-first pitch, the throughline is consistent. The 62 percent wage premium is not paid to people who learned to click a new interface. It is paid to people who understand the process well enough to know what the tool should be doing and to catch it when it is not.

Jamaica's Own Government Is Learning the Same Lesson

The sequencing problem is not just a boardroom habit. It shows up at the national level too, and Jamaica has the paperwork to prove it. On April 1, 2026, the government launched its UNESCO AI Readiness Assessment Methodology report at Jamaica House, drawing on close to 200 stakeholders, researchers, entrepreneurs, officials, youth and community representatives, with funding support from the European Union. The assessment scores a country against five dimensions: legal and regulatory, technological and infrastructural, economic, social and cultural, and scientific and educational. It found real foundation to build on, the Data Protection Act, the Cybercrimes Act, a National AI Task Force established in 2023, National AI Policy recommendations delivered in 2024, alongside real gaps. Dr. Andrew Wheatley, who holds ministerial responsibility for Science, Technology and Special Projects in the Office of the Prime Minister, named the gap directly: the country needs a National AI Oversight and Implementation Council, AI education stretching from early childhood through tertiary and vocational levels, and a closed gender gap in STEM. He has since directed the National AI Task Force, of which I am a member, to prioritise AI literacy across government, on the record stating that "Jamaica does not need AI hype; we need AI governance."

That is the same instruction Valenta is giving Kingston boardrooms, issued at the scale of a country instead of a company. Build the institutional infrastructure and the literacy first, or the AI capability the government wants to deploy will sit on top of the same gaps the assessment just documented in writing.

Photo via Unsplash

Even the Frontier Is Optimising for Fit, Not Just Power

What strikes me most is that the same discipline is now visible at the very top of the industry, not just in how Caribbean businesses should buy AI. OpenAI released its GPT-5.6 lineup publicly on July 9, 2026, and it did not ship as one bigger flagship model. It shipped as three priced tiers: Sol at $5 input and $30 output per million tokens for frontier reasoning, Terra at $2.50 and $15 for GPT-5.5-level quality at roughly half the cost, and Luna at $1 and $6 for fast, high-volume work. A frontier lab with effectively unlimited engineering budget is now selling fit-for-purpose capability instead of one maximum-power product for every job. That is the exact judgment Valenta wants a Kingston finance team to apply before it signs a contract: match the tool to the actual task, not the biggest number on the spec sheet.

The same discipline shows up in how the industry is handling interoperability. In December 2025 the Linux Foundation formed the Agentic AI Foundation, with Anthropic donating the Model Context Protocol as a founding contribution alongside Block's goose framework and OpenAI's AGENTS.md, backed by platinum members including AWS, Google, Microsoft, Bloomberg, and Cloudflare, with Snowflake at the Gold tier. The protocol had already passed 97 million monthly SDK downloads by the time of the announcement. A shared, open standard for connecting AI tools to business systems is genuinely good news for a small market like ours, because it means a Caribbean company no longer has to fund a bespoke, expensive integration for every vendor it works with. But that advantage only pays off for an organisation that has already done the unglamorous work of knowing which systems and which processes the integration needs to serve. The standard removes a cost. It does not remove the mapping.

What This Means for the Board Table

None of this requires a research department. It requires discipline a board can actually apply the next time a vendor walks in with a deck. Map the manual process end to end and time it honestly before the request for proposal goes out, not after. Ask every vendor for a pilot measured against that documented baseline, not a demo measured against nothing. Treat data governance as a line item that comes before the contract is signed, not a compliance task bolted on afterward. Keep the judgment calls, not just the repetitive tasks, under human ownership, the way Valenta structures its own deployments. Budget in weeks for the simplest automations and months for anything that touches governance, and expect your own board to be able to question a vendor invoice clause by clause, the way I am told one GraceKennedy director did in a live session this month, pricing each line against what an internal build would cost.

"Mi tell client dem the same thing every time: no bother buy the shiny AI tool if the process underneath it a mess. The AI don't fix the mess. It just run the mess faster, and now it cost more to unwind." - Adrian Dunkley, AI Boss

AI does not fix a broken process. It automates whatever is already there, at whatever speed the infrastructure allows, faster than a human ever could catch the mistake compounding. That is true whether the process sits inside a finance department in New Kingston or inside a national AI strategy in Jamaica House, and it was true before generative AI made the automation layer cheap enough for almost any company to buy. What has changed is how cheap the mistake now is to make, and how expensive it has become to unmake once it is running at machine speed. Valenta got the sequence right this month. The businesses that listen will be the ones whose AI numbers actually look like the ones PwC is measuring, not the ones explaining eight months from now why the tool made the problem worse, faster.