What actually blocks the sale in emerging markets.

Most companies that struggle to sell in an emerging market think they have a demand problem. Usually they have a trust problem, a pricing problem, or a signature problem. Each one has a different fix. Telling them apart is most of the work.

The most expensive assumption a well funded company carries into Nairobi, Lagos, Manila or Jakarta is that a good product demonstration moves a deal forward. Every emerging market has a long memory of failed technology projects. In that context a demo reads as marketing material. Your buyer has seen a beautiful demo before. They also saw the implementation that followed it and the vendor who stopped answering the phone eighteen months later.

What moves the deal is a reference the buyer already trusts, describing an outcome the buyer already wants, at an institution the buyer treats as comparable to their own. That takes a long time to manufacture. It also means your first five customers are sales infrastructure. Pick them for how well they will vouch for you.

In most emerging market institutions, the person who wants your product, the person who can approve the spend, and the person who can kill the project are three different people. Only the first one takes your meetings. Enthusiasm from the head of operations tells you something about the operations department and almost nothing about the deal.

Ask in the second meeting who else has to be comfortable. Ask what has to be true for this to reach the committee, which committee, meeting on what cycle, requiring what paperwork. Ask what killed the last vendor at this stage. People answer these questions freely because nobody else asks them. The answers usually delete half your forecast, which is the point of asking.

Your incumbent competitor is rarely another product. It is a spreadsheet, a clerk, and an arrangement that has worked well enough for fifteen years. It sits on no budget line. Price against a competitor that does not exist and you produce a number that works in your board deck and means nothing in the buyer's finance meeting.

Companies that price well in these markets do three things. They make the cost of the current arrangement explicit: the hours, the errors, the losses nobody has added up. They set an entry price under the approval threshold that avoids a board, and they design the expansion path from day one. And they hold price when the objection is really about trust. Discounting against a trust objection tells the buyer you invented the first number.

Pilots are how conservative institutions say yes without saying yes. They work when you treat them as commercial instruments. An unpaid pilot with no success criteria, no named executive sponsor and no agreed conversion path will absorb a year of a small company's capacity.

A pilot worth running is paid, even nominally. It is scoped to one number the sponsor has personally agreed matters. It has a date. The conversion terms are agreed before it starts. If the institution will not agree conversion terms up front, you have learned something about whether they intend to buy, early enough for it to be useful.

This inverts the sequence most venture backed companies follow. The instinct is to build the product, hire salespeople, then look for the pattern in the deals they close. In markets that buy on trust, the founder has to find the pattern first, in person, across a small number of deals worked slowly. Then you hand it to a team that can repeat it.

Companies that skip this rarely fail loudly. They spend eighteen months building a pipeline that never converts, decide the market is not ready, and leave. Often the market was ready and the motion was missing.

Monetising what you have already built.

Most organisations that ask me for a growth strategy have plenty of assets. What they lack is prices. Somewhere in the building sits a platform that took four years to build, a dataset nobody outside can reach, a training programme that produces the only qualified practitioners in the category, or a network a competitor could not assemble in a decade. None of it appears in the revenue line.

List everything the organisation owns that somebody outside would pay for. Ignore whether it is currently sold, packaged, or thought of as a product at all. Platforms and sub-platforms. Proprietary data. Accreditation and certification. Methodologies. Convening power. Distribution into an audience other people want to reach. The list runs longer than anyone expects, and the reaction in the room is always the same. We have always given that away.

The reason nobody took this inventory is structural. Assets built to serve a mission, a community or an internal process end up with custodians, and a custodian is measured on continuity. Nobody in the organisation is wrong. Nobody is looking at price either.

Propose charging for something that has always been free and three objections arrive within the hour. It will damage the relationship with the people who use it. A grant or a mandate required it to be free. It is not good enough to charge for.

The first is usually wrong and cheap to test. The second is a real constraint and a much narrower one than institutional memory suggests, so read the actual agreement. The third is often true and is the only one that should slow you down, because it is a product problem with a product fix. Separating the real constraint from the inherited habit is most of the value an outside operator adds. Everyone inside has lived with the habit for years.

The temptation is to price the largest asset. Price the one closest to a signature. I look for three things. Somebody outside has asked for it, unprompted, more than once. It can be delivered without new engineering. There is a buyer whose approval threshold it fits under. All three means you have a first commercial line. Two means a second quarter project. One means a slide, and it should be resourced like a slide.

The unprompted request is the strongest of the three signals and the most consistently ignored. Organisations sitting on valuable assets are already receiving inbound requests for them. The requests get routed to whoever picks up, handled as a favour, and recorded nowhere. That is free market research sitting in somebody's inbox.

A first commercial line needs four things. A named buyer. A price with a stated basis. A delivery mechanism the current team can execute without heroics. One person whose job includes selling it. The common failure is the fourth. The line gets designed, approved, published, and assigned to nobody in particular, which works out the same as assigning it to nobody.

The second failure is over-design. Organisations that spent years building sophisticated products build sophisticated commercial models to match. Tiering, usage bands, partner margins, enterprise variants. All of it is legitimate and none of it should exist before your first ten customers, because you cannot calibrate any of it until you have watched ten people decide whether to buy.

Ninety days is enough time to take the inventory, test the objections against reality, pick one asset, price it, and get a paying customer live. It is short enough that the work cannot quietly turn into a strategy exercise. Monetisation projects fail by dissolving into analysis far more often than by picking the wrong asset. A live pilot with an invoice attached settles arguments that another quarter of internal debate will not.

What comes out of ninety days is rarely the biggest revenue line the organisation will eventually build. It is proof inside the institution that the thing they built can be sold, and a team that has now watched it happen once.

The translation layer: deploying AI in regulated emerging markets.

The technical problem of deploying AI is largely solved. The harder problem is translation. How do you move a model built in San Francisco into a regulatory context in Nairobi, Accra or Manila and keep the supervisor, the institution and the end user with you?

Every AI deployment I have worked on has the same structural problem. The technical team understands the model. The compliance team understands the regulation. Almost nobody understands both, and almost nobody has navigated both at once under operational pressure.

This is a training problem. Model builders think about performance, latency and accuracy at scale. People who handle regulators think about precedent, ambiguity and relationships. Organisations that assume one team can hand off to the other find out otherwise during an incident or an examination.

There is a version of compliance that is procedural. File the forms, write the policy, complete the audit. In stable high income regimes that is often enough. The regulator has seen the playbook, has capacity to evaluate it, and the relationship is established.

In emerging markets that breaks down. The Central Bank of Kenya, Bangko Sentral ng Pilipinas and the Financial Sector Conduct Authority in South Africa are sophisticated institutions operating where the norms for AI in financial services are genuinely new. They are writing the playbook in real time while supervising institutions deploying technology with no local precedent.

So regulatory work in these markets looks closer to policy co-creation than to compliance management. You are often helping the supervisor decide what standard to set and why. That work is relationship dependent, light on documentation, and much slower than most deployment timelines assume.

Technical to institutional. Model cards, evaluation frameworks and uncertainty quantification are the language of the machine learning team. They do not map onto concentration risk, consumer protection or fit and proper assessments. The people who speak both are the most under-resourced part of almost every deployment I have seen.

Headquarters to market. Global AI policy written in London or San Francisco rarely accounts for the regulatory texture of the markets where it lands. The EU AI Act is a different instrument from the Kenya Data Protection Act. Organisations that apply one global framework everywhere under-comply where expectations are highest and over-comply where the overhead costs real money.

Model to institution. A technically sound system can still fail at the last step, when a human operator inside a regulated institution reads the output and acts on it. A credit model that backtests well will underperform if loan officers do not understand its confidence intervals, or if the incentive structure rewards ignoring it. Deployment finishes when the operating culture has absorbed the outputs.

Institutions that navigate this well invest in the translation layer early. They hire people fluent in both languages, or they build processes that force the two disciplines into the same room every week. They treat regulatory engagement as an input to product design and start talking to supervisors months before deployment. And they build for the maintenance cycle, because AI systems need ongoing governance, monitoring and recalibration in ways older financial technology does not.

None of this is new in principle. What is new is how fast deployment timelines have compressed, and how hard that compression pushes teams to skip the steps that decide whether a deployment survives in a regulated environment.

Governance as a feature: what supervisors actually look for.

Most organisations treat AI governance as an audit event. You prepare for it, demonstrate compliance, and go back to work. The organisations that build durable trust with their regulators build governance into the product, where it is visible in the system's behaviour and maintained continuously.

The incentives are misaligned. The team shipping the model is measured on velocity, accuracy and adoption. The team responsible for governance is measured on the absence of incidents, which is invisible when things go well and catastrophic when they do not. Governance therefore gets resourced after something breaks and defunded once the pressure clears.

That pattern is consistent enough across industries and geographies to be the default. In my experience it is also the single best predictor of which organisations hit serious regulatory trouble within three to five years of going to production.

I have built products that supervisors examined and advised on the frameworks supervisors evaluate. The questions that decide an examination are remarkably consistent across jurisdictions.

The first is what your model does when it is wrong. Supervisors have watched enough technology fail to know that all systems fail. Well governed AI is distinguished by having mechanisms for detecting, containing and learning from failure when it happens.

The second is who is responsible when the model makes a consequential error. It sounds like a legal question and it is also a governance question. Organisations that cannot answer it cleanly have usually built decision systems without defining the human accountabilities around them. In regulatory terms that ambiguity is itself a failure.

The third is how you know the model still performs as intended. Supervisors are increasingly fluent on drift and distributional shift. Monitoring infrastructure, plus evidence that the organisation actually uses it, is the clearest signal available about maturity.

Organisations that treat governance as a product feature define failure modes before success metrics. They build monitoring and explainability into the architecture. They assign named human accountability for every consequential decision the model makes. And they document why the system was designed the way it was, which turns out to be valuable when an examination asks you to explain decisions made by a predecessor team three years ago.

This costs time, engineering and launch velocity at the outset. It usually pays that back within eighteen months of production, in avoided incidents, shorter examination cycles, and a supervisory relationship that speeds up your next deployment.

Good governance in a regulated environment looks like operating practice. A team that can answer questions about subgroup performance without preparing for them. A monitoring dashboard reviewed in the weekly operating meeting. A short chain from any model output to a named person accountable for the decision it influenced.

Mostly it looks like an organisation that is genuinely curious about how its model might be wrong. Supervisors are very good at spotting the difference between that and an organisation defending a claim.

The quiet consumer, and why incumbents systematically miss them.

Every major market research framework undercounts one consumer. They have access to products and the income to buy them. What they lack is legibility. The products built for them are consistently worse than the products built for consumers who are easier to observe and easier to survey.

The quiet consumer is defined by what they do not do. They engage very little with social media. They do not respond to promotions in ways attribution models can see. They are not early adopters and they are not loudly loyal. They do not post reviews, use referral links or attend brand events. They buy something once, use it for years, and expect it to work. When it stops working they leave without complaining, and churn models never register the departure because it generates no signal.

They are invisible to the instruments most organisations use to understand their customers. So they get under-resourced in product cycles, under-targeted in acquisition, and deprioritised in favour of consumers whose engagement is measurable and whose preferences fit the feedback loops that drive iteration.

A/B testing needs users who behave consistently enough that a product change produces a measurable change in behaviour. The quiet consumer engages too rarely to produce reliable results. Net Promoter Score needs people who answer surveys. They do not. Lifetime value models need transaction history dense enough to project from. Theirs is sparse because they buy infrequently and expect what they buy to last.

The result compounds. Every iteration is calibrated against the consumers who generate the most signal. Quiet preferences get averaged away or never enter the sample at all. Over time products converge on the tastes of engaged users and drift away from everyone else.

Their preferences are easy to discover once you commit to looking. Fewer, better things. Products that work reliably over products with broad feature sets. Willingness to pay more for quality they expect to last. Deep scepticism about novelty as a value proposition. Discretion in design, in marketing, and in how their data is handled. They do not want aggressive personalisation. They want to be left alone with something that works.

That maps badly onto frameworks optimised for engagement, retention and virality. Building for them requires inversion. Design for easy exit, for durability, for restraint in feature surface. It requires believing that a customer who stays quietly for ten years outweighs one who engages loudly for two.

Quiet consumers are not a niche. In most categories they are a large plurality of users and often the majority by volume. They look like a niche because their silence makes them invisible to the systems that size markets. Survey them and they do not answer. Analyse social media and they are absent. Look at your power user data and they are not there.

Look at long term retention and at the revenue cohorts that carry the business over multi-year periods, and they are everywhere, carrying more of it than the organisation's own picture of its customer base suggests.

Organisations that commit to the design discipline and the measurement investment required to serve them consistently find the market is larger, more loyal and more profitable than they expected. The difficulty is building conviction inside a company about a customer it cannot easily see.

Slow architecture: building for return, repair, and gathering.

The buildings that last are rarely the most technically innovative of their moment. They were designed for return: the assumption, built into the brief and the material choices, that people would want to come back, and that the building would be glad to receive them.

Modern construction is very good at delivering buildings that perform at the moment of occupation. The specification targets energy efficiency, the certification scores wellness, the photography is beautiful on handover day. The specification rarely asks whether anyone will want to be there in twenty years, whether the building will acquire the patina that makes a place feel lived in and trustworthy, or whether it can be repaired and reconfigured without demolishing the original intention.

Those questions are about time. Buildings that last get designed by people who think across generations, choose materials for how they behave under decades of use, and understand that the brief for a lasting building differs in kind from the brief for a building that has to photograph well and sell quickly.

There is a literacy that comes from studying buildings that are genuinely old and still in use, adapted continuously over centuries. Medieval guildhalls that became warehouses and then apartments. Farmhouses that absorbed an addition from every generation without losing the coherence of the original structure. Harbours whose stone quays carry the marks of rope, keel and salt water in ways no specification anticipated and that make the place unmistakably itself.

What they share is material honesty. They were built of things that age predictably: stone, timber, lime plaster, brick. And almost without exception they were designed to accommodate gathering. The hall, the courtyard, the common room, the threshold that works as a social space in its own right.

Modern programming treats gathering as a function to accommodate. The conference room is a specification item. The lobby is a circulation element. The breakout space is what happens when you run out of private offices. All of that is reasonable and it misses what the best buildings get right almost instinctively. Gathering is a relationship between a space and the people who inhabit it, and it needs design attention of a different kind from a private workspace or a technical utility.

The threshold is the most underdesigned element in most contemporary buildings. Arrival is the moment where a building either welcomes you or merely admits you. The buildings people linger in at the end of the day tend to have thresholds designed with the same care as the primary spaces. The courtyard at the Aga Khan Museum in Toronto and the entry sequence at the Kimbell in Fort Worth are the result of designers who understood that the first impression of a building is the moment of crossing into it.

Buildings designed for repair are usually cheaper over twenty years than buildings designed for replacement. Construction economics optimises first cost and discounts lifecycle cost as a risk. The buildings that need repeated replacement, with curtain walls that cannot be reglazed and proprietary systems that vanish when the manufacturer moves on, show consistently that the cost of repair aversion is real, substantial, and usually borne by owners who were never the original developer.

The design vocabulary of repairable buildings has been understood for centuries. Material systems with long supply chains and broad craft knowledge. Structural redundancy that lets sections be replaced without compromising the whole. Spatial hierarchies that let secondary spaces adapt without disturbing primary ones. And an assumption that the building will still be in service long after everyone who built it has retired.

The projects worth committing to, in architecture as in technology, answer yes to three questions. Will someone want to return here? Can this be repaired when it breaks? Does it create the conditions for gathering, for the unscheduled conversation and the shared space that builds trust between people in different roles?

These are structural questions. They tend to produce buildings, organisations and systems that outlast the conditions that created them.