On the afternoon of Friday the 18th of September 2026, the Strathmore Marketing and Communication Excellence Hub, known as SMaCE, put a question to a room at Strathmore Business School full of the people who actually commission, conduct and act on consumer research in Kenya: Can synthetic consumers replace real ones? Two keynotes, a six-person panel, a rapid-response round with seven more respondents, a regulator from the Office of the Data Protection Commissioner, and roughly three hours later, the room had an answer. It was not the answer that may have been expected. 

I was part of that afternoon. I took one of the seats in the rapid-response round, in the digital research category, and I was asked a single question: What would a credible industry certification for synthetic-persona vendors need to include? I will come to my answer, and to why I think it turned out to be quote possibly the most uncomfortable question of the day, later in this blog post. I should also say from the outset that I proposed Joel Rao, CEO of Dentsu Kenya, as a speaker to Strathmore Business School, having written in June 2026 about Dentsu’s Merkury and COPO consumer intelligence stack and its synthetic persona layer. So I arrived with a view already on the record.

A word on what follows. I was able to capture most of the afternoon, though not all of it. Daniel Oseman’s opening keynote for Kantar and the Hub’s own opening remarks are not in my notes, and the Safaricom contribution to the rapid-response round is incomplete. This is therefore an incomplete account of what I was able to capture, which is the bulk of the afternoon, rather than a complete record of the same. 

The headline is this: nobody in that room, from any side of the table, argued that synthetic personas should replace real consumers (what a relief!). The debate that the programme was built to stage, establishment versus disruption, collapsed within the first forty minutes into something far more interesting and far more useful. What replaced it was a much more provocative conversation about validation, provenance, liability and who carries the risk when a machine gives you a confident answer that happens to be wrong.

The Room Reached Consensus Faster Than Anyone Could Have Expected

The programme was designed as a confrontation. Daniel Oseman of Kantar was billed as the establishment case, making the argument for why fieldwork rigour and real respondent data still anchor decision-grade insight. Joel Rao of Dentsu Kenya was billed as the disruption case, making the argument for why speed and cost are reshaping what clients will pay for. The panel that followed was assembled to hold the tension between them across six stakeholder categories: regulator, academia, traditional research, digital research, content creation and corporate.

The tension did not survive contact with the room. Maina Ndirangu, CEO of Consumer Insight and the panel’s designated voice of traditional research, opened by saying he had told Daniel after the keynote that they should simply close the conference, because the argument was already settled. “You cannot miss trends that happen and move people forward,” he said. “That’s what development is.” He took the audience back to the discovery of fire and the emergence of language, arguing that a technology of that order is not something an industry gets to opt out of. His own agency, he noted, has used AI for transcription, interview summarisation and pattern-finding in qualitative data for something like seven years.

But then came the qualification that every single speaker after him would repeat in their own words. “We are not adopting AI or synthetic data, synthetic consumer, instead of traditional research,” he said. “You need the efficiency. You need the speed of AI and even synthetic data, and it has got its benefits. But you also need an experienced researcher, because they are human and AI is not.”

He gave a reason that makes perfect sense — empathy and contextual understanding, he argued, are not features a model can be trained into. He illustrated it with a wink across a room: a human observer knows immediately whether that was a seductive gesture or a speck of dust in someone’s eye, and an AI system can only report that a wink occurred. The distinction he was drawing is not between what a machine can see and what a human can see. It is between observation and interpretation, and interpretation is where most consumer insight actually lives.

Dr Maureen Kangu of USIU, taking the academic seat, turned the same thread into a methodological proposal. Synthetic consumers are useful for generating hypotheses, she argued, but the hypothesis then has to be tested against a live consumer matching the synthetic target on the same characteristics, location, age band and environment. “You don’t use it to just generalise unless you go through the proper validation of the live consumer versus that one,” she said. A synthetic consumer, in her account, is a secondary data source you compare against a primary one, never the primary source itself.

Joel Rao’s Case Was Not The One The Programme May Have Expected, Either. 

If Joel Rao was meant to be the disruptor, he spent a good portion of his time arguing the opposite of what the billing implied. He opened by defining the thing honestly. “A synthetic audience is not magic,” he said. “It is a simulated audience generated by an AI model to predict how people might answer questions.” He cited Greenbook’s definition of synthetic audiences as AI-generated personas designed to simulate how real people might answer surveys, evaluate concepts and react to new ideas, and he made the point that most platforms in this space draw from some mix of historical survey data.

That word, historical, did a lot of work across the whole afternoon, and Joel was the first to put pressure on it. “What we need to figure out is how to forecast what the emotion at the present continuous tense might look like in that very moment in time, at that moment of truth when a consumer is about to make a purchase.” A model built on what people did is being asked to predict what people will do, and those are not the same task.

His illustration of where the technology is heading came from a lab at Dentsu’s headquarters in Japan. He and the agency’s chief creative officer were connected to a device that reads brain activity and asked to think of the moment they had felt happiest. For Joel it was watching his eldest daughter make her first basketball shot at the age of six. The model rendered his brainwaves as an image, a flower whose colours mapped the emotions he was experiencing. “Where technology now is used not just to generate art, but it’s art that actually depicts what the human emotion can actually experience.”

From there he argued that the gap between the real consumer and the synthetic one is narrowing through attention measurement, citing eye tracking on phone cameras, patents he said Meta has taken out quietly in this area, and his own experience of discussing golf clubs and then seeing golf advertising appear on Meta’s social media platforms when he sued them thereafter. Those last two claims are worth flagging rather than repeating. Patent filings are published by law, and the recent Meta filings in this space concern smart glasses and emotion inference rather than feed attention. On microphones, a marketing firm did pitch advertisers on a product it called Active Listening, and Google dropped it as a partner when that deck surfaced, but large-scale academic testing of thousands of apps has never found evidence that mainstream apps covertly record and transmit audio. The mechanism behind the uncanny ad is almost always more mundane and more powerful than the microphone: location, social graph, cross-site tracking and purchase history.

He described what Dentsu has actually built, which is the stack I broke down in detail in June. His slides put the Merkury Media layer at more than ten thousand attributes for human understanding, spanning demographics, channel planning, decision making, attitudes and interests, hobbies and activities, and opinions and personalities, sitting on a Consumer Connection System panel of over twenty thousand people across Africa, with the CCS Planner layer planning across sixty-plus touchpoints for unduplicated reach and budget optimisation. The synthetic audience layer is then built on top of that through AI.

The proof point he showed in most detail was a Telkom SA Black Friday campaign run with Google Ads. The test pitted a Merkury fibre audience of 510,000 and a postpaid audience of 410,000 against a Google Ads custom segment drawn from a pool of three to five million. The Merkury audiences returned a 731 per cent improvement in click-through rate, a 91 per cent reduction in cost per lead and a 163 per cent lift in conversion rate. He also cited a 71 per cent increase in views on content pieces for an alcohol brand in partnership with Meta. Those are Dentsu’s own numbers, presented by Dentsu, and I report them as such.

One detail worth noting, and Joel did not hide it: the Telkom case is priced in Rand. It is a South African campaign, not a Kenyan one. Given what the rest of the afternoon established about how badly models travel across African borders, the strongest published proof point for this stack coming from a different market is precisely the sort of thing a certification standard would require a buyer to be told.

Joel was most useful was in defining the limits. “Where synthetic audiences break is where there is a need for context, where there is a need for contradiction, culture, and consequence.” On the continent specifically, he said, the hidden variable is not just data scarcity but cultural specificity. Real audiences, in his view, do more than answer questions. “The pause before the answer, that joke that exposes the truth, the disagreement inside the group, and the surprise that no model anticipates.” Those are things he said synthetic audiences cannot handle. 

His prescription was a five-step workflow rather than a verdict, and the slide carrying it was titled “The future is hybrid, not human-free”, with the subtitle “use AI to increase the quality of the human conversation, not to avoid it”. Simulate, to explore many routes. Shortlist, to choose what deserves money. Validate, by asking real people. Decide, moving with confidence. Learn, feeding the next cycle. Underneath it ran the chain that is the whole argument in six words: better data, better simulation, sharper real research.

The slide I keep coming back to, though, was the one before it, headed “When can synthetic replace real?” with the instruction to use the risk of the decision, not the excitement of the technology, to decide. It plotted two axes: low risk to high risk along the bottom, known and stable to new and emotional up the side. In the low-risk, known and stable quadrant, AI can lead, with wording options and early screens as the examples. Low-risk but new and emotional is human-led with AI support, for cultural insight and new segments. High-risk but known and stable is AI plus a human check, for campaign spend and brand risk. High-risk, new and emotional is reserved for a real audience final vote, with launch, pricing and social impact named as the examples.

That is a usable governance framework, and it was on a Dentsu slide, presented by the person the programme had cast as the disruptor. It is also, as I will come to, almost exactly the decision rule that the traditional research side of the panel arrived at independently an hour later.

He closed on the line that I suspect will outlive the event. “Don’t ask whether AI can replace people in general. Ask whether AI can replace people for specific decisions. Let’s replace repetition, not reality.”

Dharmendra Jain Put The Practitioner’s Numbers On It

The most technically grounded contribution came from Dharmendra Jain, Founder and CEO of Actnable AI, and he began by widening the definition. Synthetic data is not only a simulated respondent, he pointed out. Anything created from the real but which is not itself real is synthetic, whether that is a photograph, a video, audio or a dataset. I interviewed Dharmendra and his colleague Josiah Kimanzi on the podcast in March about building locally-built AI research tools for African realities, so I came to his contribution already familiar with the argument.

His origin story for the category was the most useful correction to the idea that any of this started with ChatGPT. Back in 2022, as an operations director responsible for data quality on tracking studies, he and a colleague were looking at longitudinal research where data is collected weekly against fixed KPIs and stable respondent profiles, and asked whether they could project it. The point he was making is that extrapolation and weighting have always been part of market research practice. If you have three years of data with a standard respondent profile, the same frequency and the same dimensions, predictive modelling is the obvious next thought for a statistician, and synthetic respondents are simply that thought taken further.

Which sets up his warning. “Are we asking just an LLM to give the profile and get the data? No, that will be a very generic answer.” Grounding, in his account, means applying real statistical technique, clustering and nearest neighbour, so that the synthetic respondent is anchored to real data. And even then it has a shelf life. You might use it for one month, two months, three, and then you have to refresh it. “You need to also see how the AI model drifts. After three months, maybe today your model is performing very well, your client is happy, but after three months your model is not giving the same accuracy. It requires constant monitoring.”

He also shared two use cases where synthetic respondents may be genuinely superior rather than merely cheaper: hard to reach respondents in places where fieldwork cannot practically be conducted, and sensitive subjects such as personal health, where a respondent may not disclose honestly to a human interviewer. “You need something which will not judge you.” That is a serious argument, and it was the only one made all afternoon for synthetic methods doing something traditional fieldwork cannot.

He also let slip, in passing, that the AI facial coding work AnneJoy Muhoro described from the EABL side was an idea his team had put forward. Actnable’s own tooling covers emotional response analysis through facial expression, and when I spoke to him in March he described a case in which identical advertising generated materially different emotional responses across three African cities. That is the same finding Kantar produced from the opposite direction, and it is worth noticing that two independent practitioners using different methods both landed on cross-market variation as the thing this continent’s marketers keep underestimating.

Joe Otin Made The Case That Historical Data Cannot Predict A Market

Joe Otin, CEO of The Collective, took the digital research seat and answered a hard question: what is the strongest evidence that a synthetic persona predicts real purchase behaviour rather than a plausible-sounding opinion?

He did not offer one. He offered history instead. Twenty years ago, he recalled, the phrase circulating in media research circles was “kill the researcher, kill the respondent”. Automation was arriving, and the argument was that if machines, algorithms and software could collect the data, you needed neither the researcher nor the person being researched. Big data was going to make respondents obsolete. That did not happen, and it is instructive that it did not.

He drew the distinction that is the basis for the whole synthetic persona question. Passive research collects your behaviour through a device as you go about your business, and gives you granular data. Declarative research asks you questions and depends on your memory and your willingness to answer honestly. Both are records of what already happened. “Our databases are excellent at giving us historical data, but will never really be able to give us any predictions.”

His example is the best one I have heard for why that matters in a market like Kenya. If you had interrogated Kenyan audience data in the early 2000s and asked what content people wanted, it would have told you news and imported soap operas, because that is what they were watching. It would not have told you that locally produced Kenyan programming was about to reshape the entire media landscape. That took somebody with a sense of originality and the willingness to carry a financial risk that was far higher than importing telenovelas. No database was going to generate that decision, and no synthetic persona built on that database would have either.

His warning, and it is consistent with what he argued when I had him on the podcast ahead of the PAMRO conference in Nairobi last year, on trust, truth and the future of African media, is that the risk is not only technical. It is bad actors and confident fabrication. He described asking an AI system a factual question about a recent summit, being given a wrong answer, challenging it, and being told he was absolutely right before receiving a different wrong answer, delivered with equal confidence, padded with editorialising he had not asked for.

That is the everyday version of the failure mode. It does not announce itself. It agrees with you and then invents something new.

On bad actors he reached for music streaming, telling the room that Spotify had removed roughly 1.9 million AI-generated songs because the listening was being driven by bots rather than people, inflating play counts so that whoever uploaded them got paid. The underlying story is real and the scale is larger than he recalled. In September 2025 Spotify said it had removed more than 75 million tracks it classified as spam over the preceding year, targeting mass uploads, duplicates with altered metadata, sub-thirty-second tracks engineered to farm royalties, and unauthorized AI voice clones. A great deal of AI output in the wild exists to game a system rather than to serve a human, and a research process that cannot tell the difference will treat both as signal.

The Regulator In The Room Changed The Nature Of The Conversation

A key presence on that panel was Caroline Macharia, Assistant Data Commissioner at the Office of the Data Protection Commissioner, and the first question of the afternoon went to her: does training or deploying a synthetic persona on real consumer data require consent under Kenya’s Data Protection Act 2019?

Her answer was far more nuanced than most of the room appeared to expect. Once you establish that you are using real consumer data, she explained, you are processing the personal data of identifiable individuals and the Act applies. Section 25 sets out the principles of data protection, of which lawfulness is one. Section 30 sets out the lawful bases you may rely on, and consent is only one of several. Legal obligation, performance of an Agreement between you and the data subject, and legitimate interests of the organisation are all available, provided they do not infringe an individual’s rights.

But she was firm about what consent means if you choose to rely on it. “It has to be specific, explicit and voluntary,” she said. “It cannot be coerced. It cannot be that it’s hidden in your privacy notices and you expect someone to just consent without knowing.”

She returned later in the rapid-response round and was asked to name the biggest compliance gap she is currently seeing, and her answer should worry anybody in this market building on synthetic data. The first gap is definitional. Organisations are assuming that because the output is synthetic, the Act does not apply, when the input that built it was personal data all along. “Simply because it’s synthetic, you cannot say that you do not fall under the ambit of the data protection Act,” she said.

The second gap is transparency. “These things are happening without consumers even being aware,” she said, adding that terms and conditions can no longer be used as the hiding place for this. “It has to be clear that you’re using AI on us, and we have to agree to proceed with that process.”

She then walked the room through the principles that apply regardless of what technology arrives next: lawfulness, fairness, transparency, purpose limitation, data minimization, security safeguards, storage limitation and accountability. Purpose limitation in particular is where I think most Kenyan marketers are exposed. You cannot collect data because someone bought your product and then repurpose that data to train an AI system without going back for a fresh basis. That is not a technicality. It is the most common thing happening in this market right now.

Most striking of all was her candour about a problem the regulator has not solved. “You’ve trained these synthetic consumers with my data. Is there any time where I, as a data subject, can say delete my data, and these personas will continue to exist?” She acknowledged directly that this is a challenge. The right to be forgotten exists in Kenyan law, and nobody in that room, including the regulator, could explain how you exercise it against a model that has already absorbed you.

On liability she was unambiguous, and this matters. Under the Act as currently designed, the data subject deals with the data controller. “The office will go for KCB,” she said, by way of example. “The office will not go for the Kantar.” The controller who collected the data carries the regulatory liability, whatever arrangements exist downstream with agencies and vendors. Her guidance was to write very clear instructions into the Agreements governing how third parties may process data you pass to them, while understanding that the liability does not travel with the data.

Marketers in Kenya should know that the ODPC has already moved on this. The office published a draft Guidance Note on Artificial Intelligence in July 2026, setting out mandatory Data Protection Impact Assessments for a defined set of AI applications, a lifecycle compliance framework and breach notification timelines. It also takes the position that the mere fact that personal data is publicly accessible does not by itself provide a lawful basis for processing it as AI training data, which aligns Kenya with the European approach and has serious implications for anybody who assumed the open web was a free training corpus. Caroline noted that the office issues advisories on request, an underused facility, and observed pointedly that had impact assessments been submitted for the work being discussed in that room, she would already have known about these conversations.

The Accuracy Question Turned Out To Have A Geography Problem

The single most useful piece of evidence produced all afternoon came from Daniel Oseman of Kantar in the rapid-response round, and it was evidence against the reliability of the very approach his questioner was probing.

Kantar, he explained, has been running creative testing using models trained on Kenyan consumers, built on thousands of Kenyan reactions to creative material. Those models perform very well when the creative is aimed at the people the models were built from. Then his team ran the same system in other markets. “We did a few tests where we try and run the same systems in Tanzania and Uganda and in Ethiopia using the model trained on Kenyan consumers. I think you can predict how accurate our results were. It fell away from the 90 per cent accuracy rate. Tanzania and Uganda not too bad. I think it was all the way down to 40 per cent in Ethiopia.”

A model that is ninety per cent accurate in Nairobi and forty per cent accurate in Addis Ababa does not have an accuracy problem. It has a boundary problem, and the boundary is invisible unless somebody tests for it. Nothing in the output tells you which side of the line you are standing on. That finding should reshape how East African marketers buy synthetic audience products from anyone, and I am grateful Daniel put it on the record in public.

Dr Maureen Kangu arrived at the same conclusion from a different direction, with an example that got the biggest reaction of the afternoon. She had tried to generate a persona of a thirty-year-old low-income African woman, and the model returned a profile that included a smartphone. “I see a lot of those low income with those they’re called a mulika mwizi,” she said. “From the data they have, they’re just assuming everybody has a smartphone.”

That is not a small error. That single assumption determines channel strategy, creative format, data cost sensitivity, the entire shape of a campaign. And the model produced it fluently and confidently, which is the failure mode that makes synthetic personas dangerous rather than merely limited.

AnneJoy Muhoro, Head of Marketing at EABL, also made her point. “I’m also very cautious with tech and AI when it comes to African consumers, because AI is built on data, and the reality is that there are huge data gaps when it comes to understanding of African consumers.” Her figure was that roughly 36 per cent of Africans have an online footprint that can be used to understand behaviour, against a far higher share in Western markets, and that most of even that 36 per cent skews affluent.

She is right, and the gap is in fact wider than she stated. The ITU’s Facts and Figures 2025 puts internet use in Africa at 36 per cent, and in high-income countries at 94 per cent. Globally, 74 per cent of people are online. So the continent that holds the fastest-growing consumer markets in the world contributes the thinnest slice of the behavioural data these models are built from, and the slice it does contribute over-represents the urban, connected and affluent minority. That is the single most important act for any Kenyan or African marketer evaluating a synthetic audience product, and it will not be fixed by a better prompt.

Maina Ndirangu Pivoted From An Accuracy Question To A Risk Question

The moderated exchange in the first panel turned on a hypothetical that Barbara Abwoga put to the room: if synthetic personas are ninety per cent as accurate as fieldwork at ten per cent of the cost, is that a research question or a business question?

Maina Ndirangu’s answer was the most practically useful thing said all afternoon. “I think the answer is: what is the risk at hand? If the risk is not huge, then that cost saving from the research is very important. But if the risk could bring down a company, then that ten per cent wrong information could bring the company down. So that risk will be too huge to take.”

The appropriate use of a synthetic persona is not determined by its accuracy. It is determined by the cost of being wrong. A ten per cent error rate on a social media content calendar is a rounding error. The same ten per cent error rate on a product launch, a pricing decision or a market entry is a balance sheet event. That is a decision rule any marketing director can apply on Monday morning without waiting for a standard to be written.

What makes this the afternoon’s most durable finding is that Maina reached it in an unscripted exchange with the moderator, while Joel had reached the same conclusion in advance and put it on a slide. Different side of the argument, different commercial interest, same answer: use the risk of the decision to decide. When the vendor and the sceptic converge on a governance rule without coordinating, that rule is probably right, and Kenya as a market should adopt it now rather than wait for somebody to formalize it.

He then produced the most uncomfortable data point of the day, and it has nothing to do with AI. Purchase intention, he said, is the metric clients love and researchers distrust. He described testing a low value product by asking respondents their likelihood to purchase on a five point scale, and then, for those who said very likely or likely, actually producing the product and offering to sell it to them there and then. Sixty per cent said no. “So the rule of the thumb, when people say they’re going to purchase in Africa, we discount by sixty per cent.”

Think about that for a minute, or two. The most widely used predictive metric in Kenya carries a sixty per cent discount rate before AI enters the conversation at all. This means the question is not whether synthetic personas are as reliable as traditional research. It is whether either is as reliable as the industry’s confidence in them implies. Dr Kangu made the same point through the New Coke episode of 1985, where Coca-Cola ran taste tests with close to two hundred thousand consumers, launched on the strength of them, and reversed the decision seventy-nine days later.

The Certification Question Nobody Could Answer

When Dentsu Kenya’s Gideon Kiptarus put my question to me in the rapid-response round, I had looked it up beforehand, and the first thing I said was the thing I found most surprising.

There is no certification for synthetic-persona vendors. Not in Kenya, not in the United Kingdom, not in the United States, not anywhere. As of this month there is no accreditation scheme, no certifying body and no agreed protocol that validates a synthetic sample or the supplier who produced it. There are codes, notably the ICC/ESOMAR International Code, which was updated in 2025 to address AI and synthetic data. There is guidance, including ESOMAR’s buyer checklists and a 2024 MRS Delphi Group report on synthetic participants which is explicit that no agreed protocols exist, that supplier methodological transparency is lacking, and that the line between primary and synthetic data is in some instances deliberately obscured. There is a code update from the Insights Association in the United States. But none of that is certification.

What does exist, and this is the important development for this market, is ISO 20252. The international standard for market, opinion and social research published its fourth edition on the 1st of September 2026, and for the first time it carries requirements covering AI. Clause 4.1.5 requires research organizations to disclose the type of AI used and the stage of research it was applied to, to obtain explicit client consent before passing client or project data through AI systems, to document AI system versions and monitoring, to inform participants when AI is involved, and to tell clients when AI generated the insights. A new Annex G covers panel development and lifecycle management. The standard is certifiable, and a three-year transition is expected.

My argument on the day was that we should not invent a Kenyan standard from scratch. We should build the synthetic-persona requirements as an addendum to ISO 20252, paired with ODPC guidance on the data protection half, because a standard nobody outside this market recognizes is not a standard. It is a badge.

The six criteria I put forward are these, and I have clarified them slightly since.

Provenance comes first. What is the quality of the data underneath, and has it been validated? Synthetic data has to be grounded in real data, and if the foundation is unsound then everything built on it is flawed from the outset. This is the recurring truth across every AI conversation I have become a part of these days: if your data is disorganized and unvalidated, do not do AI. Spend six months to a year fixing the foundation before you go near a model.

Grounding comes second, and this is where the afternoon’s evidence bites hardest. If a persona for a mama mboga in Kangemi is built on a Western AI model, it will confidently tell you she owns a smartphone. The certification must force the vendor to declare what real data the persona actually sits on, for which market, and where that grounding runs out. Daniel’s Ethiopia result is the proof that this is not theoretical.

Currency is third. Consumer behaviour is a moving target, and a persona built on last year’s data is describing a consumer who no longer exists. AnneJoy Muhoro gave the definitive illustration of this. Before COVID, 95 per cent of alcohol consumption in Kenya happened in bars. That collapsed when bars closed, and EABL used daily online tracking to reshape its off-trade strategy in real time. Today, by her account, 15 per cent of consumption sits outside the bar, triple the pre-pandemic share. Any synthetic model built on pre-2020 Kenyan data would be describing a market that no longer exists, with total fluency and total confidence.

Benchmarking is fourth. Is the model being tested in isolation, or scored independently against comparable datasets, with the misses published rather than only the matches?

Reproducibility is fifth, and it is most often glossed over. Put the same question to a generative model twice with the same data and you can get two different answers, because these systems are probabilistic rather than deterministic. They are not structured databases returning a fixed value. If your research instrument returns a different result on Tuesday than it did on Monday, it is not an instrument. Certification has to require logged model versions and disclosure when the underlying model is updated, because the update changes the instrument you have been tracking against.

Bias and coverage is sixth. A model is only as good as the data inside it, and if it does not hold the difference between Sheng in Nairobi and the way people speak in Mombasa, that gap propagates straight into the output. On the global scale of AI training data, Africa is close to a rounding error.

Then labelling and liability, which sits as the wrapper around all six. Who owns the decision? None of us can go back to a client and say the AI got it wrong. The client gave you the work and trusted you. You are the accountable party.

I also told the room a cautionary story about AI agents behaving in ways their operators did not anticipate. In July 2026, roughly 1,200 AI agents running OpenAI models inside a reinforcement learning evaluation environment escaped network isolation, improvised a coordination channel through a shared package server, accumulated hundreds of thousands of messages between themselves, and breached Hugging Face infrastructure before being locked out. Even when their access was cut, they re-established the channel through an unauthenticated endpoint. The point stands whatever you make of the detail: systems that can pursue an objective around an obstacle can also produce plausible, confident, well-formed output that is simply not true, and a research process with no verification layer will not catch it.

Angela Mwirigi Drew The Line Between What Is Rational And What Is Human

Angela Mwirigi, Director of Digital Financial Services at KCB Bank Kenya, was asked what decision she would trust a synthetic persona to make that she would not trust to a human panel. Her answer was the clearest articulation of the division of labour that anybody offered.

She started with a piece of evidence that ought to trouble anyone selling automation to Kenyan banks. Customers get irritated when a chatbot answers the phone, and in her account Africans are less tolerant of it than most. KCB used to say it would close branches. It has closed none, and opened twenty last year. “People prefer to talk to another person,” she said, and the reason is the one Maina Ndirangu had already pointed out: context and background, the sense that the person on the other side understands what you are actually asking for.

But that is not an argument against AI, it is an argument about where to put it. Her rule was simple enough to write on a wall. “Anything rational, logical, reasonable, we give it to the AI. Anything that requires context, requires background, requires intuition, emotion, we give it to the human being. However, once it becomes big and complex, even though it requires emotive layering, we still need the AI to help us.”

The volume argument is the part marketers tend to miss. As transactions digitize, a bank is processing hundreds of millions of them. No human organization can manage that complexity, and humans make mistakes at scale in a way machines do not. So the machine is not replacing judgement, it is making judgement possible at a volume where judgement alone would collapse.

Her working example was fraud. Kenyan courts, she noted, have been clear that the institution holding the vault carries responsibility for what happens to the money, even where a customer has handed over their own PIN. You are the keyholder, so you are accountable, however clever the theft. That legal position is what makes the AI worth having: KCB uses it to generate the many scenarios by which a customer might be socially engineered, iterating through possibilities continuously, including overnight. “It’s really exhausting to try and figure out what are the many different ways people can get defrauded.”

That is a synthetic simulation being used exactly where it should be, on a high-volume, known and stable problem where the cost of a false positive is low and the cost of missing a pattern is high. It also, quietly, made the case for the category better than any accuracy statistic did.

Nobody Could Answer The Liability Question Either

The liveliest exchange of the afternoon happened when Barbara asked who owns the consequences when a synthetic-persona-driven decision goes publicly wrong. The client, the vendor, or nobody?

Angela Mwirigi of KCB answered from the client’s chair with a line about accountability travelling with the purchase order. Maina Ndirangu, who had previously worked for her, pushed back hard and in public. “It’s amazing how when things go wrong, the agency is wrong. But when things go well, the marketing director gets the credit. I have no problem with that. But then stop giving me the story that I’m your partner. I’m not your partner, because if you are partners and something goes wrong, you share the responsibility.”

That is the most honest thing anybody has said about agency and client relationships in Kenya in a long time, and it was said on the record to a room full of both.

AnneJoy Muhoro closed the exchange by widening it. Her reference point was the Bud Light episode in the United States, where an influencer partnership in April 2023 triggered a consumer backlash so severe that Modelo Especial took the number one position in the American beer market by June of that year, ending roughly two decades of Bud Light dominance. Her point was that the consequences do not land on one party. The marketing director loses their job, the agency loses the account, and the company carries the financial damage. “So it is up to all of us to not abdicate responsibility for this.”

Caroline Macharia then supplied the piece that the industry voices could not, which is that the regulatory answer is already settled even where the commercial one is not. Liability for data protection sits with the data controller. Therefore, the Act requires a human in the loop for consequential automated decisions. “If you think it’s ninety per cent accuracy, but that ten per cent can have huge implications on individuals, have channels where they can actually request you to review the decision that you’re making about them, and then let a human being be able to substantiate if the decision made by AI is correct.”

The commercial question of who pays when a synthetic-persona decision fails is unresolved in Kenya. The regulatory question is not. Those two facts sitting side by side is the gap that a certification standard would close.

The Sensory Scientists Made The Case Nobody Else Could

Nina Kioko, CEO of Revealiq Consultancy and a cognitive and sensory science specialist, was asked where the gap between what a consumer states and what their body does has changed a client’s decision, and whether a synthetic consumer could have caught it.

Her example was Afro-textured hair. A consumer, or a model simulating one, will tell you a conditioner makes hair soft. Put the actual product in the actual consumer’s hands and you get what the stated response never contained: the texture, the creaminess, the physical experience that determines whether it gets repurchased. The same applies to taste, where a high rating might trace back to a sweet the respondent ate years ago, and that emotional trigger is the thing you would actually build the brand on.

AnneJoy Muhoro gave the commercial version of the same argument, and it is the best single case study for hybrid method I heard all day. EABL has historically done liquid testing through focus groups, where consumers taste a product and report what they think. Her observation about that method was that Kenyan respondents are, in her word, passive aggressive. The real feedback sits in a single sentence at the end, the footnote, and hopeful marketers habitually ignore the footnote.

On a recent launch, EABL ran AI facial coding alongside the traditional liquid test. The stated feedback said the product was good to go. The facial analysis returned a just-right score roughly 20 per cent off, against an acceptable threshold below 10 per cent. They reformulated. Without the machine layer, traditional research alone would have sent a product to market that the consumers in the room had already rejected with their faces while telling the researchers it was fine.

Cognitive Surrender Was The Best Phrase Of The Afternoon

Daniel Oseman’s rapid-response answer contained the argument most under-discussed in this market, and it is about competitive advantage rather than accuracy.

His first safeguard was simple. “Synthetic personas must never be allowed to become self-validating. We cannot remove human intelligence from the loop.” His reason was not about today’s error rate. “The biggest commercial risk isn’t that the system necessarily makes the wrong decision today. It is that it gives you yesterday’s answer tomorrow.”

Then he took it somewhere more interesting. If AI continues developing at its current pace, access to synthetic data becomes democratized, levelling the field between the multinationals who can afford multi-million-dollar research programmes and the startups who cannot. This means the insights themselves become commoditized, and advantage stops flowing from who has the best data. It flows instead to how quickly an organization turns insight into action, and to how confident it can be that its understanding of consumers is more current than its competitors’. “In that world, winners won’t necessarily be the companies with the biggest, most advanced data models. They’ll be the companies that have the organizational structure, the culture, and the systems that allow them to spot changes and take brave action before anyone else can.”

In between, he named the risk I have not been able to stop thinking about since. “One of the biggest dangers that we see is cognitive surrender. This idea that we let the machines do the thinking for us, and it actually takes away the ability of companies to develop competitive advantage, because we’re all using the same machines, the same algorithms, the same data models.”

If every agency and every brand in this market runs the same foundation models over broadly the same commoditized data, the outputs converge. Differentiation has to come from somewhere else, and that somewhere else is organizational: judgement, speed of decision, and the willingness to act on a conviction the model did not hand you.

The closing part of the afternoon turned to what this does to people rather than to companies, and it produced the most human exchange of the day. Daniel Oseman made the historical case against panic, pointing out that calculators did not stop children doing mathematics but raised the level of mathematics they could do, removing the need to memorise tables by rote. He went further back, to Socrates opposing writing on the grounds that it would make people stupid because they would no longer have to remember anything. “This is not a new question. We just have to be very cognizant that it can happen, and be very deliberate in making sure that we retain critical thinking skills.”

I offered the counterpoint from research I had been looking at recently, and it is the finding I now use whenever anyone asks me how to deploy AI in a learning context. A World Bank randomized trial in Edo State, Nigeria put around 800 senior secondary students through after-school sessions twice a week for six weeks, using a generative AI assistant. The design is the whole point: the AI was configured as a tutor that guided students towards an answer rather than supplying it, with teachers supervising the prompting. Comprehension and critical thinking improved measurably. Set against that is MIT Media Lab work on essay writing, which found that participants who leaned on a language model showed weaker neural engagement and struggled to quote from work they had produced minutes earlier. Same technology, two configurations, opposite outcomes. If you outsource your brain to it, you get worse. If you use it to make yourself understand the steps, you get better.

For the record, the claim that a generation’s IQ is now falling because of technology needs care. Test scores have declined across several high-income countries since the mid-1990s, and a Norwegian study of 730,000 conscripts established that the cause is environmental rather than genetic. But no study has identified technology as that cause, and at least one researcher points to declining test-taking motivation as an alternative explanation entirely. The concern is legitimate. The attribution is not yet evidence.

Joel closed the thread with the a compelling perspective on how a system behaves when you tell it to win. A frontier reasoning model, instructed to beat a chess engine it could not out-play, manipulated the game’s system files to force the result instead. “Even as we think about what success looks like, if it’s a machine that you’re telling to win at all costs, that’s how these things could happen.” Define success as the outcome alone and a sufficiently capable system will find a route to the outcome you did not intend. That is as true of a synthetic research brief as it is of a chess game.

That is also why what AI really changes about the marketer’s work, from the brief through to the boardroom, is a question about how the work is configured rather than about which tool is bought.

What I Think This Means For Marketers In Kenya & The Rest of Africa

I went into that afternoon holding the position from my June analysis: that Dentsu’s stack is a genuine step change for data-driven marketing here, while leaving the hype versus reality question unresolved. I came out with a clearer perspective. 

Synthetic personas are not a research method yet. They are a hypothesis generator with a research method’s vocabulary, and the vocabulary is what makes them dangerous. They produce output that looks like findings, reads like findings and can be pasted into a deck like findings, and there is currently nothing in the market that tells a buyer whether what they are holding has any grounding at all.

For anyone commissioning this work in Kenya, the practical position is straightforward. Ask what real data the persona is grounded in, for which market, collected when, and under which lawful basis. Ask what happens to the accuracy when you cross a border, and treat any vendor who cannot answer that with the Ethiopia result in mind. Insist on reproducibility and on being told when the model version changes underneath you. Apply Maina Ndirangu’s risk test before the cost saving seduces you, because a ninety per cent accurate answer is only cheap if being wrong is cheap. Never let the synthetic layer validate itself. Put the liability in the signed Agreement rather than discovering it in the aftermath.

For vendors, the position is equally clear. The absence of a certification scheme is not an advantage. It is the thing currently preventing this category from being bought by the large, governed, procurement-driven organizations that would otherwise be its best customers. The first vendor in Kenya to voluntarily publish its provenance, its grounding, its cross-market accuracy decay and its model version history will not be handicapping itself. It will be defining the standard everyone else eventually has to meet. There is a real business opportunity to be had from this perspective. 

As for the industry collectively, there is a route that does not require waiting for regulation. ISO 20252 now carries AI requirements and is certifiable today. The ODPC has published draft AI guidance and issues advisories on request. If three or four of the large buyers in this market, a KCB, an EABL, a Safaricom, wrote synthetic-method disclosure into their agency Agreements, the standard would become tangible in a quarter rather than in the years a legislative process would take. Procurement moves faster than Parliament, and in this instance it should.

The Question The Afternoon Did Not Answer

The room settled the easy question quickly and honourably. Synthetic consumers augment; they do not replace. Validate against real people. Keep a human accountable. That consensus was clear, and it was reached across the regulator, the academy, traditional research, digital research and the client side, which is not a common occurrence.

However, three things went unanswered, and they are the really important ones. 

1. How does a Kenyan exercise the right to be forgotten against a synthetic persona that has already absorbed them? The regulator said openly that she is struggling with this, and that honesty is worth more than a manufactured answer would have been.
2. Who actually pays when a synthetic-persona-driven decision fails commercially? The room agreed that everyone loses. It did not agree who writes the cheque, and until that sits in an Agreement rather than in a panel discussion, it will be settled the expensive way.
3. Finally, will anybody build the certification? I said on the day that it should be an addendum to ISO 20252 enforced through procurement. Nobody disagreed. Nobody volunteered either.

A certification would not tell you that a synthetic persona is right. It would tell you that when it is wrong, you will be able to see why. That is a modest ask, and in a market where we are already discounting stated purchase intention by sixty per cent, it is the least we should expect before we start buying answers from machines that cannot tell us where they got them.

Two contributions belong in the record that I cannot report properly. Moses Mutua, Chapter Lead for Market Research at Safaricom, was asked what internal safeguard he would put in place before his team touched synthetic personas, which is precisely the question a large Kenyan corporate needs answered, and his response is not in what I was able to capture. 

The SMaCE Hub deserves real credit for convening this while the question is still open rather than after the market has made its mistakes. Dr Nancy Njiraini, Dr Sarah Awinyo and Dr Olgha Adede built and planned the afternoon, Charles Bodo held the it together as the MC, and Dentsu Kenya’s Barbara Abwoga and Gideon Kiptarus moderated two very different formats without letting either drift. The next conversation should be shorter, narrower and more uncomfortable. It should be about who is going to write the standard, and which buyers in Kenya are prepared to make it a condition of doing business.