The Valley of Death Between a Working Demo and a Working Business

In late August Digital News Asia published a long account of how MIMOS, the national applied research centre under MOSTI, gets technology out of the laboratory and into companies, and its central image is a "valley of death" between proving that an innovation works and making it commercially viable. The piece is MIMOS describing its own work and discloses no sales or adoption figures, so it cannot tell us how well the method performs. What it does set out, quite clearly, is the order in which a cautious organisation reduces the risk of adopting something new, and that order is more useful to a small firm considering AI than most of the vendor material I am sent.

I generally feel that our government—or semi-government—agencies are under-appreciated. I remember MIMOS doing credible work with our Internet infrastructure in the old days, even with many complainants over the years. For SMEs, having an agency that helps with ideation, technology, and moving entities towards commercialisation is no easy feat, and many Malaysian companies have benefitted from MIMOS. However, MIMOS can't do everything, nor should they. Ultimately, success for a technopreneur lies in accountability for the whole journey and not just the parts the founders are good at.

What the report actually describes

The article sits under a "Powered by Mimos" banner and ends with a note that AI was used to generate the first draft, with an editor responsible for the published version, so it is best read as a sympathetic institutional account and not as independent assessment. Within that limit the facts are specific enough to be worth passing on: MIMOS says it holds more than 1,400 national patents, runs a Semiconductor Technology Centre and an AI Innovation Hub, and licenses technology to local companies it calls technology recipients. Three examples are given, a handheld haemoglobin screening device (REVA, which uses spectroscopy and AI) licensed to Recove, an inline palm oil quality monitor (INSPECTRA) with Rodatech, and a graphene conductive ink with Tanjung Plus. Only the first of these is described as using AI, which I mention because the queue headline that brought me to the article suggested a firmer AI connection than the text supports.

The part I would underline is the sequence MIMOS recommends to companies that find advanced technology intimidating, because the worries it lists (the cost of implementation, uncertainty over performance, integration difficulty, and whether customers will pay for the result) are the same four an SME owner raises when someone proposes an AI project. CEO Razman Jaafar says the firm first gives businesses access to validated technologies, prototypes and testing before they commit, then encourages a phased approach starting with a proof-of-concept, a pilot or a sandbox deployment, with technical support to limit integration risk, and offers licensing, technology transfer or co-development only after that. The stated aim is to "progressively reduce uncertainty at every stage".

Why the same gap opens for AI

In my experience the valley is, if anything, wider for AI than for hardware, because a demonstration of a language model on clean sample data looks convincing in a way that a prototype device on a bench never quite does, and so the temptation is to skip the middle stages altogether. The businesses I see struggling are rarely those that chose the wrong tool; they are those that moved from a good demonstration straight to a company-wide subscription, and found afterwards that nobody had tested it against their own invoices, their own staff habits or their own customers' patience.

There is a fair objection, which is that an SME has no national research centre to provide validated technology and a sandbox, and a monthly subscription hardly needs a de-risking framework. That is true of the subscription, and I would not wrap a process around something you can cancel next month. It is less true once the tool touches customer data, replaces a step someone is accountable for, or is rolled out to a team, at which point the cost of being wrong is no longer the fee. The discipline is the same one, only scaled down: a named person, a defined trial on real work, a short list of what would count as success, and a date on which you decide.

What This Means for You

The report is evidence that a phased route from proof of concept to pilot to wider deployment is how a serious national institution manages adoption risk, and it is not evidence that the route works as well as claimed, since no results are given. I would take the method and leave the claims. If you are weighing an AI project this quarter, ask which stage you are really at, whether you have tested on your own work or only watched a demonstration, and who in the business owns the result if the pilot goes well. Most firms can answer the first two honestly within a few minutes, and the third is usually where the hesitation shows.