Essay · Trust

Designing for trust before scale

Institutions race to scale what they have not yet earned the right to scale. Trust is not the reward for scale. It is the precondition.

There is a reflex in every transformation programme. A pilot works, or seems to work, and the next question is always the same: how fast can we roll it out? Scale is treated as the proof of success, and speed as the proof of seriousness.

The reflex is understandable. It is also, increasingly, the reason so many promising ideas die quietly in their second year.

What the numbers are telling us

BCG's 2025 research found that 60% of companies report little or no value from AI, while just 5% achieve value at scale. Gartner expects that over 40% of agentic AI projects will be cancelled by the end of 2027. These figures are usually quoted as warnings about technology. We read them as warnings about sequence.

Most failures we see described are not failures of the first demonstration. The demonstration works. What fails is everything the demonstration did not have to face: real data in its messy entirety, real users with real incentives, real exceptions, real auditors, real customers who complain. Scale exposes the parts of a system that were never designed, only assumed. When those parts are about trust, the exposure is fatal.

Scale does not create trust. It multiplies whatever trust, or distrust, was already designed in.

Trust is an engineering requirement

Trust is often treated as a communications task, to be handled after launch with a good explanation and a reassuring tone. That is backwards. Trust is a property of the system, and it can be specified like any other requirement. We think of it in four layers.

Purpose. Everyone involved can state, in a sentence, which decision or outcome the system exists to improve, and how improvement will be measured. A system without a stated purpose cannot be trusted, because nobody can tell whether it is doing its job.

Ownership. A named person is accountable for the system's outcomes, not merely its uptime. When something goes wrong, nobody has to ask whose problem it is.

Evidence. The system leaves a trail: what it was given, what it did, and why. Evidence is what lets a board, a regulator or an affected person check the institution's claims rather than take them on faith. Where the system makes decisions about people, South African law already sets a floor here: POPIA section 71 requires that people subject to qualifying automated decisions can make representations and receive enough information about the underlying logic to do so.

Reversibility. The institution can switch the system off, roll back a change or route around a failure without crisis. A system that cannot be stopped safely should not be started at scale.

Each layer costs something to build. Each is far cheaper to build before scale than after.

The sequence: prove, protect, then scale

A disciplined sequence looks less exciting on a slide and works far better in practice.

  • Prove the value on a narrow, real problem, with real data, measured against an agreed baseline. If the value is not there at small scale, it will not appear at large scale.
  • Protect the value by building the four layers of trust while the system is still small enough to change easily.
  • Scale only what has been proven and protected, at the pace the institution's people and controls can absorb.

This is slower in the first quarter and faster in the second year. More importantly, it means the institution is scaling something it understands.

Trust is also how you are judged

The same logic applies to how an institution presents itself. The Edelman-LinkedIn 2024 B2B Thought Leadership Impact Report found that 73% of decision-makers say an organisation's thought leadership is a more trustworthy basis for judging its capabilities and competencies than its marketing materials.

We take two lessons from that. The first is that serious buyers want to see how you think, not only what you claim. The second is more demanding: thinking in public creates an obligation. If an institution says trust comes before scale, its own work has to show it. That is one reason Possible Art exists, and one reason we hold it to a simple rule: name the source of every fact, and label every concept as a concept.

What gets designed out

Designing for trust also means choosing what not to build. The dashboard nobody will read. The automation that removes the one human who noticed the strange cases. The model that is accurate on average and unaccountable at the edges. Restraint is not timidity. It is the discipline of building only what you are prepared to stand behind. Every capability added at scale is a capability that must be governed at scale. Often the cheapest control is the feature you chose not to ship.

What we imagine

Imagine a programme review that asks, before any rollout decision: what is this for, who owns it, what evidence does it leave, and how do we stop it? Imagine boards that treat an unanswered question on that list as a reason to wait, and teams that are praised for raising it. Imagine scale that arrives a little later and stays.

The future we want to build is not the one that arrives fastest. It is the one that people are still willing to rely on when it does.

The Possible

These essays now feed a fortnightly edition, one industry at a time.

Edition 01 takes them into South African financial services: the AI value gap, automated decisions under POPIA section 71, identity and KYC at scale, and the operating model.