Rob Hornby
Co-Chief Executive Officer - London
My parents told me that, as a child, I tended to do everything “the hard way”.
Rather than buy programs for my home computer on tape or cartridge, I insisted on typing them in from magazines, a feat that typically required several hours of laborious transcription.
Similarly, my experiments with electronics involved buying individual transistors and resistors and plugging them into a solderless circuit maker called a “breadboard”.
One rogue character in the program, or a single misaligned electronic component, would result in a rather anticlimactic non-event, rather than the promised rotating 3D shape or photo-activated light.
I thought I had left those days behind, but AI has brought them back. Getting results from models and agents is another endeavour that must be done the hard way, and it is unforgiving of misalignment.
Getting real returns from AI
These characteristics are frequently glossed over in public discourse, where the impression is often given that corporate AI implementation is like a self-pitching inflatable tent that pops up unaided and ready to use.
However, my observations, now backed by good research, attest that companies are struggling to get adequate returns on their AI investments. They, too, are experiencing anticlimactic non-events for anything but narrow applications.
Some commentators claim to know exactly what the problem is: just fix this one thing, and returns will abound. Sadly, these mono-solutions are typically necessary but not sufficient, often aligning with whatever the company is selling.
The reality of AI returns is that they fall into the category of “nothing works until everything works” and consequently can only be achieved the hard way. This is not always a welcome message amongst business leaders when so many better-sounding alternatives are in circulation.
But it does explain the all-too-common state of “pilot purgatory”. These proofs of concept tend to involve an AI model delivered in a vacuum. The mock-ups look impressive and almost ready for prime time, but that is an illusion.
My new agentic engineering colleagues at Artium estimate that pure AI engineering may account for as little as 20% of the total effort required to drive outcomes. Reuters similarly suggests that AI platforms account for a minority, even of the necessary technology work.
The rest consists of selecting and scoping the best business problem or opportunity to work on, building a compelling case for it and communicating this well to those affected, assessing and mitigating risks, wrangling poor-quality data to avoid wild hallucinations, and adding new interfaces to old technology so the AI agent can actually make something happen.
The “longest pole in the tent” may well be the Chief Technology Officer approvals required to alter legacy SaaS systems, or the work of convincing a sceptical Chief Risk Officer that an agent is safe to deploy.
Forward-deployed transformation
Many claim that forward-deployed engineers are the unlock we have been waiting for, and there is no doubt that experienced technical resources who know the AI platforms inside out can make a big difference in corporate implementation. But anyone who talks to one of these rare creatures will find that they are the first to explain, “It’s messy out there”, and they do not just mean the technology.
Orchestrating all the necessary tasks requires something more than even the best engineers. Forward-deployed transformation is also needed because creating complex alignment across an organisation during AI implementation is just as much a feat of change management as it is one of technical brilliance.
This multi-dimensional complexity also explains why taking one clearly scoped AI project from idea to scaled live service is far more effective than multiple pilots. This route galvanises the ecosystem of “everything” required to achieve a result, rather than repeating the minority part of the work over and over again.
Which leads to a bold recommendation: no more AI pilots unless they are a stepping stone to scaled, live adoption and in the middle of the plan, not at the end.
What should business leaders do?
As a result, an assessment of current AI portfolios is needed:
- How many of your projects are scoped to go all the way to live service versus being characterised as a “phase 1” that ends in a pilot? A vague follow-on aspiration does not count.
- Do all AI investments come with a documented business objective and scope, a quantified value case, a change and communications plan, clearly identified risks and mitigations, a strategy for remediating poor data, a defined schedule for the legacy build, a credible route to deployment, and provisions for running the service once live? If any element is missing, expect a “non-result”.
- Can the business leaders who are responsible for the areas covered by the investment explain exactly what is happening? Do they say they are personally responsible for the outcome?
- Are employees affected by the change informed and positive about the initiative? Cases of corporate AI sabotage are increasing when AI is considered a threat to workers’ futures.
- Who is orchestrating the transformation? If the answer is a technical person, then it is imperative to add experienced business change leadership. This lesson is at least 40 years old, and AI does not defy gravity on this point.
- Finally, do you have a consolidated picture of all these initiatives and their benefits expressed in clear chronological terms? Are you confident discussing this with the Board or investors? Simply asking for one will separate serious projects from AI fluff.
Conclusions
AI is no self-pitching tent. Getting results is closer to building a circuit on the breadboard: hours of patient specification and assembly, everything in its place, each component in sync with every other, and nothing happening until it is.
Treated as a destination, the pilot was always a false dawn – too easy to make it look compelling, but too much of the real work left for afterwards.
Retiring it as the goal is more than symbolic. It signals that AI is now judged on measurable results, not positive messaging and corporate learning.
However, “hard” is not the enemy of achievement, as long as we know what we are getting into.
My old coding magazines are available for anyone who wants to train up.