Client Zero: The AI Principle That Scales from IBM to Your Substack

Why your first client should always be yourself

🎥 Some of this article is also covered in a video on Exploring ChatGPT’s YouTube channel - take a look and subscribe!

The term “patient zero“ comes from epidemiology: it’s the first identified person to carry a disease in a given outbreak. Naturally, someone has to go first. In technology, that principle was formalised as “dogfooding” — a deliberately inelegant term for using your own product first, before shipping it to anyone else.

What has emerged more recently, as organisations attempt to deploy AI in earnest, is the concept of “client zero.” This term refers to the internal group or individual that tests an AI solution on in-house operations before it touches a single customer or external stakeholder.

The distinction from ordinary dogfooding is significant. Traditional software either works or it doesn’t: a bug is reproducible and, in principle, fixable before it turns into disaster. AI systems behave differently:

• Their outputs are probabilistic rather than deterministic, meaning the same input can produce a different result on different occasions.

• They can hallucinate, producing confidently wrong answers in ways that are hard to predict and harder to catch in a controlled test environment.

• They can encode biases present in their training data, or fail in edge cases that no simulation anticipated.

The cost of discovering any of these problems in front of a client rather than in your own operations is reputational, and reputation is far easier to damage than to build.

The client zero approach addresses this by making internal deployment a genuine pilot and proving ground before the real launch. You run the AI in your own workflows, with real consequences. You document what breaks, what disappoints and what the editing or oversight process needs to look like when the output is not quite right. Only then do you deploy it externally.

This sequence is gaining traction across industries, and the reason is straightforward: the organisations most capable of credibly recommending AI are those that have lived with it first.

What “client zero” looks like in industry

IBM is one of the most visible corporate exponents of this approach. The company describes itself publicly as “client zero,” and its Vice President for AI Transformation, Radha Plumb, has used the phrase “drinking our own champagne“ to describe the logic (note the upgrade from dog food!). This positioning frames internal deployment as genuine consumption of the product, with real stakes attached.

The results IBM has published from this internal programme are specific enough to be instructive. Within its HR function, the company built an agent called AskHR to handle routine employee transactions: it now resolves more than 90 per cent of HR interactions without any human involvement and has contributed to a 40 per cent reduction in the HR operating budget, which frees the time that HR staff need to focus on complex and nuanced cases. Across procurement, IBM has saved more than 20,000 work hours annually. These, alongside projected savings of billions of dollars per year, are the numbers that IBM now brings to federal agencies and enterprise clients as evidence. These numbers are credible because they were generated at IBM’s own expense and on IBM’s own operations.

A second case, useful for its specificity about sequence, comes from Insight, the IT solutions and managed services firm. Insight was among the earliest organisations worldwide to implement Microsoft Azure OpenAI at the enterprise level, deploying a private instance that, by design, would not feed employee prompts or data back into the public model; they rolled this out globally to their staff under the name InsightGPT, within eight weeks. The approach they took was to let early adopters across departments use the tool freely, generating real use cases and feeding their experience back into how the tool was refined. HR teams used it to analyse internal surveys, saving between one and two weeks of analytical work; a sales team used it to sort and categorise a large dataset, saving more than a hundred hours across the colleagues involved; a warehouse team used it to eliminate manual errors in a repetitive process. Only once this internal cycle had run, generating operational improvements and surfacing the limits of the tool, was InsightGPT offered to clients.

The counter-example is instructive in a different way. McDonald’s ran a pilot AI voice-ordering system across approximately 100 US drive-through locations. Customers began posting videos on TikTok of the system adding bacon to ice cream orders, confusing one car’s order with another’s, and adding over a hundred chicken nuggets to a single order before the customer gave up. The failure was not catastrophic in scale (the trial had, by design, been limited to a relatively small proportion of locations) but the public embarrassment was disproportionate.

The Art of Asking Questions is a reader-supported publication. To support my work, please consider becoming a paid subscriber.

The consulting version of the same problem

Many consulting firms face a specific and somewhat sharper version of the client zero question: they are selling AI transformation as a service and advising large organisations on how to implement technology at scale. If a firm has not demonstrably lived that transformation itself (if the advice is theoretical, derived from other engagements or literature rather than internal practice), the gap is a commercial liability.

BCG has been very upfront about positioning itself as its own first client. The firm has used its internal operations as a deliberate testing ground for understanding AI deployment in practice, describing this as a necessary precondition for advising clients credibly. The argument is that leading by example and reshaping internal work before recommending the same to others is what makes the advice substantive rather than speculative.

Deloitte has taken a similar approach. The firm’s AI Institute launched an internal generative AI copilot in January 2024, designed to assist with tasks including coding and project planning, with plans to extend it to 100,000 employees across the organisation. Again, the sequence is internal deployment first, client-facing application second.

The discipline is the same whether you are IBM saving costs in HR or a consultancy firm automating elements of project delivery: you build the evidence base on your own operations before you stake your professional reputation on it in someone else’s.

🎥 Check out our video discussing this subject on Exploring ChatGPT’s YouTube channel!

What this means if you write about AI or share AI products through your newsletter

The client zero principle is, at its core, about having skin in the game before you ask anyone else to carry the risk. Interestingly, it scales down completely from the corporate world to your own publication.

If you publish about AI and write regularly about AI tools and workflows, your readers are betting that your recommendations reflect genuine experience rather than a brief demo or secondhand enthusiasm.

That trust functions much like the reputational capital at stake in the IBM and McDonald’s examples. The difference is that a solo creator has no PR department to manage the fallout and no hundred other locations still running smoothly.

Before you recommend a particular tool (especially if you have built it yourself), the client zero principle suggests running it through your own actual processes, tracking where the AI output needed significant revision, where it produced something you would have been embarrassed to publish under your name and where the time savings were genuine and where they were illusory. That lived experience is what separates a useful recommendation from one that is, at best, accurate in general and wrong in practice.

There is a useful secondary consequence. The process of being your own client zero leads to content that is actually worth reading. A clean statement like “I tried (or made) this tool and it is great” is one of the least interesting things a writer can produce. The honest account of what the tool got wrong, how you caught it and what the editing overhead actually looked like is useful reporting. The failures, edge cases and surprises that emerge from internal testing are more instructive than the headline result (or a product’s Features page), both for the writer and for the audience.

The practical standard that the client zero principle suggests is straightforward: you should not publish a recommendation or guide that hasn't been genuinely stress-tested on your own work.

👇 Steal My Toolkit

The newsletter covers ideas, frameworks and honest takes on consulting practice. The paid tier is where those ideas become usable: you get templates you can open in a project, prompts you can run today, and research tools built from over a decade of fieldwork.

Paid members get instant access to:

• Consulting with AI Series: the custom AI prompts I use in my own management consulting work, explained and ready to adapt.

• Full Course: Asking Questions Like a Pro - A Practitioner’s Guide to Mixed Methods Research ($199 value): practical research techniques refined across 10+ years of consulting engagements, covering design, facilitation and analysis.

• Practical Guides, Handbooks, Workbooks and Templates: premium resources from paid posts or 70% off discounts for my Gumroad assets.

• Cozora discount (up to $360 off): up to 50% off a Cozora subscription to learn AI from practitioners who use it in real work.

Free subscribers aren’t left out either.

• Leader Tools Cards: 10% off with code ANDREA10 at leadertools.co/ANDREA10

• Cozora welcome discount: 10% off via https://cozora.substack.com/andre10

• Demystifying Academic Research - A practical guide for navigating academic research without a PhD: Available for free (pay what you can, $0+) via Gumroad

Disclosure: I earn a referral fee if you purchase through my links, at no extra cost to you.