<- all tokdocs
2026.09.16@jenna_gardner_ai

Only a Small Fraction of an AI Project Is the AI, and the Rest Decides Whether It Ships

AI
Watch on TikTok

View on TikTok ->

Google engineers published the core of this argument in 2015, and their Figure 1 caption reads "Only a small fraction of real-world ML systems is composed of the ML code, as shown by the small black box in the middle." The account posts as @jenna_gardner_ai under an "Applied AI advice" banner. A Jenna Gardner carries the byline "Director of Applied AI, Media" on DEPT's own site, which matches the name and the field, though nothing on that page ties the byline to this TikTok account. The video is 30 seconds of selfie footage shot in a parked car, one static "Applied AI advice" sticker at the top of the frame, and auto-generated karaoke captions. No slides, no b-roll, no screen recordings. Her claim is that roughly 10% of an applied AI solution is actually AI, that the bulk of the work is process redesign and data plumbing, and that a company should exhaust the features it already pays for before commissioning a custom build.

The 10% figure is an estimate, and the paper behind it never gives a number

Treat the 10% as a practitioner's rule of thumb from someone who does this work, not as a measurement. She says "typically like only 10%," which is the cadence of an estimate. There is no study I could find that puts a percentage on it.

What does exist is Hidden Technical Debt in Machine Learning Systems by D. Sculley and nine co-authors at Google, published at NIPS 2015. Section 5 states: "It may be surprising to the academic community to know that only a tiny fraction of the code in many ML systems is actually devoted to learning or prediction. In the language of Lin and Ryaboy, much of the remainder may be described as 'plumbing.'" The accompanying diagram shows a tiny ML box surrounded by configuration, data collection, feature extraction, serving infrastructure, process management, and monitoring.

Two caveats matter here. The paper measures lines of code, not hours of project effort, and those are different things. It also predates the current generation of hosted models, which removed some of the training and serving infrastructure while adding evaluation, prompt management, and retrieval plumbing in its place. The shape of the claim survives. The specific 10% does not have a source.

The failure data points at workflow integration

The GenAI Divide: State of AI in Business 2025, from MIT's Project NANDA, reports that 95% of organizations are getting zero return on $30 to $40 billion of enterprise GenAI investment. The executive summary is explicit about cause: "This divide does not seem to be driven by model quality or regulation, but seems to be determined by approach." Elsewhere it states the core barrier "is not infrastructure, regulation, or talent. It is learning," meaning tools that fail to retain feedback or adapt to how people actually work.

The report is preliminary findings covering a January to June 2025 research period. Its evidence base is a review of more than 300 publicly disclosed AI initiatives, structured interviews with representatives from 52 organizations, and survey responses from 153 senior leaders collected at four industry conferences. It is not peer reviewed, and the 95% is a headline number that has been repeated far more often than it has been examined. I am citing it because it is directionally consistent with Gardner's point, not because 95% is a precise figure.

The shelfware problem has a number attached to it

Gardner's Adobe example is the practical half of her advice. Ask whether you are using the features you already bought before you pay to build new ones.

Zylo's 2026 SaaS Management Index found that organizations leave an average of 36% of their SaaS licenses unused. That figure comes from analysis of more than 40 million SaaS licenses and $75 billion in spend under management. The same research puts median SaaS spend at $9,455 per employee and finds companies adding roughly 21 applications per month. Zylo's own shelfware writeup widens the definition and reports that 53% of licenses are unused or used too rarely to justify the cost. Zylo sells SaaS management software, so the incentive runs toward a large number. The 36% is the one traceable to their stated methodology, so that is the one I would quote.

The MIT report adds a version of the same pattern for AI specifically. Only 40% of the companies surveyed had purchased an official LLM subscription, while workers at more than 90% of them reported regular use of personal AI tools for work. Capability that already exists inside the organization goes unmanaged while budget gets allocated to new builds.

Why the integration work eats the schedule

Gardner lists three things that consume the bulk of a project: updating the process, connecting the data, and getting two teams to talk. Each one has a cost profile that a model does not.

Connecting data means reconciling schemas that were designed by different teams for different purposes, and negotiating access with whoever owns the system of record. Getting two teams to talk means someone has to change what they do daily, which requires a person with the authority to make that change stick. Updating the process means the output has to land somewhere a human already looks, because a correct answer delivered to a dashboard nobody opens produces nothing. None of this is compute-bound. All of it is calendar-bound and politics-bound, which is why it dominates the timeline.

When building custom is the right call

Her framing allows for this, and the description says so: "Sometimes it's the right solution, but explore others first." Three conditions make custom defensible.

The vendor feature genuinely does not exist. Configuring a product to do something it was never built to do produces a more fragile system than writing the thing yourself, and it leaves you exposed at every vendor release.

The workflow is your actual differentiator. If the process is how you win, handing it to a tool that ships the same capability to your competitors gives away the advantage. The MIT report supports this directly. Buyers who succeed "demand process-specific customization," and the report's own quadrant places internal builds in the high-customization column, with fragility as the failure mode rather than the concept being wrong.

The license renewal is what you are trying to escape. If the platform is the cost you want to remove, then "use the features you are paying for" argues in a circle.

The counterweight is a self-reported number from the same MIT report: external partnerships with learning-capable, customized tools reached deployment roughly 67% of the time, compared with roughly 33% for internally built tools. The report flags that these are self-reported and may not control for confounders. A 2:1 gap in deployment odds is worth pricing into a build decision even at that confidence level.

Key Takeaways

  • Sculley et al. 2015 documents that ML code is a small fraction of a production ML system, which supports the shape of Gardner's claim. The paper measures code volume rather than project effort.
  • Zylo's 2026 index, drawn from more than 40 million licenses and $75 billion in spend, puts average unused SaaS licenses at 36%. That quantifies the "use what you already pay for" advice.
  • MIT's Project NANDA attributes enterprise GenAI failure to approach and workflow integration rather than model quality, across 52 organization interviews and 153 leader surveys.
  • Custom is defensible when the vendor feature does not exist, when the workflow is your differentiator, or when escaping the license is the goal. The MIT data still shows internal builds reaching deployment at about half the rate of external partnerships.
  • The whisper transcript matches the audio with no substantive errors I could detect. The burned-in captions spell "realisation" while the transcript renders "realization," which is a captioning artifact rather than a transcription error.
  • Unverified: the 10% figure has no source and should be read as a practitioner estimate. The Adobe stack example is generic and not tied to a disclosed engagement. Gardner's identity and current title. DEPT's site bylines a Jenna Gardner as Director of Applied AI, Media, and secondary listings describe a promotion to Principal, AI Strategist, Media, but LinkedIn returned HTTP 999 and no source links either record to this TikTok account. The MIT report PDF I read is a mirrored copy rather than a file served by MIT.

Resources

Published September 16, 2026. Writeup generated from a favorited TikTok.