Your teams are bringing AI proposals into next year's plan, and most of them will look impressive. The decision in front of you is which ones to fund. My test is one question: is this the simplest thing that does the job, or is it a Rube Goldberg machine with a better press release?

Companies have become fond of proclaiming an “AI first” approach or philosophy in all aspects of their business. I think this is leading a lot of life science companies astray. It starts with the tool and goes looking for a problem. The standard I'd hold every proposal to is human intelligence first. Start with the judgment of the people who own the outcome, define what fit for purpose means, and add technology only where it changes the result.

Rube Goldberg finished his engineering degree at Berkeley in 1904, then left engineering for cartooning and later won a Pulitzer Prize [1]. His best-known series, The Inventions of Professor Lucifer Gorgonzola Butts, ran in Collier's Weekly from 1929 to 1931 [2]. One of those inventions used a parrot, a pail of seeds, a skyrocket and a pendulum to wipe a man's chin with a napkin [3]. Goldberg called the inventions a "symbol of man's capacity for exerting maximum effort to accomplish minimal results" [4]. Adam Gopnik put it more plainly: these are things that need doing, but they don't need this much doing [5].

The cartoons were a tongue-in-cheek critique of technology that promised to simplify life and often did the opposite [6]. Look closely and every machine depends on a living thing doing its part: a parrot that has to want the seed, a person who has to react on cue [7]. The gears get the credit. The chain only works because of one very human step that nobody engineered.

Goldberg was mocking overbuilt technology almost a century ago, so the habit isn't new with AI. The pattern even has a name now. In 1995, Gartner analyst Jackie Fenn introduced the Hype Cycle, which tracks how expectations for a new technology climb to a peak, fall into disappointment and only later settle into everyday use [8]. AI first mandates tend to get written near the peak.

AI first programs repeat the same flaw. The tools get the budget and the attention, while the people whose judgment makes the output usable get treated as a given. That gap stays invisible until something goes wrong. Then someone has to explain how an expensive chain of tools produced a result nobody can defend, and why no one owned the step that should have caught it.

Gartner research shows that by the end of 2025, more than half of generative AI projects were abandoned after proof of concept, for reasons that included poor data quality, escalating costs and unclear business value [9]. In Gartner's 2027 CIO and Technology Executive Agenda survey, only 13% of CIOs said their enterprise had realized significant or transformational value from AI so far [9]. Gartner also expects more than 40% of agentic AI projects to be canceled by the end of 2027 [10].

These are cross-industry figures. I'd be careful reading them as a life science failure rate. The diagnosis behind them is more useful than the percentages. Kristin Moyer, a Gartner vice president and analyst, said, "The biggest drag on AI momentum is using AI when it isn't needed" [9]. That line could caption most of Goldberg's cartoons.

An AI first mandate tells teams the answer before anyone has defined the question. Once AI is the expected answer, teams get rewarded for using it, whether or not it solves the problem. A team that fixes a problem with a better template and a clear owner has nothing to announce. A team that builds a multi-agent workflow with three integrations has a demo, a line in the board materials and a story about being an AI-enabled organization.

I understand the pull. Boards ask about AI, as do investors and the talented people you want to keep. Saying you use AI across the enterprise feels like a strategy. It's an activity count.

I've seen this pattern up close. Early in one engagement, I led a working session on the groundwork an AI-enabled workflow would need: whether the data was ready, how outputs could be audited, and how the pilot's results would tie back to the business outcome it was meant to improve. The team was eager to get to the AI part. Afterward, one of the implementation leads asked me how valuable I thought the session had been. Before I could answer, he made it clear he thought it had none. He pushed ahead with a separate pilot. It stalled as one unplanned problem after another surfaced, including gaps in the groundwork the team had skipped, and it was eventually folded into another initiative. The simpler path would have been slower to start and faster to finish.

In the life sciences, complexity carries a cost other industries don't pay in the same way. Software used in GxP-regulated work has to be validated for its intended use, documented and kept under review. For device makers, FDA's computer software assurance guidance scales that effort to risk [11]. In Europe, Annex 11 covers any computerised system used in GMP work, and a longer, more prescriptive revision has been in draft since July 2025 [12]. A chain of tools feeding drafts into MLR review adds steps to a process whose bottleneck is usually human judgment in the first place. Each added part is one more place for the chain to break.

Human intelligence first still leaves room for AI. It changes the order of the inquiry. You start with the people who own the outcome and what they already know: the purpose, what good looks like, where the constraint actually sits and what a capable person would do with simple tools. Then you ask whether a machine improves that answer enough to justify what it costs to run and defend.

Sometimes the answer is a sophisticated model. In discovery, imaging or safety signal detection, the complexity may be the point. Often, the answer is a cleaner data feed, a decision right that finally has an owner, or one well-scoped tool. I made a version of this argument in “Fit for Purpose,” where I wrote that chasing the gold standard often slows strategy and that fit for purpose is frequently the better bar [13]. AI raises the stakes on that argument because the cool factor is so much stronger.

When an AI proposal reaches you, I'd ask four questions before the funding decision.

Each answer carries a trade-off. The simple option is cheaper and faster to prove, and it can look unambitious to a board that wants an AI story. It may also need rework if volume grows tenfold. The complex option builds capability your teams might need later, at the price of a longer path to value and more ways to fail. I'd still default to simple and make the complex case earn its way in.

For medical affairs leaders, the test hits hardest on scientific content and medical information. If your constraint is reviewer capacity or evidence quality, an AI layer that generates more drafts can lengthen the queue without improving a single answer. I'd want to know exactly where the human judgment sits before anyone adds a step in front of it.

For commercial and market access leaders, the pressure shows up in launch and field tools. An HCP engagement engine with five data sources and a next-best-action model sounds like a launch advantage. If the problem is a value story your field teams can't tell clearly to payers or physicians, the machine is tuning the wrong thing. Start with what your best people already know about why physicians prescribe or hold back.

This standard isn't free or without risk. Choosing the simplest fit for purpose answer means telling capable teams no. It means a thinner AI narrative in the next investor update. It can mean letting a competitor build a capability first, and occasionally that'll be the wrong call. Getting it right takes judgment about which challenges deserve the simple answer and which justify the complex one. Applied well, that judgment keeps the downside small and the return large.

A company that funds ten Rube Goldberg machines learns little from most of them and spends years maintaining all ten. A company that funds three fit for purpose bets, each owned by someone who can explain how the business was impacted, builds the judgment to know when complexity is worth paying for.

When next year's plan reaches your desk, count the owners before you count the AI initiatives. Every project worth funding has someone who can explain how the business was impacted by it. If nobody can, the machine is decoration.

Facing a consequential strategy decision? Bring it to Talon Group Consulting.

Our Hidden Readiness Gap study is open through December 6. Take part and get your answers benchmarked against peers.

1. Quintero, Fernando. "Fantastic Inventions." Berkeleyan, University of California, Berkeley, 6 Mar. 2002, newsarchive.berkeley.edu/news/berkeleyan/2002/03/06_gold.html.

2. Tumey, Paul. "Rube Goldberg Butts In." The Comics Journal, 24 Feb. 2014, tcj.com/rube-goldberg-butts-in.

3. VanHemert, Kyle. "Rube Goldberg: The Man Behind the Machines." Gizmodo, 26 Aug. 2010, gizmodo.com/rube-goldberg-the-man-behind-the-machines-5622673.

4. Daley, Jason. "Queens Museum Brings Rube Goldberg Machine to Life." Smithsonian Magazine, 8 Oct. 2019, smithsonianmag.com/smart-news/queens-museum-brings-rube-goldberg-machine-life-180973297.

5. "New York Times Article 'Rube Goldberg's Greatest Machine Is His Legacy.'" Rube Goldberg Institute, rubegoldberg.org/about-the-institute/new-york-times-article-rube-goldbergs-greatest-machine-is-his-legacy.

6. Wilson, Emily. "The Story Behind Rube Goldberg's Complicated Contraptions." Smithsonian Magazine, 1 May 2018, smithsonianmag.com/history/story-behind-rube-goldbergs-complicated-contraptions-180968928.

7. Crimmins, Peter. "It's Complicated: Life and Work of Rube Goldberg at Philly's Jewish Museum." WHYY, 11 Oct. 2018, whyy.org/articles/its-complicated-life-and-work-of-rube-goldberg-at-phillys-jewish-museum.

8. Pollock, Neil, and Robin Williams. "Managed Channels in the Wild Sea of Hype." After Hype: The Business of Taming the Digital Economy, Cambridge University Press, 2026, cambridge.org/core/books/after-hype/managed-channels-in-the-wild-sea-of-hype/250607FD67AF719DA9C15AE420524765.

9. Amsler, Sarah. "Chasing AI ROI? Here's Why It Keeps Slipping Away." TechTarget, 29 Oct. 2026, techtarget.com/news/chasing-ai-roi-heres-why-it-keeps-slipping-away/873195.

10. "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027." Gartner, 25 June 2025, gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027.

11. "Computer Software Assurance for Production and Quality Management System Software." U.S. Food and Drug Administration, Feb. 2026, fda.gov/regulatory-information/search-fda-guidance-documents/computer-software-assurance-production-and-quality-management-system-software.

12. Henderson, Trevor J. "EU GMP Annex 11 and Computerised Systems." Lab Manager, 2026, labmanager.com/eu-gmp-annex-11-and-computerised-systems-35922.

13. Sharma, Aditi. "Fit for Purpose." Talon Catalyst, 8 Sept. 2025, newsletter.talongroup.consulting/p/fit-for-purpose.