The Upstream River Parable: Origin and What It Means for AI Automation
By Matt Cullerton, founder of Mavric and Apso
The river parable, in which someone keeps pulling problems out of a river until they walk upstream to find the source, was published by John B. McKinlay in 1975, crediting medical sociologist Irving Zola. A TikTok from @jenna_gardner_ai retells it for AI workflow builders: stop automating the cleanup downstream and fix the source upstream. The advice is sound, and it has named methods behind it, but there are also clear cases where building the downstream filter is the right call.
Where does the upstream river parable come from?
McKinlay published it in "A Case for Refocusing Upstream: The Political Economy of Illness" (1975), arguing that health systems fixate on downstream problems that are short-term and specific. His version begins: "There I am standing by the shore of a swiftly flowing river and I hear the cry of a drowning man ... who the hell is upstream pushing them all in."
The story is older than the paper. Terence Eden traced a 1971 version used by Dr. Reginald Lourie in congressional testimony, and Zola's own telling appears to come from a 1970 speech. It is often attributed online to Desmond Tutu, with no evidence.
The business version is Dan Heath's 2020 book "Upstream: The Quest to Solve Problems Before They Happen." His term "problem blindness," the belief that a recurring problem is just how things are, matches the line in Gardner's story: "I don't know. It's always been there."
What methods help you find the upstream cause?
The video gives the instinct without a method. Several established ones exist:
- Five Whys. Developed by Sakichi Toyoda and used in Toyota Production System training. Taiichi Ohno described asking "why" five times as the basis of Toyota's approach. Former Toyota managing director Teruyuki Minoura called it too basic for real root cause work, and Alan J. Card has argued the fifth "why" rarely reaches an actual root cause.
- Ishikawa (fishbone) diagrams. Kaoru Ishikawa popularized them in the 1960s at the Kawasaki shipyards. Because they branch, they surface several contributing causes instead of one chain.
- Jidoka and the andon cord. Toyota stops the line the moment something goes wrong, so a defect never moves downstream.
- Poka-yoke. Shigeo Shingo formalized mistake-proofing in the 1960s. In software, the equivalent is a database constraint, a typed schema or a required field.
Software has its own version. Larry Smith coined "shift-left testing" in a September 2001 Dr. Dobb's Journal article, arguing for moving testing into design and coding.
An example of fixing the source: Expedia
In 2012, Ryan O'Neill, head of Expedia's customer experience group, found that 58 of every 100 customers who booked travel called afterward for help. The top reason was getting a copy of their itinerary, about 20 million calls a year. At roughly $5 a call, that was a $100 million problem.
The causes were mistyped email addresses, spam filters and accidentally deleted confirmation emails. Expedia added an automated resend option to the phone system, changed how emails were sent so they cleared spam filters, and built a self-service portal. The call rate fell from 58 percent to about 15 percent.
The downstream version, a better call center or an AI voice agent answering "where is my itinerary," would have left millions of unnecessary calls in place.
Where AI automation lands downstream
The same pattern shows up in enterprise AI projects:
- An LLM pipeline that parses vendor invoices from unstructured PDFs, when a supplier portal with typed fields would remove the need.
- Training a classifier to re-route misfiled support tickets, when the ticket form offers 40 ambiguous categories.
- A nightly reconciliation agent that catches data mismatches, when a validation rule on the original form would prevent them.
- Summarizing 60-page status documents with an agent, when a reporting standard would shorten them.
Each downstream build is easier to fund. It touches one team, needs no negotiation and shows a metric within a quarter.
When is cleaning up downstream the right call?
Gardner's story has one polluter, no defenders and a free fix. Real problems are messier.
Ford's 8D problem-solving method, developed in the 1980s, separates D3, an interim containment action, from D5, the permanent corrective action. Containment protects the customer while root cause work continues. Building the filter is D3. The mistake is treating D3 as finished.
Downstream cleanup makes sense when:
- You do not control the source. The bad data comes from a customer's ERP, a regulator's filing format or another division's system.
- The fix upstream takes 18 months and the damage is happening now. Ship the filter, and give the upstream fix a date, an owner and a plan to retire the filter.
- The source is a paying customer. If messy spreadsheets come from 30 percent of revenue, absorbing the mess is part of the product.
- Many sources contribute. A few defects from each of many systems is the usual case, and an Ishikawa diagram shows that spread better than Five Whys.
- Upstream is acting rationally. Sales may skip CRM fields because the form costs nine minutes per lead and they are measured on volume. Fixing that means changing their incentives.
The quick test is to ask who owns the source, what it would cost them to stop, and what you are measured on. If no one you can reach owns it, build the filter and label it containment. If one person could fix it in two weeks, have that conversation first.
Bottom line
The river parable comes from McKinlay and Zola, not Desmond Tutu, and its lesson applies to AI automation: automating cleanup of a broken process keeps the process broken. Use Five Whys, fishbone diagrams or poka-yoke to find and remove the source when you can. When you cannot, treat the automated filter as temporary containment with a plan to replace it.
Sources
- Reclaiming public health's defining metaphor (PLOS Global Public Health): the Zola and McKinlay origin.
- Terence Eden on the parable's origin: the 1971 Lourie version and the Tutu misattribution.
- How Upstream Thinking Saved Expedia Millions (Qualtrics): the Expedia case.
- Five whys (Wikipedia): origins and criticism.
- Ishikawa diagram (Wikipedia): fishbone diagram origin.
- Autonomation (Wikipedia): jidoka and the andon cord.
- Eight disciplines problem solving (Wikipedia): Ford's 8D method.
- Shift-left testing (Wikipedia): the software version of upstream thinking.
- Shift-Left Testing, Dr. Dobb's Journal, September 2001: Larry Smith's original article.
- Business @ the Speed of Thought (Wikipedia): Bill Gates's 1999 book.
- @jenna_gardner_ai on Instagram: the creator's account.
Published September 19, 2026. Based on a TikTok by @jenna_gardner_ai. Full video notes (markdown)