An agriculture technology company built a crop yield prediction model that combined satellite imagery, soil sensor data, weather forecasts, and historical yield records. The model predicted per-field yields with an accuracy that exceeded the agronomists who had been advising farmers for decades. In validation against three years of historical data, the model’s yield predictions were within eight percent of actual yields, compared to the agronomists’ average error of fourteen percent.

The agronomists did not trust it.

The model was deployed to support agronomists in making fertilizer application recommendations. An accurate yield prediction allowed the agronomist to calculate the right fertilizer rate — too little and the crop underperformed, too much and the farmer overpaid while creating environmental runoff. The model’s recommendation was displayed alongside the agronomist’s own recommendation. The agronomist chose which to follow.

For two growing seasons, the agronomists overrode the model’s recommendation seventy-eight percent of the time. They used their own judgment, which was less accurate, because they did not trust a system that could not explain its reasoning in terms they recognized.

The expertise mismatch

Agronomists make recommendations based on observable crop indicators: leaf color, plant height, root development, pest pressure, and soil moisture measured by hand. These indicators are concrete, visible, and verifiable. An agronomist can walk a field, observe that the corn is showing nitrogen deficiency in the lower leaves, and recommend a sidedress application. The reasoning chain is short, observable, and grounded in decades of field experience.

The model’s reasoning chain was different. The model correlated satellite-derived vegetation indices with historical yield outcomes, adjusted for weather patterns and soil composition. The model could not walk a field. The model could not see nitrogen deficiency in lower leaves. The model saw spectral reflectance patterns that correlated with yield outcomes at a population level.

Both approaches were valid. They operated on different evidence and at different scales. The agronomist’s evidence was local and observable. The model’s evidence was statistical and population-level. When the two disagreed — which happened frequently — the agronomist defaulted to their own evidence because they could verify it with their own senses.

The trust problem was not about accuracy. The agronomists acknowledged that the model was more accurate in validation. The trust problem was about evidence type. The agronomists trusted evidence they could observe. The model’s evidence was invisible to them.

The failed approach: accuracy arguments

The data science team’s first instinct was to present the accuracy comparison. They showed the agronomists the validation results: the model’s eight percent error versus the agronomists’ fourteen percent error. They showed that in 217 out of 300 validation fields, the model’s prediction was closer to the actual yield than the agronomist’s prediction.

The accuracy argument backfired. The agronomists interpreted it as an attack on their expertise. Some pointed out that the model had access to data they did not — satellite imagery processed by algorithms they could not inspect — and that the comparison was therefore unfair. Others pointed out that the validation was retrospective and that the model’s performance in novel conditions — a new pest, an unusual weather pattern, a soil anomaly — was unknown.

Both objections were legitimate. The accuracy comparison was retrospective, and the model’s behavior in novel conditions was uncertain. But the deeper issue was psychological. Telling experts that a model is more accurate than they are does not build trust. It builds resentment. The experts disengage or actively resist, not because they are irrational, but because the framing positions the model as a replacement rather than a tool.

The approach: augment, not replace

We redesigned the interaction model so that the model’s output was presented as additional evidence for the agronomist’s decision, not as a recommendation to accept or override.

The combined evidence view showed the agronomist four things: the model’s yield estimate with its confidence interval, the satellite-derived vegetation index map for the field, the agronomist’s own field observations from their most recent visit, and the soil test results. The agronomist weighed all four sources of evidence and made a recommendation.

The critical design decision was that the model’s output was positioned as one evidence source among several, not as a recommendation. The agronomist was not asked “do you agree with the model’s recommendation.” The agronomist was asked “given the model’s yield estimate, your field observations, and the soil tests, what fertilizer rate do you recommend.” This framing preserved the agronomist’s agency and expertise while making the model’s evidence available for consideration.

Over time, agronomists started using the model’s evidence to sharpen their own recommendations. An agronomist who observed good crop conditions but saw a declining vegetation index trend in the satellite data would investigate whether there was a localized stress factor they had missed during their field visit. The model’s population-level evidence complemented the agronomist’s local-level evidence. Neither was sufficient alone.

The calibration partnership

We also built a feedback mechanism that allowed agronomists to flag when the model’s yield estimate seemed inconsistent with their field observations. These flags were not overrides — they were calibration signals. When an agronomist flagged a prediction, the data science team investigated whether the model was missing a feature that the agronomist’s observation implied.

In the first season, agronomists flagged forty-three predictions. Investigation revealed that the model was systematically overestimating yields for fields with compacted soils, because the soil sensor data did not capture compaction — it captured moisture and nutrient levels, which were adequate even in compacted soil. The model saw healthy soil chemistry. The agronomist saw poor root development. The agronomist was right, and the model was missing a feature.

The data science team added a compaction indicator derived from historical tillage records and soil type classifications. The model’s accuracy for compacted fields improved from a twelve percent error to a seven percent error. More importantly, the agronomists who had flagged the compagation issue saw their feedback directly improve the model. This was the turning point for trust.

What we gave up

The evidence-based interaction model was slower than a direct recommendation. An agronomist reviewing the combined evidence view spent an average of six minutes per field, compared to two minutes for a direct recommendation. At 200 fields per week, this was an additional thirteen hours of agronomist time.

The team accepted this because the agronomists were making better recommendations. Fertilizer application accuracy improved in the second growing season, with the combined approach producing yield outcomes within six percent of optimal, compared to eight percent for the model alone and fourteen percent for the agronomist alone. The combination was more accurate than either source of evidence independently.

The second trade-off was model development velocity. The calibration partnership meant that the data science team spent approximately twenty percent of their time investigating agronomist flags and integrating feedback. This slowed feature development but produced a model that the agronomists trusted, which was a precondition for adoption.

Results

In the third growing season — the first after the interaction model redesign — agronomists used the model’s evidence in eighty-three percent of their recommendations. Override rate dropped from seventy-eight percent to nineteen percent. The overrides that remained were concentrated in fields with known data gaps — sensor failures, satellite image cloud cover, or historical records older than five years.

Fertilizer cost savings reached twenty-two percent across the customer base, exceeding the eighteen percent that the model alone could have achieved if it had been trusted from the start. The additional four percent came from the agronomists’ local knowledge correcting the model’s population-level blind spots.

The decision heuristic

When domain experts reject a more accurate model, the problem is not their irrationality. The problem is that the model’s evidence is invisible to them. Do not present the model’s conclusion and ask them to agree or disagree. Present the model’s evidence alongside their own evidence and let them synthesize. Position the model as an additional data source, not as a recommendation engine. And when their expert observation reveals a gap in the model’s features, fix the gap. The fastest path to trust is demonstrating that the expert’s knowledge makes the model better, not that the model makes the expert obsolete.