Three data uses that waited for a better answer

Training an asset-detection model only on consented imagery

In early 2023 our product team proposed a model that would detect hydrants, manholes, valve boxes, catch basins and poles from street-level imagery and mobile lidar point clouds. Customers had uploaded about 41 terabytes of such material to our hosted platform, from 214 municipalities and utilities. The model would save customers' technicians weeks of manual digitising when updating asset registers.

Our standard terms, written in 2016, allowed us to use customer data to improve our services. Marcus Feldane, our chief product officer, believed that clause covered training, and Priyanka Hollestad, the Halvercroft director on our board, wanted the feature launched in the first quarter of 2024. I disagreed. Municipal procurement officers had not read the clause as permission to build a product sold to other municipalities, and imagery contained residents, vehicles and private driveways.

I recommended a separate consent addendum: participation by choice, training in Canadian hosting regions, faces and plates blurred before any image entered the training set, and a plain statement that withdrawal stops future training but cannot reverse a released model. The chief executive accepted the recommendation after a board discussion. We gave up speed and data volume. Launch moved to November 2024, seven months later than planned.

By launch, 73 customers had signed. About forty smaller municipalities declined, most because the addendum needed a council resolution, not because they objected. The model reached 92 per cent precision for hydrants and manholes but performed worse on rural gravel roads, where our consenting set was thin. By June 2026, 126 customers had consented, 58 were paying for the feature and it contributed C$2.7 million of annual recurring revenue.

The consent register is now a standard part of our model pipeline, and every model version lists its contributing customers. I would do one thing differently: prepare a short briefing note for councils at the outset. Many small municipalities wanted to participate, and the administrative burden excluded exactly the rural imagery the model most needed.

Pausing a records integration after an acquisition revealed shifted positions

In 2022 we acquired Dunsgarrow Locate Records, a Winnipeg product that utilities use to record where buried cables and pipes have been marked before excavation. Our sales team had promised an Alberta natural-gas co-operative a single asset view combining Dunsgarrow records with its Pellarwick network map by mid-2023.

Early testing showed positions disagreeing by between 0.4 and 2.1 metres for some older records. Dunsgarrow stored a mix of original datums and applied a transformation only when displaying the map; our platform stored positions in a current Canadian reference frame. The former Dunsgarrow founder, by then our general manager in Winnipeg, considered the display transformation adequate. Our chief revenue officer wanted the integration released on schedule with a caveat. For buried assets, a metre can decide where a crew digs.

I recommended pausing the release for two quarters. Every migrated record would carry its source datum, the transformation applied and a confidence class. Records whose lineage we could not establish would be labelled unverified rather than moved to a best estimate. The chief executive agreed. We deferred about C$1.6 million of expected first-year integration revenue, gave the co-operative a C$240,000 service credit and spent roughly fourteen engineer-months on the work. Responsibility for excavation decisions stayed with the co-operative and its locators throughout.

We migrated 3.1 million records by December 2023. About 412,000, or 13 per cent, were labelled unverified. Over the next eighteen months customers re-verified 280,000 of them through routine field locates, and the co-operative renewed its contract in 2025. The unverified label initially generated complaints from customers who preferred a clean map, and two smaller utilities delayed adoption.

Our acquisition process changed. When we bought Ambrevick Stormwater in 2023, a data-lineage review of sample customer records was part of due diligence before signing. I would also have told the co-operative about the discrepancy in the first week of testing, rather than after we had settled on a remedy.

Withdrawing precise locations from a benchmark product

In 2025 our commercial team proposed a pipe-condition benchmark built from the closed-circuit inspection grades that municipalities record in our platform. The first design showed condition by pipe segment with coordinates. Engineering consultancies and two insurers had expressed interest, and the forecast was C$3.4 million of annual revenue by 2027.

My concerns were threefold. Municipalities had licensed inspection data to us for their own asset planning, not for sale. Segment-level condition with location could reveal vulnerable infrastructure and, near private connections, individual properties. And insurers could use the data to price cover for the very municipalities that supplied it, without their knowledge. Our chief revenue officer and Priyanka Hollestad argued that the data was anonymised because municipal names were removed. I pointed out that a segment's coordinates identify its owner immediately.

I proposed a narrower product. We would publish benchmarks only by pipe material, installation decade and broad region, with at least twelve contributing municipalities behind every figure. Utility network data would be excluded entirely. Municipalities would contribute only by opting in, and contributors would receive the benchmark without charge. Insurers would not be offered the product. The chief executive took the question to the board, which supported the narrower version.

We also found that a demonstration file with street names had been shared with one insurer under a non-disclosure agreement before the decision. We requested its deletion and received written confirmation, and I reported the matter to the audit and risk committee.

The product launched in March 2026 with 64 contributing municipalities. The first-year revenue forecast fell to C$1.1 million, and by September nine consultancies had subscribed, worth C$380,000 a year. Contributing municipalities have used the benchmark in their own capital plans.

I would involve customers' privacy officers before the commercial team begins external conversations. The demonstration file existed because the data question reached me after a prospect had already been shown something attractive.