Designing energy hardware for the truck roll you won't need.

The economics nobody puts on the datasheet

A distributed fleet's true cost model has a line that swallows the others: sending a technician to a rooftop, a substation or a paddock three hours from the depot. Each avoidable visit is fuel, hours, scheduling drag and a customer watching an asset not work. Serviceability isn't an aftersales virtue. It's a unit-economics decision made in CAD, months before anyone drives anywhere.

Designing the visit out

The hierarchy. First, don't fail: derate honestly, protect against the site's real conditions (heat, salt, surge, wildlife), and let simulation find the margins before the field does. Second, fail remotely recoverable: the OTA and telemetry backbone that turns a fault into a reset instead of a drive, and diagnostics honest enough that support, or the software triaging for it, knows which part to bring when a drive is unavoidable. Third, fail fast to fix: modules that swap rather than components that solder, wear parts reachable without dismantling the enclosure, fasteners chosen for a gloved hand on a ladder in wind, and spares that fit a van's shelf rather than a workshop's bench.

Design for the installer too

The first truck roll is the install, and it sets the fleet's economics: lift points where the weight actually is, cable paths that don't demand a contortionist, commissioning flows that de-skill the site work so the schedule doesn't hinge on one senior tech. Commissioning should also leave a clean record of what was installed where, on which firmware and with which settings, because every later remote diagnosis starts from that record. Watching real installs (field research in high-vis) reliably pays for itself in the first design round.

The agent-native lens: deciding whether anyone drives

When the buyer sends an assistant. A fleet operator comparing hardware may hand the maintenance comparison to an assistant, and it can only fill in the columns you publish. Make the service design legible as data: which units are field-replaceable and their part numbers, which wear parts need attention and when, what can be recovered remotely and what can't, and the documented swap procedure for each. "Low maintenance" fills no cell. The same holds after installation. A technician's assistant matching a fault code to a part before the van leaves needs every code mapped to a field-replaceable unit in current documentation it can reach, not in the memory of one senior tech.

When the business runs on agents. An AI-native service desk turns the three tiers into a triage loop. Agents watch telemetry across the fleet, line faults up against site conditions and firmware versions, and run only the remote steps people have approved in advance as safe, such as pulling health data or a controlled restart. When a visit is unavoidable, they draft it: the likely part, site access notes, nearby units due attention. A person approves every dispatch, and anything that changes how equipment behaves on the grid or near energised parts stays a human decision. The loop's last output goes back to design: faults grouped by root cause, so the next revision designs out the visits this one couldn't.

The compounding return

Serviceability compounds like interest: fewer visits per unit, times the fleet, times the years. It also sells: operators buying their second hundred units have run the maintenance spreadsheet on their first hundred, and increasingly they'll have software run it across every vendor they buy from. Hardware that publishes its service facts and reports its own health honestly gives that comparison real data to work with. This is standing doctrine in our energy & climate work, and a core argument of the engineering practice: the field is the real spec, and the truck roll is its unit of failure.