From Harvard Business Review · José Parra Moyano · · 4m
The gap between AI prototypes and real tools
Building software with AI is fast. Making it work at scale is hard. Here's how companies that succeeded did it.
In brief
Building software with AI is fast. Making it work at scale is hard. Here's how companies that succeeded did it. Companies building their own software with generative AI often struggle to move prototypes into production. Three practices close the gap: involve engineers early, limit system connections, and resist scope…
AI makes building software feel almost free
Generative AI lets product managers describe what they want and get working tools in an afternoon.
The deployment gap
Building a working prototype is one thing.Turning it into a reliable solution that works in everyday operations is another.
A European bank's pilot success
A Swiss AI startup deployed seven coordinated agents to automate administrative work for relationship managers at a major European bank.Client-facing time increased by more than one hour per day per relationship manager.
The bank tries to scale internally
Bank leadership asked why they couldn't hire a team, give them the same AI tools, and scale the solution themselves.
The vendor rebuilds from scratch
The bank re-engaged the vendor, who determined it would take more time to fix what the bank had built than to build out their original pilot from scratch.
Research across 11 companies
Over two years, researchers studied how organizations that choose to build their own software solutions meet the deployment challenge.
Three steps to close the gap
Successful companies follow three steps.First, before the build, they discuss which parts a non-technical person can build and where experts should weigh in.Second, they avoid tightly wiring the solution into other core systems.
Myopia: the blind spot
Generative AI lets domain experts build solutions by handling decisions about structure.The human supplies intent and the AI determines architecture.
When updates become impossible
When a solution must be updated to handle an edge case or evolve, the people who created it do not know what elements of the architecture to fix because they never learned how it was…
Engage engineers early
Non-technical teams should not avoid using AI to build, but should engage software engineers early to anticipate true complexity.
Roche's review process
At Roche, the Swiss pharmaceutical company, all 18 members of one data team use generative AI to build dashboards and automate analysis.
Dependency: keep connections few
Every link between a new solution and the rest of the company creates an obligation.Each data feed, integration, and downstream process that depends on the tool is something that breaks when the tool changes.
A food company's video pipeline
A food and beverage company headquartered in northern Europe brought in-house the video advertising it used to commission from outside agencies.
Easy to update when connections are few
Because the food company's pipeline stops before publishing systems or social media feeds, it is easy to update and improve.
Scope: when functionalities widen risks
After a successful pilot, the temptation is to push it everywhere: more teams, more regions, more use cases.
A packaging company's failed expansion
An international packaging company built a tool to schedule warehouse shifts.It worked so well that management asked it to also handle payroll exceptions, union rules, and expand to three more locations.
Scale the approach, not the solution
Rather than expanding a successful solution to serve more users and purposes, companies can use it as a model for other teams to build narrowly scoped solutions to problems they themselves own.
Governing the zoo of AI solutions
A large consumer-goods multinational draws a firm boundary between tools built for a team's own use and those intended to travel across the organization.Individuals and teams are encouraged to build solutions to their own problems.
At a glance
Companies building their own software with generative AI often struggle to move prototypes into production. Three practices close the gap: involve engineers early, limit system connections, and resist scope creep.
