Electricity didn’t change the economy by being installed in factories. It changed the economy when factories were redesigned around it. AI is at the same inflection — and most mid-market companies are still wiring their old factory layout.
The historical pattern
When electricity first arrived in factories at the turn of the 20th century, productivity gains were modest. Factories had been designed around steam power: a single central source, with mechanical shafts and belts distributing power throughout the building. When electricity was installed, factories used it the same way — a central generator, mechanical distribution, same factory layout.
It took 30 years for productivity to surge. As economist Paul David documented in his 1990 study of the productivity paradox, factories ran on “group drive” — one big motor turning the same shafts and belts the steam engine used to — from the mid-1890s until around 1920. The breakthrough came when factories were redesigned around electricity’s actual properties: “unit drive,” a separate motor on each machine, decentralized power at each workstation, layouts optimized for workflow instead of mechanical distribution. Only then did the gains arrive — electrification accounted for roughly half of manufacturing productivity growth in the 1920s, decades after the first power stations came online.
The technology didn’t change. The operating model did.
The same pattern with AI
Almost every mid-market company is in the equivalent of the “electricity in a steam-era factory” phase. AI tools have been installed; the factory hasn’t been redesigned.
Some signs you’re in this phase:
- People in your business use AI as individual productivity tools, but no workflow has been redesigned around AI
- The team uses AI for “summarize this” and “draft that” but no specific business outcome has shifted
- License costs are real; license utilization is high; business impact is invisible
- The CFO can’t point to a number that changed because of AI
You’re not alone. MIT’s NANDA initiative studied 300 deployments and found only about 5% of AI pilots achieve rapid revenue acceleration — the vast majority stall with no measurable hit to the P&L. High adoption, invisible impact. That’s the signature of AI installed in a factory nobody redesigned.
This is normal. It’s also the point where the historical pattern says the operating model has to evolve, or the technology won’t deliver.
What the redesigned operating model looks like
A few specific changes:
Workflows designed around AI properties. AI is good at: structured extraction from unstructured input, drafting from context, classifying and routing, summarizing across volume. Workflows that exploit those properties — quote drafting, document processing, customer service triage, knowledge capture — get materially faster. Workflows that don’t get incremental improvement at best.
Decisions about what humans should keep doing. In a redesigned operating model, humans keep the judgment, the relationship work, and the edge cases. AI takes the routine, the volume, the synthesis. Specific human roles change — but the humans stay.
Measurement keyed to workflow KPIs, not tool usage. “Quote turnaround dropped from 5 days to 24 hours” instead of “120 people are using a chat tool weekly.”
Operating cadence built around AI’s continuous improvement. AI workflows get better with use. The operating cadence has to include feedback loops — captured corrections, refined prompts, retrained context. That’s a new operational muscle most companies don’t have yet.
What’s not the operating model change
A few things companies do that feel like operating model change but aren’t:
- Hiring a Chief AI Officer. Helps in some cases, doesn’t change anything in others. Title without redesigned workflows is theater.
- Standing up an AI committee. Almost always slows things down rather than speeding them up. Committees don’t redesign factories.
- Buying enterprise AI software. Buying a platform doesn’t redesign the workflow on top of it.
- Sending the team to AI training. Useful, but workflows are still designed the old way after training ends.
The operating model change requires the workflows themselves to be redesigned. Tools, training, and titles are downstream.
What this means for the next 24 months
The historical analogue suggests we’re roughly 18 months into the “AI installed in steam-era factory” phase. The factories that get redesigned in 2026–2027 will pull ahead structurally — the way the early-adopter factories of the 1920s pulled ahead of the laggards.
Three implications for mid-market operators:
-
The window for getting the redesign right is now, not later. Companies that wait for “stability” will be redesigning while their competitors are scaling.
-
The redesign work is operations work, not IT work. It belongs to the COO, the operations director, the GM — the people who own the workflows that need redesign. RAND, studying why more than 80% of AI projects fail — twice the failure rate of non-AI IT projects — traced the root causes to misunderstood intent and bad data, not weak models. The failures live in the operating model, not the technology.
-
The first redesign is the hardest. Subsequent ones compound. The companies that ship one workflow redesign in 2026 will ship 5 in 2027. The companies that don’t will still be debating the strategy.
Where to start
If you’re sitting in the “AI installed, factory not redesigned” zone, the question isn’t whether to redesign — it’s which workflow to redesign first. That’s the conversation worth having: pick the one workflow where AI’s actual properties line up with a real business outcome, and rebuild it end to end. The first one teaches you everything you need to scale the next five.