Six stages from business case to optimization
PPDIOO is Cisco's network lifecycle model. Each stage feeds the next, and lessons from Optimize feed back into Prepare. The meter shows the average share of each stage's tasks an LLM can do in the selected year.
Decommissioning (secure wipe, removal, inventory update) is folded into Optimize. Cisco later condensed PPDIOO to Plan → Build → Manage.
What share of each task an LLM can do
Each cell is the estimated share of the task's effort an LLM can complete in that year, with its band. The highlighted column is the selected horizon. Tasks that stay human the longest involve physical work, legal accountability or relationships.
Which roles are most exposed
Exposure is the average of the tasks each role owns. Desk-based execution roles (L1 NOC, remote implementation) cross into delegation first. Physical and relationship roles hold out longest. The Solutions Architect becomes the approver of record.
Share of work an LLM can do
Project Manager is split into planning work (WBS, schedules, RAID log, cross-PM coordination), which is quickly delegated, and people work (stakeholders, escalations, scope), which stays human longer. One PM approving LLM-generated plans can cover several times today's project load.
Converting automation into workforce shrinkage
The automation share isn't a headcount cut. People still review what the LLM does, firms adopt slowly, and demand grows when the work gets cheaper. The model accounts for all three. Change the factors or pick a scenario and the chart updates.
H = 1 − α × A × (1 − r)The part of the task the LLM doesn't do, plus the time spent reviewing the part it does.
W = (Hyear × Dyear) ÷ H2026Shrinkage = 1 − W. A negative value means the role grows.
- A · capability
- Share of the role's tasks an LLM can do (section 3).
- α · adoption
- How much of that capability firms actually use.
- r · review overhead
- Human time still spent checking LLM work, as a share of the work it took over.
- D · demand
- Volume of this work vs. today, from more networks, more complexity and cheaper services.
| Factor | 2026 (baseline) | 2031 | 2036 | 2046 |
|---|
Headcount change vs. 2026
Takeaways
Model capability isn't the main limit
Physical work depends on robotics. Sign-offs need someone with legal accountability. Letting an LLM push changes to production is a governance decision. These three keep the human share up long after the models can do the work.
Data quality sets the pace
Planning and operations automate fast only where configs, topology and documentation live in a reliable source of truth. Messy environments run 10–20 points behind these estimates.
Most of the drop comes in 2031–2036
Adoption lag and review overhead keep early shrinkage modest. The big drop comes when trust grows, so adoption rises and review time falls.
Demand is the biggest unknown
If cheaper network services lead clients to buy more of them, shrinkage is moderate and field roles grow at first. If demand stays flat, the 2046 cuts are 15–40 points deeper.
The career ladder breaks
L1 NOC and junior remote engineering are where future architects learn. Without a deliberate path, such as apprenticeships built around supervising AI agents, expect a senior-talent shortage around 2036.
Services firms need to price differently
Billable hours for design, implementation and documentation shrink first. Value moves to accountability, outcome-based and managed-service pricing, and advisory work. These are the parts hardest to delegate.
New roles appear
Agent supervisor, network automation and source-of-truth engineer, and AI governance or change-approval lead. None are in the tables, and they absorb part of the shrinkage.
Shrinkage doesn't have to mean layoffs
Much of it can be handled through attrition and fewer new hires, especially for L1 and junior roles.