What problem does it solve?
Building a fibre network means choosing, street by street, where cable goes: whether to reuse an existing duct, dig a new trench, or string cable along poles, and whether to route around a river, a protected tree or an area where paving would be expensive to dig up and restore. Telefónica Spain describes the traditional process as complex, with many variables that were difficult to predict, and says conventional tools were not always able to identify physical obstacles such as rivers or hard to reach areas, which increased execution costs.
Deutsche Telekom makes the trade off explicit: "the shortest route to the customer is not always the most economical." Paving stones take longer to dig up and restore than a dirt road, and effort and cost depend on what is already there; roots near a tree change what a crew can safely do. Working this out by hand, area by area, from survey notes and a planner's experience, does not scale to a national rollout that has to pass millions of homes on a tight timeline, and inconsistent estimates lead to budget overruns and rework once construction starts.
How does it work?
- Collect the ground truth. A survey vehicle or aerial and satellite imagery captures the terrain: 360 degree panoramas, laser scan point clouds and existing duct, pole and cabinet locations, referenced to GPS coordinates.
- Classify what is in the data. A computer vision model trained on the imagery and point clouds recognizes surfaces and obstacles (asphalt, cobblestones, gravel, trees, buildings), so the planning step does not start from raw pixels.
- Generate and cost candidate routes. An optimization model proposes one or more routes between the network and the homes to connect, reusing existing infrastructure where possible and minimizing new civil works, and estimates the cost from the route's characteristics and the cost of comparable past projects in similar terrain.
- Rank against the rollout plan. Where many areas compete for the same construction crews and budget, the system can also weigh which routes bring homes online soonest for the investment.
- A planner reviews and approves. The proposed route and cost estimate go to a human planner, who checks it against local knowledge, adjusts where needed and approves it before it becomes a construction work order.
- Audience
- Employee facing
- Autonomy
- Copilot
- Adoption
- Emerging
- Channels
- Internal tools, API and system to system
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Cycle time reduction | Too few to pool | about 75% | 1 | 1 vendor |
Value drivers: Speed and cycle time, Employee productivity, Risk and loss reduction.
Indicative value
A fibre operator planning 500,000 new homes passed a year
USD 240,000 to USD 1.7 million
Planner time cost avoided per year
How this is calculated
Formula: (homesPerYear / 1000) * plannerHoursPerThousand * timeSaved * costPerPlannerHour. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| New homes passed planned per year homesPerYear, homes per year | 500,000 | 500,000 | The reference operator. |
| Planner hours to design routes for 1,000 homes, without AI assistance plannerHoursPerThousand, planner hours per 1,000 homes | 40 | 80 | Editorial assumption, replace with your own time studies. Steep's case study for Deutsche Telekom says planning and approval processes used to take several months. |
| Share of planning and approval time the assistance saves timeSaved, fraction of planning time | 0.3 | 0.6 | Conservative against the deployment evidence on this page (Deutsche Telekom's technology partner reports planning and approval about 75% faster once its geospatial data pipeline was in place), because that figure covers the whole data platform at scale, not automated route determination on its own, and this range should also fit an early rollout. |
| Fully loaded cost of a network planning engineer costPerPlannerHour, USD per hour | 40 | 70 | Editorial assumption for a mid to senior network planning role. Replace with your own cost. |
What it leaves out: Gross planner time saved only. It leaves out the cost of building and running the AI pipeline (survey vehicles or imagery, model training, integration with GIS and costing systems), any change in civil works cost accuracy, and the value of passing homes sooner.
Who already uses it?
3 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Telefónica España
Spain · Telecommunications · 2025
Telefónica Spain built an AI based on genetic algorithms that designs fibre deployment routes from geographical coordinates, estimating cost and reusing existing infrastructure where possible. The company says the tool is already being used by its planning teams and has reached what it calls level 4 autonomy in the fibre planning process, under its Autonomous Network Journey program. No quantified time or cost saving is published.
No outcome disclosed.
Openreach
United Kingdom · Telecommunications · 2026
Openreach has built a digital replica of the UK's transportation corridors on Google Cloud's Vertex AI, combining data on 35 million homes and businesses with road, rail and waterway networks and its existing broadband infrastructure, so planners can see where full fibre can be extended soonest. The same announcement names a reduction of upwards of 50% in "time to insight" from using Gemini Enterprise to convert legacy data queries, but that figure is about cloud engineering work, not the route planning itself, so it is not recorded as a metric for this use case. Openreach's managing director for fibre first, James Tappenden, said the collaboration brings practical, measurable benefits, from "connecting more families to gigabit broadband faster" to cutting vehicle emissions across its workforce.
No outcome disclosed.
Deutsche Telekom
Germany · Telecommunications · 2018
Deutsche Telekom's own newsroom describes a pilot, first run in Bornheim near Bonn, in which a measuring vehicle with 360 degree cameras and laser scanners collects environmental data for its FTTH rollout, and software developed with Fraunhofer IPM automatically classifies around 30 categories of surface and obstacle (pavement type, trees, root structure) to determine the optimal cable route, which a Deutsche Telekom planner then checks and approves. A separate account from Fraunhofer IGD's Steep workflow platform, which has processed this imagery and point cloud data at scale in Deutsche Telekom's cloud since the collaboration began in 2017, says the program has since planned more than 10 million households and connected more than 8 million of them to FTTH, and that planning and approval work which used to take several months is now finished within a few weeks, on average about 75% faster. Steep credits that speedup to the whole geospatial data infrastructure it built for Deutsche Telekom, including large scale point cloud classification and the Fibre3D tool that lets planners present plans to municipalities with fewer site visits, not to automated AI route determination alone: Steep says the classification output "can later be used to automatically determine optimal routes", describing that as a future use rather than a capability it says is running at this scale today.
- Cycle time reduction: about 75%
"Now they are on average about 75% faster and can be finished within a few weeks."
Claimed by: vendor
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Geospatial survey data (aerial or satellite imagery, or vehicle mounted panoramas and laser scans) covering the planning area
- An existing infrastructure and asset inventory (ducts, poles, cabinets, exchanges) that is kept current
- Historical civil works costs by area and terrain type, with enough volume to estimate from
- Local exclusion rules (protected sites, environmentally sensitive land, heritage paving)
Systems to integrate
- GIS or geospatial mapping platform
- Civil works costing or ERP system
- Network asset inventory database
- Permitting and municipal approval workflow
Complexity: High
The hard part is not the routing algorithm, it is the data: accurate, current geospatial imagery or point clouds, a reliable existing infrastructure and asset inventory, and enough historical project cost data by terrain type to make estimates trustworthy.
- 1
Start with one region and one clean data source
Pick a bounded rollout area and one reliable data source (an existing GIS layer or a single survey pass) before trying to cover a whole country, so data quality problems surface early and small.
- 2
Ground routes in real infrastructure and real costs
Feed the model the actual duct, pole and cabinet inventory and historical civil works costs for comparable terrain, not just a map. A route that looks efficient on a satellite image can still be the expensive one to build.
- 3
Keep an editable obstacle and exclusion list
Rivers, protected trees, heritage paving and other no go or extra cost areas should be a list a planner can add to, not something buried in model weights, so local knowledge fixes the model's blind spots quickly.
- 4
Require planner sign off before any route becomes a work order
The system proposes; a named planner approves. Deutsche Telekom's own account of its pilot describes exactly this: a planner double checks and approves the route before it is used.
- 5
Track estimate against actual, per job
Compare the AI's route length and cost estimate against what construction actually spent and built, and use the gap to improve the model and to flag terrain types it handles poorly.
Guardrails
- Planner approval required before a recommended route becomes a construction work order
- An exclusion and obstacle list a human owns and can edit, covering protected and environmentally sensitive areas
- Reconciliation of the infrastructure data the model uses against a physical audit sample, so stale records do not drive routes into ducts or poles that no longer exist
- Version control on the routing model and the data it was trained on, so a bad estimate can be traced back
KPIs to instrument
- Estimated versus actual civil works cost, per job
- Planning and approval cycle time, from data collection to approved route
- Share of recommended routes a planner approves without changing
- Change orders raised during construction that the plan did not anticipate
Human in the loop
A named planner reviews and approves every recommended route and cost estimate before it becomes a construction work order, and can override it with local knowledge the model does not have. On a portfolio of routes, planners also review a sample of accepted recommendations against the obstacle list to catch model blind spots before they reach many jobs.
Common failure modes
- Stale infrastructure data
- The model routes through a duct, pole or cabinet that was never recorded or has since been removed, causing rework once construction starts. Reconcile the asset inventory against a physical audit sample before trusting it at scale.
- Cost estimates that do not transfer across terrain
- A model trained mostly on urban routes underestimates rural terrain costs, or the other way round. Segment training data by area type and monitor estimate error by region before widening the rollout.
What are the risks and rules?
EU AI Act
Minimal risk
Designing and costing a proposed fibre route is planning for new construction, not the management or operation of an existing network, and it does not decide anything about an individual person. It is not a use listed in Annex III. Annex III point 2 (safety components in the management and operation of critical digital infrastructure) targets systems that are safety components in running live infrastructure; a route and cost planning assistant for new construction is not part of managing or operating that infrastructure, so it does not fall under it. Because it plans investment in a public electronic communications network, general obligations for network operators under the NIS2 Directive still apply to the operator, separately from the AI Act.
Rules that apply
Guidance
- Annex III, high risk AI systems referred to in Article 6(2) (European Union, Europe). Point 2 covers safety components in the management and operation of critical digital infrastructure; planning and costing new construction, rather than a component that manages or operates live infrastructure, is outside that test.
- Directive (EU) 2022/2555 (NIS2) (European Union, Europe). Sets cybersecurity risk management obligations for essential and important entities, including public electronic communications network operators, covering the systems and data they run, separately from any AI specific duty.
Controls to put in place
- Planner sign off recorded against every route before it becomes a work order
- An accountable engineering owner for the exclusion and obstacle list
- A documented review cadence comparing estimated and actual cost and route length
- Change control and version tracking for the model and its training data
- Automated anonymization of people and vehicles in survey imagery and point clouds before storage or review, under GDPR
Frequently asked questions
- How much faster is AI assisted fibre route planning?
- It depends heavily on how mature the data pipeline is. Deutsche Telekom's technology partner reports that once its geospatial data platform was in place, planning and approval that used to take several months came down to about 75% faster, finishing within a few weeks, after planning more than 10 million households and connecting more than 8 million of them to FTTH. That figure covers the whole data platform, not AI route determination alone. Telefónica Spain, which said in December 2025 the tool was already being used by its planning teams, describes its deployment as reducing human error and accelerating service availability, without publishing a percentage.
- Does this replace network planners?
- Deutsche Telekom's account of its pilot says a planner double checks and approves the route before it is used. Telefónica says the algorithm is already being used by its planning teams and, in its own words, "multiplies the capacity of our teams". The AI narrows a large search space and produces a consistent first estimate; the planner still owns the decision.
- Is this the same as mobile network capacity planning?
- No. This use case plans where to build new fixed line, mostly fibre, infrastructure and what it will cost. Forecasting mobile data traffic and tuning radio parameters on an existing network is a related but separate job, covered on the network planning and capacity optimization page.
- What data does an operator need before starting?
- Reasonably current geospatial imagery or point clouds of the planning area, an existing infrastructure inventory that is kept up to date, and a useful history of past project costs by terrain type. Without the last two, the AI can still draw a plausible route but its cost estimate will not be trustworthy.
How to cite this page
Blits.ai AI Use Case Library, "AI for fibre network route and build planning", last verified 29 September 2026, https://www.blits.ai/ai-use-cases/fibre-network-route-and-build-planning. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 29 September 2026: First published