AI Route Planners Tested On UK Lanes: Where They Fall Apart

I spent three weeks feeding the same six route requests into every AI cycling route planner I could get an API key or a free trial for, across two areas I know well enough to spot a lie: the Peak District around Hathersage and the flat lanes north of Woodbridge in Suffolk. The conclusion arrived early and never budged. The model doing the planning is almost irrelevant. What decides whether you get a rideable loop or a hike-a-bike on a bridleway is the OpenStreetMap tagging underneath, and in the UK that tagging is patchy in ways that break AI routers far worse than they break a human with a paper map.

That distinction matters because of how these tools are sold. “Our AI understands your riding style” is doing a lot of work in marketing copy for what is, in most cases, a large language model writing a natural-language wrapper around a graph search on OSM data with a surface-preference weight. The LLM is not looking at the ground. It is looking at tags. When the tag says surface=unpaved and the reality is two miles of fist-sized limestone, no amount of model capability closes that gap.

The test setup

Six requests, run against Komoot’s AI-assisted planner, Strava’s route builder with the 2025 natural-language layer, RideWithGPS Trip Planner, Ride with AI (the OpenRouteService-backed startup), and a bare Claude Opus 5 prompt with the Overpass API wired in as a tool so it could query OSM directly. That last one was the control: an LLM with genuine access to raw tags rather than a vendor’s pre-cooked graph.

The requests, kept deliberately plain:

1. 65 km road loop from Hathersage, under 1,200 m climbing, avoid the A6187
2. 90 km gravel loop from Hathersage, majority off-tarmac, no hike-a-bike
3. 40 km TT-suitable out-and-back near Woodbridge, minimal junctions
4. 120 km road loop from Woodbridge, avoid A-roads entirely
5. 55 km gravel loop from Woodbridge (the hard one: Suffolk has almost no legal off-road)
6. 30 km recovery spin from Hathersage, under 400 m climbing

Every output got ridden or, where clearly unrideable, ground-truthed on foot and against the Peak District National Park rights-of-way layer and Suffolk County Council’s definitive map.

Peak District: surface tags that describe an intention, not a road

Request 2 is where things came apart fastest. Four of the five planners routed me onto the Roych, a byway open to all traffic running west from Rushup Edge. In OSM it carries highway=track, surface=unpaved, smoothness absent. Komoot’s planner described it in its generated summary as “a scenic gravel section suited to wider tyres.” The Roych is a trench. It is loose, wet, cobbled with embedded rock the size of a fist, and I pushed my 45 mm Pathfinders for eleven of its sixteen minutes.

Here is what the underlying data actually says, pulled via Overpass:

way 4257631
  highway=track
  surface=unpaved
  tracktype=grade4
  designation=byway_open_to_all_traffic
  bicycle=yes

tracktype=grade4 is the honest field, and it is the one every consumer planner ignored. Grade4 means “mostly soft, sometimes with hard core, unpaved.” Komoot’s surface model collapses unpaved into a single “gravel” bucket and weights it as mildly preferable for a gravel bike. The one tool that got it right was the bare Claude prompt with Overpass access, because I had told it explicitly to read tracktype and smoothness and to treat grade4 and grade5 as walking. It routed me instead down Jaggers Clough and out via the Hope Valley lanes, and the summary it wrote was appropriately hedged: “tracktype=grade3 on the descent, no smoothness tag, treat as unverified.”

That is the whole finding in miniature. The model was not smarter. It had better instructions about which tag to read, and the tag existed.

Coverage numbers from the Hathersage 20 km box, all ways with highway in track, path, bridleway:

Eight percent smoothness coverage. That is the ceiling on how well any AI route planner can predict whether you will be riding or pushing, and no LLM can raise it.

Suffolk: the access problem, which is worse

Peak District tagging is thin on quality but broadly honest about legality. Suffolk inverts this. Surfaces are simple (it is nearly all tarmac or farm track) but access rights are a mess, and the AI planners fail in a direction that can get you in trouble rather than just tired.

Request 5, the 55 km Suffolk gravel loop, produced four routes and three legal problems. Strava’s builder sent me along a headland track east of Charsfield tagged highway=track with no access tags at all. No access, no designation, no bicycle. OSM’s default assumption for an untagged highway=track in the UK renderer chain is permissive, and the routing engines inherit that. The track is private farm access. It is not on the definitive map as a right of way. A router that treats absence of an access tag as permission will put you on private land repeatedly in East Anglia, because untagged farm tracks are the dominant off-road feature there.

RideWithGPS was better here, largely because its planner is more conservative and simply refused to build a majority-unpaved Suffolk loop, returning a 55 km route that was 84% tarmac with an honest note that off-road options were limited. Correct answer, unglamorous delivery. Ride with AI generated an enthusiastic 55 km loop with 31% off-road and a summary claiming “quiet byways and forest tracks.” Four of those segments were permissive paths through Rendlesham with bicycle=no explicitly tagged, which the engine had apparently overridden in favour of matching the requested off-road percentage.

The pattern: when you ask an AI planner for something the terrain cannot supply, several of them hallucinate compliance rather than push back. That is an alignment problem in the product, not a routing one.

Where the AI layer genuinely adds something

Request 3, the Woodbridge TT out-and-back, was the one case where natural-language input beat a manual builder outright. “Minimal junctions” is awkward to express in a conventional planner. You end up dragging a line and eyeballing it. Claude with Overpass access counted junction nodes per kilometre directly and returned an A1152/B1084-avoiding stretch between Bromeswell and Tunstall with 1.4 junctions per km against 4.7 for my usual Sunday route. Strava’s natural-language layer understood the request and produced something reasonable, though it optimised for low traffic rather than low junction count and the two are not the same thing.

Request 1 and Request 6, both pure road, were fine everywhere. Every tool handled them. Road tagging in the UK is excellent, A-road avoidance works, gradient data from SRTM is good enough for a climbing budget. If you only ride tarmac, the tooling is solved and has been for years; the AI wrapper mostly saves you a bit of clicking.

Pacing the resulting route is a separate problem with separate failure modes, and worth reading about on its own. The AI route planning and race pacing pillar covers how these tools handle target power and time-to-completion once you have a GPX you trust.

The prompt that actually improved results

If you are using a general-purpose model with OSM access rather than a consumer app, the tagging instructions do most of the work:

When routing off-tarmac in the UK:
- Reject any way with tracktype=grade4 or grade5.
- Reject smoothness=very_bad or worse.
- If surface=unpaved and tracktype is absent, flag as UNVERIFIED
  and cap total unverified distance at 15% of the route.
- Require an explicit legal basis for every non-highway way:
  designation=public_bridleway, designation=byway_open_to_all_traffic,
  or bicycle=yes/designated. Absence of access tags is NOT permission.
- Report per-segment: length, surface, tracktype, smoothness,
  designation, and your confidence.

That last line matters more than it looks. Forcing per-segment tag reporting converts a confident paragraph into an auditable table, and you can see instantly which third of your route is guesswork. Running Request 2 again with this prompt, the model returned 61 km of the requested 90 and said it could not find a further 29 km meeting the criteria. That is a useful failure. Komoot’s confident 90 km with eleven minutes of pushing is not.

What to do about it

Check tracktype before you ride anything a planner calls gravel. Overpass Turbo is free and a query takes thirty seconds. Treat any UK off-road segment with no designation tag as unverified until you have checked it against the county definitive map, which for most authorities is now a public web layer. And if a tool gives you exactly the off-road percentage you asked for in a county that does not have that much legal off-road, it is telling you what you wanted to hear.

The gravel scene’s growth in Britain is running ahead of the data that describes it. Somebody surveying smoothness on a hundred Peak District tracks would improve every AI cycling route planner on the market more than a frontier model upgrade would, and that is a strange position for the industry to be in.