Static routing assumptions fail in congestion, Tarot Analytics brings time-aware optimisation to last-mile planning.
Tarot Analytics has launched a new traffic-aware route optimisation engine that allows last-mile operators to plan multi-stop delivery routes based on how congestion changes across the day, rather than relying on static average travel times. The new system uses machine learning trained on large volumes of historical traffic data to model how travel speeds shift across five-minute intervals, enabling routes to be optimised around real-world driving conditions instead of theoretical averages.
Co-Founder, Tarot Analytics.
The new engine represents a major upgrade to Tarot’s optimisation platform since the company was founded by Simon Webb and Jesse Treharne. Over the past decade, Tarot has focused on multi-stop delivery and field service operations across sectors including retail distribution, food and beverage, pharmaceuticals, automotive spare parts and courier networks. The company’s software is designed to help operators decide which vehicle should visit which jobs, and in what sequence, while also managing driver communication, proof of delivery and customer tracking.
Jesse, co-founder of Tarot, says the decision to rebuild the algorithm was driven by a recurring customer question.
“We kept getting asked whether we take traffic into account,” he says. “Historically, our answer was that you did not really need to, because the normal variation in driving time often mattered more. But we kept hearing it again, so we decided to take a serious look at it.”
From static averages to time-of-day routing
That led Tarot to develop a new generation of its optimisation engine that treats traffic as a dynamic variable rather than a fixed assumption. Instead of planning routes as if roads behave the same way all day, the new system recognises that congestion evolves hour by hour and even minute by minute.
“We are still solving the same multi-stop delivery problem,” Jesse says. “But now we are considering time of day and traffic. That means the algorithm is no longer working with hypothetical mathematical distances. It is working with what happens on the road.”
Tarot is built for classic milk runs, where each vehicle makes many stops across a shift. This could involve delivering groceries from a distribution centre to multiple stores, supplying spare parts to workshops, delivering furniture to households or sending technicians out to perform field service visits. In these environments, the order in which stops are made can be just as important as the distance between them.
“If an algorithm sends you into the city at 9am without knowing it is peak hour, you are going to get stuck,” Jesse says. “Our new algorithm is aware of that and changes the sequence so it reflects what the roads will actually be like.”
Why traffic is hard to model at scale
While consumer navigation apps have shown traffic-aware routing for years, applying that concept to large-scale route optimisation is much more complex. A consumer app only needs to calculate one route at a time. A fleet optimisation engine may need to calculate tens of thousands or even millions of travel times in a single run.
Simon, co-founder of Tarot, says this has been a major barrier for the industry. “If including traffic saves you 20 minutes across the fleet, but getting the data means the calculation takes 30 minutes to run, then you have not actually gained anything. The challenge has always been getting the data fast enough to be useful.”
Many existing systems work around this by building routes first and then adjusting for traffic afterwards. Others only account for traffic within short windows, such as assuming a single traffic state for the next half hour. That approach breaks down when routes run across several hours and through multiple traffic phases.
“Our view was that traffic is not a static thing,” Jesse says. “A route might start in early morning, go through peak hour, then into the middle of the day. Every decision you make at the start affects what happens later, so the algorithm must understand that.”
Using machine learning to predict congestion
To make this possible, Tarot built a hybrid data and machine learning pipeline. The starting point is open-source road network data, which provides baseline travel times based on road geometry, speed limits and intersections. On top of that, Tarot layers in large volumes of historical time of day traffic data from commercial providers.
“We take that historical data in bulk for a whole city,” Jesse says. “Then we train a neural network on it so it learns how traffic changes compared to the baseline. Instead of trying to learn every travel time from scratch, the model learns the percentage impact of traffic at different times of day and in different parts of the network.”
The day is divided into five-minute intervals, creating 288 time slices across a 24 hour period. That level of granularity allows the model to capture the difference between, for example, a 7.30am run into an industrial area and a 10am run along the same roads.
Training the model is computationally intensive and can take more than a day for a large city. However, once trained, it allows Tarot to generate traffic adjusted travel times at the scale required for route optimisation.
“When a customer runs the algorithm, we need thousands or millions of travel times instantly,” Jesse says. “We cannot wait for external providers to calculate them one by one. So, we use the trained model to generate those values very quickly. It is slightly approximate, but in our testing, it is very accurate.”
The result is an optimisation engine that can build routes in advance using realistic traffic patterns, then continue to adjust them during the day as conditions change. For many operators, planning still happens the day before or early in the morning, based on expected conditions. As vehicles are on the road, the system can respond to delays, cancellations, accidents or urgent new jobs.
“The answer is both,” Jesse says. “You need a good plan at the start of the day, and you also need incremental updates during the day as reality changes.”
What this means for fleet operators
The benefits of this approach extend beyond faster routes. By reducing total driving time and kilometres travelled, operators can lower fuel consumption, vehicle wear and tear and the number of drivers required to complete a given workload.
“In general, our customers see around a 30 per cent reduction in those costs,” Jesse says. “It depends on what they were doing before and the geography they operate in, but you are getting a route that matches reality much better.”
Simon adds that there are also back-office gains.
“We have worked with customers who have reduced their daily planning time from three hours to about 15 minutes. That means they can take on more work without adding more staff.”
There is also a sustainability dimension. Fewer kilometres driven means lower emissions, without requiring major capital investment.
“You are reducing emissions and saving money at the same time,” Jesse says. “It is a very practical way to be greener.”
For Tarot, the launch marks a shift from theoretical optimisation to what Simon describes as reality-based planning.
“This is about making sure the algorithm reflects what actually happens on the road, not what a spreadsheet thinks should happen,” he says.
As last-mile networks become more complex and customer expectations continue to rise, the ability to plan around real-world conditions is becoming less of a nice to have and more of a requirement. By embedding traffic directly into the optimisation process, Tarot is aiming to give operators a clearer view of what their fleets can really achieve each day.
