Optimizing your AI investment with decision services

Identify and link operational areas that benefit from decision models such as vehicle routing, worker scheduling, and order fulfillment – empowering your teams to deliver more decision AI projects faster.

The wave of generative tech and LLMs over the last year has prompted businesses to think differently about their AI investments. And this doesn't just mean how to start using ChatGPT. There's a wide ecosystem of technologies that can fit into the AI bucket that includes machine learning (arguably the term du jour before "AI" stepped in) and decision science (optimization, simulation, and heuristics for decision making). 

This leaves many technology teams and leaders asking: What’s our strategy? Where are we missing opportunities? How can we get the most ROI? In industries from last-mile delivery to retail, companies are using AI tech such as machine learning and optimization to improve everything from operational efficiency to sustainability initiatives. 

We’ve helped teams (big and small) map out strategies to optimize their budget and accelerate their AI projects. We start by asking them to consider the following: 

  • Are there areas where manual decision-making can be automated? Where are you spending a lot of time on repeatable tasks?
  • Where are you already using declarative and generative AI tech like optimization and ML? Can (or should) these areas be linked? 

In this post, we'll work through these questions and map out (literally) what the AI tools are and how they fit together to deliver improved outcomes. We'll tackle this mapping exercise through the lens of decision services since they're an effective way to apply AI to your operations by automating processes that take in operational data and output plans. Whether a decision is made via a business process or an optimization model, mapping it will help understand the flow. Let's take a look.

Interested in chatting about an AI strategy for your specific use case? Get in touch with our team to optimize your AI investment and tackle more projects faster.

Understanding decision services

What does a decision service do? Do I have a decision service? A decision service automates the process of making a decision like matching a driver to a delivery order. If you’re automating the process of creating routes or scheduling shifts in any way (e.g., with a rules-based system, optimization, and/or machine learning), you likely have a decision service!

What are common decision services? Examples of decision services in the logistics space include: 

  • Vehicle routing
  • Shift scheduling
  • Order fulfillment
  • Resource allocation

What does a decision service look like? A decision service is an app, container, or program that solves an operational problem by reading input data and producing a solution (or plan) as an output. We often talk about decision services in the context of optimization, where there’s a model used to define the problem and a solver used to find feasible solutions.

Where does a decision service fit? How does it integrate with other tech in your systems? Often, inputs come from a database, the model runs and solves the optimization problem with an algorithm or set of algorithms (open source, commercial, custom), and then the output goes to ops tooling.

While every system architecture is unique, there are similarities in terms of the major components of these systems. We can use that framework to lay out where decision services can sit in order to increase operational efficiency of the business.

Mapping decision services

Finding where decision services exist, or could exist, at your organization is a helpful exercise when deciding how to effectively use AI tech. In addition to pinpointing these areas, it’s also important to understand how various functions/departments can work together. Let’s look at an example of a decision service map for a fictional farm share company that delivers produce from local farms to customers’ homes. In this exercise, we’ll identify existing decision services, add a new service, and split a large service into smaller ones.

Identifying existing decision services 

At the farm share, we start with historical order demand that serves as the input into a forecasting (machine learning) service. The output (worker demand forecast) is then used as the input to the next service (driver shift planning). That output (scheduled shifts) alongside order data is used as the input into the next service (order distribution) to create assignments.

Creating and linking decision services 

So what’s next after sketching out the current state? We take a look at surrounding areas that are using manual processes. We know that order data is read from a database and involves a manual process – using a static, rule-based approach for determining the number of orders to outsource to third-party drivers and the number of orders to assign to our in-house fleet.

We can easily identify this as an opportunity to save money by creating an order fulfillment service with inputs like carrier rate data, third-party orders, and customer orders to create the in-house fleet orders.

Let’s add this new decision service to our map and link it to our existing services. With the addition of our order fulfillment service, we are now: 

  • Optimizing order fulfillment as an automated service to minimize costs
  • Accounting for how carrier rate data impacts the number of in-house fleet orders and therefore driver assignments
  • Factoring in multiple, sometimes competing business rules and constraints that need to go into the decision

Splitting one decision service into multiple services

Another way to create more efficient operations is by splitting one service into two or more. Let’s take a look at how to do this at the farm share company. Currently, driver shift planning is a single service that both creates shifts and assigns workers to those shifts. 

We’d like to split shift planning into two decision services (shift creation and shift assignment) for the following reasons:

  • Shift creation only needs to happen once every two weeks while shift assignments need to be done daily due to more frequent changes in driver availability. 
  • We’d like to use different objectives (maximizing coverage vs maximizing driver happiness) as well as different model formulations for shift creation and shift assignment. 
  • Breaking one application into smaller applications has historically been better for our teams when performing maintenance and/or troubleshooting.

Finding the right approach

This simplified example outlines the flow for a delivery use case – demonstrating how decision services can link together to eliminate silos and streamline operations. Similar maps can be drawn for any organization from home services providers to ride-share companies.

From a high level, what should you consider when adding, consolidating, or linking services?

  • Your unique KPIs: On-time deliveries, vehicle utilization, driver happiness, etc. 
  • Complexity: Are we under or over-constraining the problem by linking or combining them?
  • Performance: What’s the solution quality? How quickly is the problem solved?
  • Costs: Are we saving money or time?
  • Revenue: Does this impact revenue?

Optimize AI investments with Nextmv

Nextmv can help demystify this process and accelerate your teams to ship more models with confidence. 

With Nextmv, every decision service looks and feels the same so it's easier to integrate. Plus we have a growing open source repo of sample apps to help teams get started faster. 

Mapping: Our team of decision scientists and engineers has hands-on experience building effective systems for large retailers, food delivery companies, transportation services, government agencies, and more. We can help you map your decision AI strategy in a matter of hours rather than weeks or months.

Orchestration: It’s easy to interact with and manage your models using a unique set of API endpoints.

Scaling: Create (and repeat) model dev workflows with Nextmv apps to cover a variety of operational use cases.

CI/CD: Use a Git-style flow to automate the management and versioning of your models. Kick off experiments like acceptance tests and shadow tests to ship to production with confidence.

Integrations: Nextmv integrates with many optimization modeling tools and solvers, so you can choose what works best for you (OR-Tools, HiGHS, Gurobi, Pyomo, and more).

Get started

Ready to get the most out of your AI investment? Sign up for a free Nextmv account to explore the platform or reach out to us directly to map out an AI tech strategy with our team.

Video by:
No items found.