The challenge

Mariner Logistics, a prominent player in the logistics industry, offers warehousing, fulfillment, and inventory management services to fast-growing companies. With over 200 employees, Mariner Logistics serves renowned companies such as Bayer, Shasta, Masienda, and Blackhive. As a tech-forward brokerage, Mariner Logistics aimed to scale its business by enhancing efficiency and capacity through technological advancements.

Mariner Logistics launched Project Apollo to integrate various technologies into a unified system to provide turnkey automation for logistics brokerage. Project Apollo focused on automating freight coverage, managing capacity, bringing extra capacity online for businesses, and automating the spot bidding process with a risk-focused approach. The project leverages advanced AI technologies for AI phone calls, ensuring seamless cross-channel communication across email, voice, and SMS.

Project Apollo is a key competitive advantage for Mariner. It is designed to provide a seamless experience by integrating multiple systems, including customers’ ERP systems (such as SAP and Oracle), carrier systems, and supporting automated workflows. Project Apollo aims to meet carriers where they are—whether through phone calls, text messages, or emails—to ensure effective interactions, increasing successful engagements without forcing carriers to adopt new technology platforms. This approach digitizes non-digital interactions, creating immense value for shippers and 3PLs, and allows brokerages like Mariner Logistics to automate a large percentage of its freight operations, significantly enhancing both capacity and efficiency.

Before Digibee:

Before adopting Digibee, Mariner Logistics’ engineers were overwhelmed with multiple tasks. They developed custom engines for their SaaS product to meet specific business needs and spent significant time coding connections to third-party carrier and shipper applications. Additionally, engineers were hand-coding API data and orchestrations, which was time-consuming and created long onboarding times for new customers and carriers. This impacted overall customer satisfaction, operational efficiency, and created resource bottlenecks.

“We integrated our digital carriers with one of our technologies. It took two engineers two months to do it, costing tens of thousands of dollars,” said Ken Apple, vice president of engineering at Mariner Logistics.

The Objective

Mariner Logistics aimed to:

• Provide seamless integration with every shipper to digitally engage carriers who have not adopted technology, meeting them where they are rather than forcing platform adoption.
• Enhance operational efficiency and completely automate a large percentage of all freight that moved through the brokerage.
• Reduce onboarding times for customers and partners to improve customer satisfaction and operational efficiency.

The solution

To overcome these challenges, Mariner Logistics embarked on a transformative journey by rethinking their engineering priorities and adopting Digibee. Adopting an iPaaS solution rapidly increased their time to market and was critical to their new strategy. They were able to ideate, iterate, and go live without any engineers because all new product ideas for the business start with iPaaS from a business perspective. Key steps in this implementation included:

Mariner Logistics aimed to:

• Realizing the need for change:Ken Apple, VP Engineering at Mariner, recognized the inefficiency of their current approach. “I’m challenging a traditional premise that we have about how we should build software. We need to quit building applications and assemble solutions. You’re spending a bunch of money on expensive engineers. Those expensive engineers are solvers, they like to build. If you put any tech problem in front of them, they’ll solve it. But that’s not a cost-effective way of solving every problem. You have to have some discernment about the kinds of problems that are worth solving by an engineer versus not.”
• Identifying tasks for acceleration:Mariner Logistics first identified the time-consuming tasks that required their most skillful engineers. These tasks were categorized into business-specific or generic tasks that did not impact their core business logic (engines). Ken identified connectivity to carriers’ and customers’ ERP systems, as well as the workflows to implement business processes and data orchestration between engines and third-party systems as primary candidates for optimization and embracing AI.
• Choosing Digibee: Mariner Logistics evaluated Digibee alongside three leading iPaaS products and their custom solution. They used carefully crafted criteria and a short proof-of-concept (PoC) based on their specific requirements. They selected Digibee because of its ease of building workflows without special training or certifications, the acceleration in building workflows due to Digibee’s fail-safe workflows, and the reduction in operational efforts from predictable performance with dedicated Docker containers for every request execution. Digibee allowed for seamless integration with third-party applications and ERP systems like SAP and Oracle without disrupting existing workflows.
• Prototyping for Production: An engineer built a prototype to create an engine for 3PLs and brokers to advertise, quote freight, and fulfill orders with dynamic pricing driven by a team. The next phase focused on incorporating automation and AI. Each channel had a slightly different API/interface (email vs. contact vs. survey). The prototype built on Digibee was not only enough to validate the concept, but more than enough for production, and is already delivering immense business value. This approach has also changed the way the Mariner team thinks about solving any new business problem, connecting existing systems using an iPaaS instead of having to code everything.
• Optimizing the use of talent: With Digibee, Mariner has one person doing the work of two engineers, which used to take three times longer. It also allows less senior engineers to handle integrations, freeing up highly skilled engineers for complex tasks.

“Digibee allows us to use less experienced software engineers [to do data orchestration and integrations], and I can take highly skilled engineers and focus them on things that only they can do, like building sophisticated logic into our engines,” said Ken Apple, Mariner’s vice president of engineering.

By reducing custom development and leveraging Digibee for faster proof-of-concept (POC) development, Mariner Logistics has significantly accelerated its product development, increasing customer satisfaction and operational efficiency.

The results

Implementing Digibee has yielded remarkable results for Mariner Logistics:

• Accelerated Product Development:Reduced development time by up to 24x on certain key projects.
• Rapid POC development:Accelerated proof-of-concept development, minimizing waste and enabling swift idea validation. Scale from PoC directly to Production without having to rebuild
• Reduced costs and fast ROI:Achieved cost efficiency by avoiding expensive integrations, resulting in positive ROI within three months.
• Optimized resource utilization:Less experienced engineers handled integrations, while highly skilled engineers focused on developing sophisticated APIs and logic.
• Improved integration speed:Faster onboarding processes have reduced integration times for customers and partners, increasing customer satisfaction and operational efficiency.

“When I talk to people about Digibee, I tell them to create a force multiplier with this mature technology. People who use this type of technology find it to be a great leverage. As an engineering leader, you should not ignore this technology. It should be part of your repertoire and we have already seen great ROI,” Ken emphasized.

CIOs and VPs running integration teams know the pressure.

Your CEO has mandated an AI-first strategy and your stakeholders increasingly believe that agents are the answer to every workflow or integration need. Meanwhile, your backlog has a year’s worth of work, you’re racing to validate where and how to use AI, and you face real system, team, and data constraints.

So, how do you take on the task of matching each problem to the best solution?

Think through the integration backlog item by item. What does each job call for? Some fit with deterministic workflows; AI would only introduce risk, cost, and delays. Others present genuinely ambiguous problems where agentic AI unlocks automation that wasn’t previously possible.

A surprising number fall somewhere in between—elements that are structured enough to pre-define alongside requirements that are messy enough to need AI.

This sorting exercise is AI strategy in action. It demands you approach each problem with curiosity and discipline to assess the best solution.

This piece gives you a framework for when AI adds value, when it doesn’t, and how to tell the difference.

Start with outcomes, not solutions

It’s natural to bias towards a particular solution—a preferred platform, language, or the latest technique. But first: ask what success requires for the process.

Here are signals we look for.

Signals to build deterministic

Zero tolerance for failure

At 99% accuracy, a process that runs 100.000 times is wrong 1.000 times. For some tasks, that cost is too high. Imagine the consequences of 1.000 failed paychecks. If a process needs to work 100% correctly XNUMX% of the time, don’t use AI.

Decision explainability

If a regulator or compliance body might ask exactly why the company made an automated decision, stick with deterministic logic. It produces an audit trail. LLMs produce probabilistic outcomes that sometimes violate instructions. When they do, it’s difficult or impossible to understand why.

Simple A-to-B logic.

If a skilled developer could write the rules in an afternoon, write the rules. Leave simple things simple, and save AI for more complicated problems. You’ll save on both cost and complexity.

If any of these apply, you need a deterministic solution. An agent won’t make it better—it will make it less reliable, harder to govern, slower, and more expensive.

Signs that agentic AI adds value.

Unpredictable inputs.

If the project calls for ingesting documents in different formats, natural language requests, or other unstructured data sources, LLMs are typically the only reasonable solution.

Contextual, run-time decision making.

An LLM can reason across ambiguous inputs in ways that rule sets cannot.

Broadly, if a process requires reasoning too complex to put into code, it needs an LLM.

Considerations that can sway an unclear case

If the above qualifiers don’t pin a solution to one end of the spectrum or the other, these factors help locate where the task falls in between.

Throughput and latency.

High-volume processes with tight timing requirements should lean deterministic. AI inference adds latency and cost at scale.

Cost predictability.

The $47 11-day long agent spiral that made the rounds represents an extreme outlier, but agentic workflows carry variable operational costs at any scale. If budget predictability matters, model it carefully before committing.

Total cost of ownership (TCO).

Building code pipelines usually costs more in employee time. LLM endpoints charge per token. Failed runs can require manual repair. Account for all of this in your math. Is an AI pipeline that takes hours to build but fails 2% of the time cheaper than one that takes a week and fails never? That depends on the application.

If your task has you asking about throughput, latency, cost predictability and total cost of ownership, your solution may call for a mixture of deterministic and AI components.

Not just black or white. Solutions can be gray.

Modern integration work increasingly sits on a spectrum from deterministic to agentic workflows. In between lays a wide middle ground where deterministic foundations are enhanced by targeted agentic steps. This framework applies at both the individual workflow level and across projects.

Today, most organizations we talk to find the majority of their work belongs at or near the deterministic end. As it should; we’re talking about proven, solved integration and automation goals.

That said, agentic workflows open up powerful new opportunities—and our customers are increasingly finding creative ways to add incremental agentic value to deterministic workflows.

Deterministic workflows

Deterministic workflows form the backbone of enterprise integration infrastructure. They deliver reliable, auditable, repeatable, and cheap execution. When requirements are stable and inputs are well-structured, code workflows are almost always the right call—and often underappreciated in the age of AI.

Use when:

• Requirements are stable
• Inputs are well-structured
• Governance matters
• Failure is not an option

Common examples:

• Password recovery workflows
• Compliance audit log exports on regulatory schedule
• Bank transaction fraud flag escalations

Agentic workflows

Agentic workflows handle what code can’t: ambiguity, variable inputs, synthesis, and reasoning. They unlock automation for processes that previously required humans (often expensive specialists with limited availability) to make boring and repetitive judgments.

Those expanded capabilities come with real tradeoffs. Their output is inherently variable, which serves as a feature in some contexts and a liability in others. Agentic workflows also cost more to run, are harder to audit, and require more robust monitoring.

For that reason, it’s worth asking if the task truly needs an LLM. Some AI engineers have replaced LLM layers with clever regex filters, which excel at simple input and logic tasks. Claude Code even uses them in its harness.

Where regex does the job, it will always be cheap and fast. LLM reasoning abilities pick up where regex falls short.

Use when:

• The problem requires judgment, synthesis, or creativity
• When some variability in output is acceptable
• The task must adjust to dynamic context

Common examples:

• Contract summarization and risk flagging
• RFP response drafting
• Generating social posts based on a prompt

Deterministic Plus

Most integration workflows start as deterministic pipelines—as they should. "Deterministic plus" describes what happens when you enhance a proven, governed workflow by adding one or more agentic steps that deliver clear, bounded value.

This is not a 50/50 split. The deterministic pipeline remains the backbone; the agentic step or steps add additional value. A workflow might pull structured records from a database, apply a transformation, and enter them into a different system. Along the way, it may pass the record to an LLM that reasons through whether the moved data requires human attention. The AI touches one step. Everything else stays predictable, auditable, and cheap to run.

A complementary variant involves separate workflows. A fully deterministic pipeline ingests or moves a batch of data, then triggers an agentic workflow to analyze the batch for insights. This agentic workflow may also be used alongside many other workflows. For example, one of our customers is experimenting with a centralized evaluation workflow to assess the performance of others in its integration estate. The integration logic stays clean. The AI operates only where variability is acceptable.

This solution lets organizations capture AI value without exposing critical infrastructure or operations to agentic failure modes. It also reflects how most integration estates will evolve—incrementally, deliberately, and with governance intact.

Use when:

• A well-governed deterministic workflow could deliver more value with agentic enrichment
• The core process must remain predictable, but edge cases or outputs benefit from AI judgment
• You want to evolve an existing pipeline rather than rebuild it

Common examples:

• IT helpdesk ticket routing with AI-assisted triage notes
• Supplier invoice processing where exceptions are flagged by an LLM for human review
• Automated release notes drafted by an agent from structured commit data

The best integration strategy is a deliberate one

The most capable integration leaders aren’t the ones “doing the most” with AI. They’re the ones who deeply understand the tradeoffs among deterministic and agentic solutions.

For that reason, it’s worth asking if the task truly needs an LLM. Some AI enAI promises to create meaningful business value. But success means staying away from “AI washing" and applying AI to the right tasks. If a deterministic workflow isn’t broken, don’t fix it. If AI doesn’t make something better, don’t add it. Look for unmined enterprise value in deterministic pipelines that can be enhanced with targeted agentic steps.ineers have replaced LLM layers with clever regex filters, which excel at simple input and logic tasks. Claude Code even uses them in its harness.

Here, platform choice becomes a real strategic variable. A platform built to handle integration, automation, and agent development in one place (like Digibee) makes it significantly easier to add agentic steps to existing deterministic workflows, govern the result, and iterate over time.

The backlog doesn't need to be an AI backlog.

It needs to be a solved backlog.

The tool follows from the problem—and that sequencing is exactly what separates integration leaders from integration followers.