New integration threads are needed to form the data fabric for AI.

Forbes

24-06-2025

New data integration threads are needed to weave a stronger data fabric for AI use cases.
Data demands direction. Left on its own, any information resource is of comparatively little value in and of itself until it is applied to a specific use case, inside a specific application (or other data service) through the specified calculations of an algorithm and under the auspices of a specific set of controls for provisioning, maintenance and management.
Nowhere is this simple enough truth more evident than in AI. The technology industry loves to remind us that in AI, it's 'garbage in, garbage out' and that we should focus on the provenance and preparation of data long before we start to think about giving it a seat at the dinner (or boardroom) table and putting it to work.
Sloppy Data In AI
While the average user might expect data injection, ingestion and integration within the bedrock of AI to be a precise science, instances of AI can experience limitations as a result of poorly aligned data channels. It's not so much a case of garbage in or out, more a case of trying to work with good ingredients, but while working from the garbage heap… with refuse getting the way and a lack of clean work surfaces. In data wrangling terms, that translates to AI being built on incomplete knowledge bases, AI being deployed through public cloud computing resources or on-premises equipment that has memory limitations, the use of misaligned large language models and a lack of fluid 'information transports' exacerbated by rigid interaction channels.
This is the pain point that AI integration and automation platform company Tray seeks to remedy. Rich Waldron, co-founder and CEO of Tray says that his firm's Merlin Agent Builder addresses the gap between agent deployment and real-world user adoption. Something like a sat-nav system designed to make sure AI agents not only get built, but actually get used and driven, he says that (all too often) agent experiences feel clunky and disconnected.
'This is because agents lose context, rely on narrow knowledge and force users to start from scratch in every session. Behind the scenes, IT and AI teams struggle to align the right LLMs to the right use cases, especially in multi-agent environments. Without flexible deployment options, it's hard to meet users where they work,' said Waldon 'To bridge the adoption gap, agents need smarter data access, built-in memory, LLM flexibility and tailored user interactions… and all that needs to be built to drive sustained usage, not just prototypes and demos.'
Not A Services Wrapper
Waldron's fellow co-founder and company CTO Alistair Russell insists that, 'Merlin Agent Builder isn't a services wrapper. It's a fundamental part of our product and is built for ease of use and scale.' By which he means that this product doesn't do what a service wrapper technology does i.e. act as an intermediary abstraction layer that enables non-native computing services to run on an operating system that they were not specifically designed for. Although service wrappers are popular for a variety of use cases (they extend functionality, enable management options and can provide granular control), Russell is suggesting that his firm's tools go deeper and work at the lower substrate layer, weaving a set of interconnections that exist far closer to where data itself is born.
'Our platform handles chunking [diving data into more manageable, digestible pieces] and embedding at the source, ensuring each data source is optimally segmented and vectorized so agents are grounded in high-signal, relevant context. That means fewer retrieval failures, more reliable decisions and agents that reason and take action," explained Russell.
He asks us to imagine building an IT help desk agent to automate ticket resolution. But the knowledge it needs (past tickets, solution articles, internal policies etc.) is scattered across siloed systems. When agents can't find the right data for the user, user trust breaks down. The answers offered by agentic AI at this level start to 'feel a bit off' and conversations fall flat.
'For the teams building agents, one of the most time-consuming, technically challenging parts is grounding the agent in the right data. They rely on custom ingestion pipelines or manual preprocessing, burning developer time just to prepare knowledge for agent use. Even then, keeping that data updated and consistently accessible across agents is a challenge,' said Russell. 'Tray's data sources [controls]
eliminate the barrier on both ends by making it easy for users to connect and sync structured and unstructured knowledge from file uploads or sources like Google Drive.'
Today he laments, we are at a point where an IT help desk agent is answering follow-ups and handling escalations, but every time a user returns, it forgets the context of their earlier issue. This is because most agents forget everything between conversations. Short-term context often gets lost in other platforms due to token limits and storage constraints. Long-term memory usually requires custom engineering or patchwork workarounds to avoid frustrated users and disconnected experiences. To address this, the Tray team says that Merlin Agent Builder now includes maximum short- and long-term memory, so agents can track session history and refer back to prior conversations automatically.
Competitive Analysis, iPaaS Integration
For all Tray's data management, data wrangling, data channeling and data integration capabilities, it's not unreasonable to call the firm an iPaaS player. Analyst house 2025 Gartner places the company in its Magic Quadrant for iPaaS after all. The integration Platform-as-a-Service market is both variagated and various in the types of firms that dominate and proliferate in this space. Top usual suspects in this arena include Boomi, MuleSoft, Workato, SnapLogic, Jitterbit, Informatica and database giant Oracle with its own Oracle Integration Cloud technology. AWS, Microsoft, IBM and Huawei Cloud also all feature in this market.
Workato is known for AI-fuelled automation technologies, some of which cover the integration space and many of which fall into the low-code tooling zone. More directly recognized as an integration purist, Boomi (for just over a decade in the 2010s part of Dell, but no longer) offers iPaaS capabilities that run across cloud-native, hybrid mixed environment and legacy data sources. Possibly the 'Windows of the iPaaS market', Boomi is thought of as user-friendly, but without the more advanced application programming interface connectivity that some vendors boast in this arena. Absolutely API-first is MuleSoft (part of Salesforce these days) with its MuleSoft Anypoint Platform, the company is all about API-centricity, API integration, API management and (logically enough these days) API AI.
Lesser-known integration brands Jitterbit and Zapier enjoy adoption with smaller to medium-sized businesses that need rather more point-and-click technology services. SnapLogic wins some over for its speedy deployment, its big data alignment and the general appeal of its Iris AI service; commensurately, it loses some prospects over what some feel is its less transparent pricing structure. Data integration is deep and complex, so customer use case contracts shouldn't be, but they sometimes are. Vendors like Celigo get their niche status by offering data itgeration for more defined uses (in this case e-commerce), so think extra governance and security here. IBM App Connect, Microsoft Azure Logic Apps and Tibco Cloud Integration also all make up the smorgasbord of specialists in this market sector.
As the iPaaS market continues to develop, AI will (perhaps obviously) feature more prominently as a purchasing decision factor. This will likely enable more iPaaS technologies to move outwards from the datacenter and work at the smart edge in the internet of things, all of which will see more real-time data streaming come to the fore as a critical must-have.
All That Agent Talk
Watch any news feed on agentic AI services and there are countless pages of new developments telling us about sparkling new agentic functions. Some will be applied to new HR use cases, some will work in disconnected air-gapped deployments such as military-grade software installations, some will offer point-and-click simplicity and some will offer new strains of cloud-native functionality so that they are well-suited to align with Kubernetes orchestration layers and so on. There is an almost infinite variety.
What will make fewer headlines are the agentic functions that offer delivery data configuration and integration advancements for smart routing and the ability to maintain continuity across most complex multi-turn interactions, but that's what's happening here. This is a question of data direction for agents so that they steer us the right way and keep us out of the garbage heap.