technologyreseller.co.uk 29 AI information; most of our competitors don’t,” said Bashir; and b) because DevRev brings in both the record and the underlying definition of the schema. “We’re not bringing in data in a flattened, lossy format and saying ‘We want to be able to observe this record and be able to understand how to answer questions with this exact information’. We bring the record in, but we also bring in the underlying definition of the schema. That’s important because you don’t just want to know the city is Carmel and the state is Indiana, you need to understand that there’s a field dependency between those two attributes and you need to understand what the other drop-down values are and not just make assumptions on what they are or simply ignore what they are. You need awareness around that so that if somebody wants to get an action answer and act on that information, they know what values are permissible. You might say ‘I want to update the city from Carmel to Birlingham’ and it could say ‘That’s not an option, but you could change it to Birmingham’. ‘Okay, fine, let me do that’. The only way to allow for that is to know what the other options are. And you only know that when you have a deep understanding of the schematics that are powering the record as it stands today.” Why AI projects fail These capabilities, he suggests, or rather the lack of them, is one reason why so many AI project fail. “There are people in every business, in every department who are eager to able to organise information so that for example your customer success team and your customer sales team or customer experience team could work from a single tool; so that the person who’s fixing the bug in the back office and the person who’s connected to the customer in the front office are able to see the same quality of information across various dimensions and various layers. “That seemed like an awfully nice-tohave feature then, but it wasn’t a pain killer. After 2023, people felt that the ability to harness the data so that facts could ground conversational answers and subsequent actions became highly desirable. That was the major shift.” Flexible schema He adds that the work DevRev was doing then gives it an advantage over competitors who generally don’t have a flexible, versioned schema engine that he says is needed to unify structured and unstructured data. “Computer’s memory is fronted by a data layer that is SQL-compliant, and to do that you need a back-end component that is a fully versioned schema that you can run an entire database operation on. Most of our competitors are not building a schema layer, nor are they building an SQL engine, in which case the information received conversationally is not going to be grounded in something highly declarative and provable like SQL. Because you don’t have that grounded fact you cannot just transition from a conversational experience to a more traditional experience, like a dashboard or a widget, that you might share in a board meeting or put on a monitor and share with the rest of your organisation because it’s grounded in something factual and declarative. We’ve been able to ground everything we’ve done because we have this schema engine, this data layer. That’s unique in the market,” he said. Another competitive advantage, Bashir suggests, is the way in which DevRev brings in information. “It’s not just that we’re bringing in structured and unstructured information, which is highly differentiated in the market. We’re also bringing in access permissions. We’re bringing in the schema of the product,” he said. This is critical: a) because in the agentic world nobody’s going to recreate the entire access provisioning layer that already exists in systems of record, which you therefore cannot afford to lose during synchronisation – “We preserve that they wanted people to travel and eat outside their homes.” Michelin’s story is analogous to DevRev’s because technologies the company created for one purpose, to break down the barriers between developers and growth teams, have the potential to be of even greater value for agentic AI, not envisaged when they were originally developed. “When I joined DevRev in 2021 we had the idea to explore the intersection between AI and the enterprise, so we got AI from the very beginning. But in 2021, even though the world had already introduced the concept of Transformers, the concept of LLMs and conversational search, the way we see AI today didn’t exist. “We had already patented our first vector database, in 2021 or 2022, and unifying information and being able to help people work better together through the power of AI was something we were already building technologies around. It required us to invest tremendous time on workflows, because we needed to be able to synchronise data from disparate systems and we needed to bring in tremendous amounts of data, or on some days very little data, and be as reliable moving data that took two hours to move as moving data that took two milliseconds to move.” From the very beginning, DevRev invested in an identity engine; an authorisation layer; a flexible data layer, enabling workflows to leverage and synchronise information from diverse databases and SaaS tools; and the ability semantically to interpret data, all of which have become foundational services for the agentic revolution we’re now in. “The world of agentic hadn’t been introduced to the world then but the concepts had and those foundational services, which we had worked on since 2021, and our eagerness to explore that intersection between AI and enterprise meant that in 2024 we were in the perfect position to really develop agentic tools, an entire Agent Studio (for building specialised agents) and to define what we call Computer, the product we now have in market.” From vitamin to painkiller Bashir describes the launch of the first LLMs in 2023 as ‘transformative’ for people’s understanding of the value of unifying information which, to use his words, went from being a vitamin to a painkiller. “When we talked about unifying information prior to 2023, people felt it was desirable but not necessary to be continued...
RkJQdWJsaXNoZXIy NDUxNDM=