Last month, Actively AI hosted two deep-dive panels at its office in New York City with builders shaping the next wave of enterprise AI agents.

The first session, moderated by Anshul Gupta (Co-Founder, Actively), brought together leaders in enterprise AI: Nick Hanley-Steemers (Engineering Leader, Decagon), Aditya Ramanathan (Engineer, Hebbia), and Mihir Garimella (Co-Founder, Actively) for a discussion on how AI interfaces are evolving beyond chat and what it takes to make agents reliable and valuable inside real enterprise workflows.

The discussion traced key trends in how the frontier of enterprise AI is shifting: from conversational prototypes toward embedded, persistent systems that collaborate with humans in real time. In this blog, we share some of our biggest takeaways.

Interfaces are evolving beyond chat

Chat remains the most natural way for users to begin interacting with AI — minimizing friction and maximizing exploration. But as teams embed AI deeper into daily work, chat alone becomes restrictive, as illustrated by the builders of ChatGPT. Our panelists discussed how they are now experimenting with multimodal interfaces that combine transparency, initiative, and contextual grounding.

For Decagon, voice introduces a new set of constraints: ultra-low latency and tolerance for messy, overlapping speech.

“You can’t force users down a predefined path with voice. The system has to adapt to how people actually talk.” - Nick

Spreadsheet-like views, such as Hebbia’s Pubmatrix, expose how models reason — transforming opaque outputs into auditable logic.

“The reason we’re big believers in Excel is that it’s a great interface for communicating why.” - Aditya

Actively’s per-account agents extend that evolution further, pushing insights and recommendations at the right moment rather than waiting for user prompts.

“Pull (chat), Push (proactive suggestions), Ambient (real-time assistance), and Autonomous (self-directed action)—designing for all four is how AI moves from reactive tools to persistent collaborators.” - Mihir

Each mode (chat, voice, grid, push, ambient) represents a distinct interaction style. The long-term goal is to blend them into continuous, contextual touchpoints across the tools where work already happens.

(For a deeper view of Actively’s perspective on human-agent collaboration, check out Mihir’s recent post)

Durable agents depend on strong primitives

As interfaces diversify, the real foundation shifts below the surface. Our panelists agreed that reliable, long-running agents depend on three primitives: memory, executable reasoning, and observability.

Memory is not a feature — it’s infrastructure. Actively’s per-account agents run indefinitely, carrying forward everything they’ve seen, inferred, and learned about each customer. Their memory stack spans procedural (how to act), semantic (what’s true), and episodic (what’s happened). But durability depends on integrity: once memory becomes corrupted — a bad fact, a wrong assumption, a mislearned pattern — every subsequent action compounds that error. Recovering from that drift is one of the hardest unsolved problems in building persistent agents.

“Memory is probably the most important component. If it gets corrupted even once, the agent’s understanding of that account can go off forever.” Mihir

Decagon reinforces durability with executable reasoning — letting models write and run code to test hypotheses rather than relying solely on text prediction.

“We give the model a sandbox. It writes code to test hypotheses, runs it, and uses the results to guide its next step. This makes reasoning transparent and reproducible.” Nick

Hebbia extends that idea by embedding multiple analytical perspectives into each query.

“Our users are professionals who live in documents. The value comes from agents that can reason through multiple analytical lenses, not just extract data.” Aditya

The enterprise layer is a human problem

Technology alone doesn’t drive enterprise adoption—organizations do. Deploying an agent into production means contending with fragmented systems, entrenched workflows, and risk-averse cultures. Every panelist described a version of the “last-mile” gap: the friction between a capable model and a usable product.

Deploying AI inside an enterprise is less like shipping software and more like fitting a suit. The base stack (models, tools, and infrastructure) provides the fabric, but it still needs skilled tailoring. Last-mile delivery must ensure the system fits the organization’s messy data (because bad data breaks good intelligence) and reflects its internal motion, such as how sales actually happen, how decisions are made, how teams collaborate. Without that customization, even the best model remains theoretical.

Hebbia relies on AI Strategists, who former finance or legal professionals translating AI capabilities into measurable ROI.

“We deploy AI to people who already have deep expertise, such as lawyers, bankers, researchers. You can’t just drop an LLM and hope it fits.” Aditya

Decagon embeds Agent Project Managers and Agent Engineers alongside client teams to tune logic and escalation paths.

“Every large deployment ends up being high-touch. You need Agent Project Managers and Agent Engineers sitting with the customer, tuning the AI to their exact workflows.” Nick

Actively mirrors this structure with Forward-Deployed Engineers and Agent PMs, who ensure each deployment reflects how the customer actually sells rather than how the system was designed.

“We’ve learned that deploying AI to sales teams is a high-touch problem. You need people who understand the domain and can configure the agent for each team’s motion.” Mihir

Modern agents thrive on messy data

As noted earlier, Enterprise data is rarely clean. Traditional rules-based systems collapse when inputs are inconsistent or incomplete. The new generation of agents turns that mess into signal.

Rather than rejecting noisy inputs, these agents reason through ambiguity — assigning confidence scores, reconciling contradictions, and inferring likely truths. Actively’s per-account agents, for instance, continuously reconcile misaligned CRM and engagement data, adjusting their internal state instead of failing the task. The result is a shift from validation to interpretation: determining the most plausible state of the world given partial evidence.

“Human sales reps deal with messy CRM data all the time. Half the fields are wrong, but they still get the job done. Agents should be able to do the same. One of the biggest opportunities with agents is that they can actually reason over messy or complex data.” Mihir

By treating imperfection as a design assumption, not an exception, modern systems transform “garbage in, garbage out” into a competitive edge.

Security and scalability are non-negotiable

As AI systems enter regulated domains like finance, healthcare, and law, security and scalability have become the entry ticket, not the differentiator. Winning enterprise trust now requires architecture that satisfies both auditors and CTOs.

This discipline shapes design from day one: single-tenant or VPC isolation, encryption at rest and in transit, and infrastructure capable of trillion-token workloads. Decagon and Hebbia have both architected for strict compliance and full auditability, while Actively isolates per-account agents to guarantee data privacy and tenancy boundaries.

“To win in finance or law, you need to prove control before you prove capability.” Aditya

Enterprise AI adoption will ultimately hinge not just on intelligence but on control, transparency, and operational scale.

Next: In part two of this series, we will dive into the second panel, featuring engineers from OpenAI and LangChain, on the machinery that makes these agents real and the emerging patterns for building reliable, evolving systems. You can find that recap here.