AUTOGENAI: Contextual Colleague

February 2025

Designing a trusted colleague: embedding conversational AI into workflows to simplify bid writing and boost retention.

01 / Project overview

At the start of 2025, AutogenAI was facing some difficult product realities. Clients, particularly SMEs, were struggling with complexity: too many ways to do the same thing, a navigation model that often confused new users, and outputs that didn’t measure up to the likes of ChatGPT, Copilot, and Claude. The problem wasn’t just usability it was business critical. SME churn was hovering around 30%, NRR was stuck at 98%, and the company had set a target of 104% to get back on a sustainable growth track.

In January, I ran a design sprint on platform consolidation. The question I framed was simple: how might we create one clear, consistent way for users to engage with AI without adding yet another layer of complexity? Over a week, we explored the problem, ideated around different approaches and converged on a vision. Early prototypes tested with a user group showed us that embedding a conversational assistant directly into workflows dramatically reduced confusion and made the product feel more natural and responsive.

As part of the sprint, I also developed a strategy for rolling out the assistant incrementally. The first step would be to evolve a current feature (Research Assistant), moving it from a basic Q&A feature into something conversational, recursive, and agent-powered. The second step was to consolidate overlapping tools such as Ideator, Document Extract, and Document Comparison into Research Assistant, reducing the number of entry points and unifying the experience. From there, we could extend the assistant into the Editor, a natural place to bring in context and memory, before finally expanding it across the wider application.

When I roadshowed these ideas with around ten clients, the response was overwhelmingly positive. They immediately saw the value of a colleague-like assistant that would guide them step by step, live inside their workflow, and learn from their organisation’s voice and history of successful bids.

Not everyone was convinced internally. The CEO and founder was openly sceptical. His preference was to keep the product as simple as possible ideally hiding all functionality behind traditional UI. While elegant in theory, that model didn’t fit the complexity of bid writing, where users often need to iterate, review, and refine their way to an answer. A traditional UI couldn’t give them the control or transparency they needed.

User research provided a counterpoint. Clients were already comparing us to ChatGPT and Copilot, and were frustrated that our tools felt clunky in comparison. They expected a conversational interface because it had become the dominant way to interact with generative AI. One of our largest enterprise clients, went so far as to say they wouldn’t renew without chat in the product. This combination of evidence and market pressure shifted the CEO’s view, creating alignment around a conversational, workflow-aware assistant.

In short, Contextual Colleague wasn’t just a design improvement. It was a retention and renewals strategy. By consolidating tools into a single, familiar interface and acting as a persistent guide, it promised to reduce SME churn and move the business closer to its NRR goal of 104%.

02 / Phase 1 – Homepage Experiment

Before building Contextual Colleague, we ran an experiment to quickly test the hypothesis that a consolidated single entry point could reduce confusion by repurposing Research Assistant as the homepage. This experiment aimed to make it easier for users to start, resume, or complete work without bouncing between tools.

Through async surveys, user testing, and moderated interviews, we found clear signals that this approach improved navigation and reduced friction. Most importantly, the new homepage adoption produced statistically significant increases in engagement compared to the old homepage : more active days, more transformations, and longer time spent in the app. This gave us confidence that contextual entry points were the right path forward.

03/ Phase 2 – Contextual Colleague Design

Designing Contextual Colleague required thinking beyond interface patterns or conversational tone — it demanded a systems approach. We needed to define how the assistant understands the world around it, how it remembers what matters, and how it shifts modes to support users in different parts of the bid workflow. These three foundations Context, Memory, and Roles work together to give the assistant intelligence, continuity, and capability.

Working closely with the Product Manager, AI Researcher, and Prompt Engineer, I helped define our initial approach. Together, we mapped how each capability would evolve incrementally, starting with lightweight experiments to validate user value before adding complexity.

Context Design — Giving the Assistant Situational Awareness

Context is what allows the assistant to understand where it is, what the user is doing, and what matters most at that moment. We designed context as a layered awareness system that pulls information from three main sources:

Automatic context, derived directly from the RFP and associated documents (requirements, deadlines, word limits).
Project-level context, added by bid managers to capture key terminology, win themes, and tone of voice.
Session context, introduced by writers as they work — uploading documents, adding links, or refining instructions.

Context flows automatically into the assistant’s reasoning so that users don’t need to restate it. When a writer asks, “Can you review this answer against the data security requirement?” the assistant already knows which section of the RFP, which requirement, and which supporting materials to reference. It can explain its reasoning, cite the sources used, and surface relevant context when needed.

Memory Design — Retaining Tone, Past Decisions, and Bid History  

A key challenge in designing Contextual Colleague was creating a sense of continuity — the feeling that the assistant “remembers” the bid, understands the tone, and learns from previous interactions. We approached memory as a layered system rather than a single source of truth. Each layer captures different kinds of information and has distinct visibility rules to ensure both usefulness and transparency.

At a project level, memory holds contextual details such as win themes, terminology, client preferences, and stylistic tone. This ensures that every section of a proposal feels cohesive, even when written by multiple contributors.

At a session level, it retains short-term interactions — decisions that were made, responses that were approved or revised, and key points raised in conversation. 

The goal wasn’t to make the system remember everything, but to help it remember what matters. Memory supports consistency and saves time, but always within clear boundaries. Users can see what’s being recalled, override it, or clear it entirely. In this way, memory makes the assistant feel like a familiar colleague — aware of past choices, aligned to the project’s voice, and attentive to the context of the current task.

Assistant Roles — Defining Orchestration Between Writer, Reviewer, Researcher, and Bid Manager  

Rather than creating a single “all-purpose” AI assistant, we designed Contextual Colleague as an orchestrated system of specialised roles — each representing a different mode of expertise. These roles work together behind the scenes but share a single, consistent persona.  

The Writer focuses on structure and persuasion, helping bid teams craft responses that are clear, coherent, and compelling.

The Reviewer checks for compliance, tone, and accuracy, identifying gaps or inconsistencies.

The Researcher retrieves and synthesises relevant information from past bids and source documents to strengthen arguments with evidence.

Finally, the Bid Manager role maintains an overview of progress — tracking deadlines, aligning outputs to win themes, and ensuring the proposal meets its strategic goals.  

This multi-agent approach allows Contextual Colleague to feel both capable and coherent. Orchestration logic determines which role takes the lead depending on user intent and workflow stage. Users don’t need to switch roles manually; the assistant adapts seamlessly. The result is one consistent colleague with multiple dimensions of expertise — a design that brings structure and clarity to complex, collaborative work.

04/ Conversational modelling

Together, context, memory, and role orchestration form the cognitive architecture of Contextual Colleague — the system’s understanding of what’s happening, what has happened, and what needs to happen next. But intelligence alone doesn’t create trust. The next stage of the design focused on how that intelligence expresses itself: the assistant’s personality, tone, and conversational behaviour. We moved from designing the system’s mind to designing its manner — defining how a capable system could also feel credible, professional, and human to work with.

Designing Contextual Colleague’s conversations required a hybrid approach.

We needed the flexibility of a large language model — so the assistant could respond naturally, adapt to different tasks, and handle open questions — but we also needed structure to keep the interaction purposeful, explainable, and aligned to workflow logic.

The result was a guided conversational framework: the LLM generates freely within defined boundaries — steps, roles, and prompts that give the dialogue shape and predictability.
Hybrid approach: combine the flexibility of an LLM with a designed conversational framework.

LLM layer: allows open, natural dialogue so users can phrase things their own way.

Framework layer: gives the exchange direction and boundaries.

Beyond that, I designed for explainability and traceability. CoCo always shows which documents it is using, explains its reasoning in natural language, and reveals agent planning steps when needed (e.g. “Step 1: extracting requirements… Step 2: mapping to past bids”). Checkpoints ensure users stay in control. This turns CoCo from a black box into a transparent partner.

05/ Defining the Assistant Persona

Defining Contextual Colleague’s persona was about more than choosing adjectives — it was about creating a consistent, credible voice that reflected the professionalism of the people it supports. Working with our AI researcher, product manager, and writers, we ran a collaborative workshop to define the assistant’s personality traits, role, and tone principles. Together, we built a persona canvas that described who CoCo is in the context of a bid team — calm, curious, concise, and confident — a colleague who listens first and guides with clarity. From there, we modelled language through sample utterances, testing how each trait translated into real dialogue. To make tone and pacing tangible, we used 11Labs to generate voice samples in different styles — adjusting warmth, rhythm, and inflection until the delivery matched the intended personality. These artefacts were then added to our Design System as a shared reference for designers, writers, and engineers, ensuring CoCo speaks with one consistent voice wherever it appears.

06/ UI experiments

Alongside conversation design, I prototyped high-fidelity UI experiments in Figma Make, sometimes connected to an LLM to create near-real experiences. Two areas of focus stood out.

First, document sourcing. The original UI for selecting and applying sources was clunky and unclear. I integrated source selection and visibility into the chat interface so users could always see what CoCo was drawing on.

Second, placement. I explored whether the assistant should live in a side rail, as a floating widget, or embedded within the editor, and how it should coexist with existing one-click transformations like “summarise” or “expand.” These experiments showed that the challenge wasn’t just functionality but trust: users needed clarity about what the assistant was doing and why.

07/ Evalution framework

To track progress, I developed a framework for evaluating both output quality and user experience.

On the output side, we combined human evaluations, automated checks, and LLM-judges. We looked for hallucinations, tone alignment, correctness, readability (measured with Flesch–Kincaid), and task-specific quality such as summarisation and retrieval relevance. Crucially, we also measured explainability and traceability: could users see which sources were used, understand the assistant’s reasoning, and follow checkpoints from skeleton to final draft?

On the UX side, we measured task success rates, iteration efficiency, trust, adoption, persona fit, learnability, and how well CoCo coexisted with traditional UI. These metrics gave us a dual lens: not just what CoCo produced, but how it felt to work with it.

08/ Insights and impact

The research and prototyping made several things clear. Contextual entry points increased engagement, as the Research Assistant experiment demonstrated. Users wanted chat not as a gimmick but as the predominant pattern for working with AI. Context and memory needed to work together. Agents worked best when orchestrated invisibly. And above all, the assistant needed to feel like a colleague: competent, proactive, transparent, and professional.

For the business, CoCo promised to consolidate complexity, reduce SME churn, and move NRR from 98% toward the 104% target. For users, it promised a single, trustworthy partner embedded directly in their workflows.

The assistant is now embedded directly in the editor, combining conversational interaction with contextual awareness.
We’ve implemented context and early-stage memory, enabling workflow-aware responses and continuity between sessions.
Conducted user interviews and in-product testing — early feedback shows strong engagement and trust in the new experience.
Scorecard evaluations show measurable improvements in clarity, accuracy, and transparency.
Next phases will expand memory, source provenance, and explainability — completing the vision of a truly contextual colleague.