IN-HOUSE GPT SOFTWARE

A voluntary initiative was launched, involving interested employees, including mainly developers, product and myself - to create an in-house GPT tool for wider colleague assistance. Given the low resources, we conducted regular informal knowledge-sharing and departmental needs mapping sessions. Based on this discovery, we executed a soft launch using 3 core GPT tokens to gather essential business-wide feedback;

A general enquiries service
A secure tool for users to ask general business questions and find problem-solving support, without the risk of data leakage

A HR policy advisor
Where users could get advice about HR policies and principles

A contract assistant
Provides assistance on obtaining key information out of long contracts, or advice on the best negotiation points

Challenges

Operating voluntarily and outside core hours presented significant scheduling hurdles, leading to inconsistent team attendance at our twice weekly 8am stand-ups. This lack of consistent synchronisation meant we struggled to maintain a clear structure or agreed-upon priorities, resulting in work being ready at different times.

Learnings / What we could have done differently

Structure: Dedicated collaboration within working hours instead of relying on fragmented, out-of-hours catch-ups.
Governance: Implementing a stricter priority list and greater adherence to jira tracking would have maintained consistency and reduced rework.
Focus Balance: We must actively guard against prioritising the "exciting" elements over completing the critical, foundational (but less glamorous) tasks.

The development team and I utilised a data analytics dashboard in order to quickly identify user issues, provide adequate solutions and seek out common trends and themes for the GPT service usage.

Dashboard homepage:

Feedback dashboard shows all submitted feedback, capturing the user, context, specific issue, and the originating upvote or downvote

Analytics surrounding specific users and team usage, themes from chats and feedback data

Data surrounding data exported, content and query breakdowns and common topics identified

Manages GPT control, administering admin rights and access to all employee teams (Head Office, warehouse, and logistics)

User activity:

As GPT usage grew, we added the data token capability after identifying this feature was highly requested from the testers - including this resulted in higher user engagement. Usage trends generally fluctuate based on seasonal cycles, business objectives, and trading periods. We tracked the data to gauge the influx of users and their activity over the course of 30 days.

Total, active, and inactive user counts, along with their identification methodology

The results of the data token addition, in comparison to other tokens available

Other tokens added as per user requests included a marketing email assistant and a capitalisation assistant

Overall token usage

Improvement Strategy

The GPT was soft launched to volunteer user testers provisionally to get direct feedback. This was crucial for identifying problems and potential improvements prior to hard launching to the whole business. The main learnings were;

More help required: a lot of users struggled to navigate the prompts to get the end result they desired
Accessibility: users who were not on the company vpn or at the head office could not access the GPT
Inaccuracy: a lot of users were not checking the outcomes and were being fed inaccurate information
One-sided feedback: users would typically only provide feedback when something went wrong, which made it difficult to gauge what was appreciated
Feedback limitations: Users only voted after their first use with one-worded critique which made it near impossible to fix any problems.

Next steps

We produced training modules to help teach those unfamiliar with AI how to optimise their results
We improved the UI to include helper text and guidance throughout the user journey
We added in a notification to inform users about the accessibility issues whilst the devs figured out how to expand the service
The feedback mechanism was revised and the form was expanded to ensure the feedback was more valuable

WIREFRAME DEMO

We established the foundational elements for the MVP, prioritising immediate user needs and token capabilities for the planned soft launch. Initial UX elements included a password indicator and a required T&C acceptance on sign-up. The 3 GPT category options were designed for easy, navigable selection, pending the addition of clear helper text detailing each token's capability.

The prompt interface was designed with clear indicators for the three stages: provisional user input, generating response phase, and generated response phase. Following prompt generation, a feedback option was included, initiating a cycle that continued based on user input and satisfaction.

PROMPT JOURNEY SCREENS

GPT Sign Up Screen

Feedback mechanism included to gather improvements at this early stage

Notification informing users of accessibility of the GPT software

Users must check the box to accept the terms, which are available for review

Instructional header

1 / 6

FEEDBACK SCREENS

Feedback form

The feedback form opens in a sheet, providing helper text to guide submissions and input boxes for supporting comments

1 / 2

USER PROFILE SCREENS

New user profile

In the absence of new user data, the screen is limited to core configuration options: dark mode preferences and a sign out option.

1 / 2

View all work

See Site Redesigns