The Gen AI Roadmap: From Idea to Full Automation

Introduction

You know that Generative AI is set to revolutionise industries by automating complex tasks and enhancing decision-making processes however it is one thing to know it, another to deliver the solutions in your business. You are not alone, we chat daily to business leaders across different industries, and the struggle is real in how to go from an initial idea, to the delivery of an effective AI solution. There has been no clear roadmap for this next frontier of AI.  

At Kortical we have been delivering AI solutions since 2016 and in the past year and a half we have adapted our AI roadmap process to Gen AI projects with huge successes.  Reed, Charlotte Tilbury, Deloitte and many other great companies, use Kortical to power their automation solutions and deliver big commercial gains with our platform. 

We have distilled the main learnings into this article with the aim to inform and inspire, guiding you through a comprehensive Gen AI roadmap from ideation to the deployment of AI agents capable of automating full job roles.

‍

Understanding the Gen AI Roadmap

The journey from an AI concept to a fully automated generative AI solution involves several critical stages: ideation, proof of concept (PoC), evaluation, development, and deployment. Each stage presents unique challenges and opportunities, however if you can get this right then you will be able to unlock huge potential and create groundbreaking generative AI tools.

‍

Stage 1: Ideation to Map Out Gen AI Use Cases

Host an Ideation Workshop

The first step is to generate and refine ideas. Conducting a structured workshop can help identify the highest value AI opportunities that are also easiest to implement. I know some companies are fully remote but if you can do this face to face it will be a lot more productive. 

During the workshop, encourage cross-functional teams to brainstorm potential AI applications that address key business challenges.

Who should be there I hear you ask, the best group we have seen and you can adapt this for your business often looks like this: 

• those that feel the pain (do the job day to day), there could be a few different teams represented in here
• those that manage the budget and will get excited about the cost savings or revenue generating value (senior manager), 
• data person who knows what the data looks like, where it is stored and how to access it for training data (might be the same person in the day to day job) 
• An open minded technical person who will be able to discuss security, integrations, API’s 
• AI / LLM lead - might be a data scientist or someone who knows the limitations of large language models and feels comfortable with prompt engineering - you could consider outsourcing this resource if there is no one in house
• Project manager - might be the day to day person / technical person / AI lead 

‍

‍

Identify High-Impact Use Cases - Be Brave

Map the high value use cases that align with your strategic goals, against those that are easier to implement from an AI delivery point of view. For example a generative AI use case that reviews internal documents and then has another step where it searches the internet for information has a lot of moving parts and will be harder to fully deliver vs a FAQ bot that reads from a knowledge base and can respond instantly in many languages.

Consider starting with a lighthouse project—an impactful yet manageable initiative that can demonstrate the value of generative AI to stakeholders.  This approach will ensure that you get buy-in from your business and will create some buzz to encourage more use cases to be implemented vs a toy use case that is quite niche where most people are not that interested in.

‍

Get this template for free

If you would like the template for the above ranking and mapping then please click this link and you can copy it for yourself :) And if you have any questions please ask us on support@kortical.com.

‍

Co-Pilot to AI Agents

There will be a range of ideas that generally fit in these 3 camps:

Co-Pilot - help someone do their job better, like those AI systems in Github or Microsoft applications helping write better code or helping write AI generated content like emails.

Prediction - this is often referred to as traditional machine learning (ML), for example AI models that forecast sales or supply chain materials more accurately.

Automation / AI Agents - these can take over parts of roles or in some cases full job roles and while they sound super ambitious they are fully achievable with the technology available today - check out this video for more detail.  To start you may have an AI chatbot answering customer queries that is hooked into your back-end to know where your orders are, that can then be hooked into more systems like product catalogues and sizing charts to be able to be a sales rep also.  You can chain more generative AI models together, doing specific tasks to deliver more automation until you have a full role automated with generative AI applications. 

The ideas here are endless, where it could be creating job postings, content generation for blogs and social media, doing tax accounts and finance tasks etc…. So the way to start is to think, what is the simplest step to prove an AI agent can work for our business and then build from there. 

‍

Roadmap Generation

To get to a starting point strip each use case back to its single most useful part e.g. for IT support, you may start with automating the frequently asked questions, then stage 2 is automating password reset by integrating with your access software, stage 3 is ability to order new hardware etc… 

Using your mapping as your guide, you will start with the high value easier to implement solution first and list off the next in order of the feasibility and potential impact of each proposed AI application. Consider factors such as data availability, technical complexity, AI capabilities that you can access and very importantly the expected business benefits. Use this assessment to prioritise projects that offer the highest return on investment and are relatively easy to implement.

Create a timings list to hold the team accountable so that decisions are made in a timely manner and the projects can quickly get started. Agree on a budget and get going! 

‍

‍

Stage 2: Proof of Concept (PoC)

‍

Develop a PoC Within 1-4 Weeks

Getting a working proof of concept is the best way to start gaining momentum and with the latest AI tools out there this is now easier than ever. Aim to develop the first PoC within 1-4 weeks, depending on the time the key delivery people have to spend on it and the complexity of the project. This rapid prototyping phase should focus on demonstrating the core capabilities of the generative AI system using a subset of your data.

Start lean, by using the single most useful part of the use case and parking integrations for a later phase then this will enable you to move fast and get a working bot into people’s hands to test. 

‍

Utilise Trusted AI Frameworks

Leverage established AI frameworks and technologies to accelerate your PoC development. OpenAI and Azure’s instance of OpenAI provide robust platforms for building and testing AI models while ensuring data security. Additionally, KorticalChat offers an AI agent framework that includes tools for data ingestion, a range of large language models, GDPR compliant LLMs, fine tuning, prompt engineering, and chatbot UI so no need to plumb them all together you can use an off the shelf platform.  There is no excuse not to get a working PoC delivered quickly that meets the security goals and can demonstrate the value to users. Yes that sounds a little harsh but honestly we hear so many excuses why a use case has not been delivered, but it need not be the case and I am here to tell it can be done!  

‍

‍

Iterate Rapidly

The goal of the PoC phase is to quickly iterate and refine your AI solution so that people are able to play around with it. Use feedback from stakeholders those inside and outside of the project as well as the initial users, to make necessary adjustments and improvements. This iterative approach helps ensure that you are moving towards a valuable solution and you are getting wider views that even if you don’t incorporate them in the POC phase you are able to capture those and add them to the roadmap. 

‍

‍

Stage 3: Evaluation

‍

Clear Objectives - Limit Scope Creep

There will quickly be scope creep, people just can’t help themselves as they always want AI to take on more of what humans are doing.  So from the outset define the use case and clear success criteria for your PoC. This might include metrics such as accuracy, efficiency, and user satisfaction. Regularly evaluate the PoC against these criteria to determine whether it is ready to move to the next stage of development. These often are most referred to documents as it helps keep everyone on the same page, as we said we would achieve x and look this delivers x in spades.

‍

Measure Success - Compare the Old vs the AI

Using comparison benchmarks between the old way and the generative AI process is very helpful, as people may say that creating that document only takes 5 mins but when they actually time themselves and note the steps the reality may be a lot longer. It is also helpful to keep a track of the success metrics and see how they change over time as you add more generative AI capabilities.  

‍

Think Long Term

Sometimes the first use case is not a game changer but it is a step that everyone can buy into, it is low risk and/or involves few integrations but what it is, is a showcase into the art of the possible.  A common example is a generative AI app that helps writes personalised sales emails and through measured testing it proves out it will save 500 sales employees 2 hours a week. That cost saving may not be huge however the impact could be much greater as with that free time the sales team will be able to spend time on activities generating more revenue so while this solution may not be ground breaking, the flywheel has started and from this seed more trees will grow. 

‍

‍

Stage 4: Development

‍

Iterate and Expand

Once the PoC is validated, the next step is to iterate on the Gen AI solution, incorporating feedback and expanding its capabilities. Focus on building a scalable architecture that can handle increasing volumes of data and user interactions.

‍

Overcome Common Challenges

‍

• Access to Data: While data has been the backbone of AI with Gen AI solutions you often need less data than anticipated for initial LLM prompts. A few best in class examples can sometimes generalise quite well.  If you are thinking of a bot that knows your business well or the market very well then you may want to consider gathering a more comprehensive dataset for building a robust RAG (Retrieval-Augmented Generation) system. This will mean that it can become a specialist in that data, like HR onboarding, you could upload 1000’s of files relating to HR frequently asked questions, host a chatbot on your intranet and your employees can use that to help self-serve 80% of their queries.

‍

• Integrations: This is where a lot of the complexity and timescale spiralling can come in and while a fully integrated system may be your end goal if you can, our advice is to avoid integrating all systems at once. Identify the valuable parts of the AI solution that can function independently in the first phase.  Then look to incorporate more over time.

‍

• Data Security: We see this blocker used a lot, however there are ways to stay secure and to drive AI innovation.  One could be to use private instances of Azure OpenAI, Google Gemini or whatever cloud provider to keep your data within your private cloud.  If that is too risky then consider starting with a use case or a part of it that uses public-facing data to mitigate security concerns or try synthetic data.  All the while ensuring that the Data Protection Officer within your organisation has access to education on best practises for Gen AI and data security so that over time you can be both secure and innovative.

‍

• Leadership Buy-In: Depending on the background of your leadership team they may be gung ho on AI and innovation projects or more typically excited but read the headlines on AI going rogue so are more cautious.  For those that need more convincing often it is best to start with internally facing low-risk projects, such as internal virtual assistants for HR, generate content bots with a manual review stage, all with an ROI, to build confidence among senior leaders.

‍

Implement Robust Testing Procedures

As with any new technology testing of your AI solution is critical to ensure the reliability and usability of your solution. Develop comprehensive testing procedures that include creating a baseline across different outputs and consider how you are going to sign off new developments coupled with user acceptance tests. This helps identify and address potential issues before they impact end users.

With the rapid development of new Large Language Models, which are often better at understanding but also come with new cost and speed dynamics, if you spend some time building your testing framework and nailing what matters to you, then testing new models will be much easier to onboard or dismiss over time.  

‍

‍

Stage 5: Deployment

‍

Change Management

‍

By now your Gen AI solution is fully tested, refined and your leadership team is fully bought in so it is time for deployment.  However great your solution is, as with any new tech, if people do not use it then it is a complete waste of time.  Successful deployment of AI solutions often requires effective change management. Here are some strategies to facilitate a smooth rollout:

‍

1. Make it easy to find: Where is this new technology going to live, how close to the current workflow can we put it e.g. it is a chatbot on the landing page of the intranet, full integrated into your CRM or a new webpage with all your AI agents in one place clearly signposted. 
2. Stakeholder Engagement: This is a no-brainer to engage with stakeholders early and often to manage expectations and address concerns.  
3. Training and Support: Provide comprehensive training and support to users to ensure they are comfortable with the new AI tool. When you are implementing generative AI solutions, think demo videos, webinars, step by step guides, however the solution should be intuitive enough that this is a quick guide vs a day training exercise.  
4. Communication: Communicate the benefits and expected outcomes, any risks or ethical considerations, as Gen AI tools may need manual review stages etc... to all relevant parties to build support, ensure best practise and also and probably most importantly build enthusiasm.  And ongoing let the users know updates, share results, keep the feedback loop open. 

‍

Performance Monitoring and Optimisation

One thing I know for sure the world does not stay still, your customer base will change, your products evolve, new competition may launch so continuous monitoring and optimisation are essential to maintain the performance of your Gen AI solution. Establish a monitoring framework that tracks key performance metrics and use this data to identify areas for improvement. Regularly retrain and update the generative AI models to ensure they continue to meet your business needs.

‍

Ensure Robust Support and Maintenance

Deployment is not the end of the journey. Ongoing support and maintenance are essential to ensure the long-term success of your Gen AI solution. Ensure someone is responsible for checking the latest generative AI models releases, testing them against your framework and understanding if there is a business benefit to deploying these new large language models e.g. a larger context window, or quicker speed, or cheaper cost may prove really valuable so having someone keeping an eye and quickly testing makes it easy to take advantage of them.  

‍

‍

Finally: Keeping ahead of the curve 

‍

Looking Glass into the User’s Mind

As you get a history of what every user types into the chatbot or Gen AI application, this is the holy grail of setting the direction of what to develop next.  You may start with a HR onboarding bot and see that through the questions they are asking the chatbot that the managers want a tool to help them with delivering effective employee feedback.  You also see that they are looking for job interview questions for a broad range of roles so you create a chatbot that is specific for that. Then you may find that the team is asking questions on social media posts and what is compliant so you create a Gen AI chatbot that delivers linkedin posts for company announcements, job postings etc.. that aligns with brand values. 

This treasure trove of user interactions will help shape the roadmap and ensure that you create solutions that your team and/or customers actually need. This is so helpful and something that all our clients have jumped on to build their roadmaps. 

‍

Keep an eye on the future

The pace of change in this space is like nothing we have ever seen, they say what used to take 10 years in an innovation cycle is now taking one and this acceleration is only going to keep going.  What that means is that new AI models, new frameworks, new technologies, are being released every quarter and some will suit your use case better than others.  However don’t worry that everything will change, start today as it is easier to build upon new technology then stand still without any of the learnings that you get from trying. 

‍

‍

Get started with your own Gen AI solution today

This is a new exciting frontier where those close to the business problem are often best placed to solve it for themselves or their customers. 

It would be remiss of me if I didn’t mention KorticalChat as many customers are using our Gen AI platform and consultancy services (from Deloitte, to Reed, to Santander) which takes care of so many of the different AI technologies you need to get going with an easy chatbot builder, RAG, GDPR compliant models and the chatbot UI ready to go. 

Take a look at this video on how easy it is to build a chatbot with Kortical. 

And if you want to sign up or contact us please do, we love to chat about how to get you a Gen AI solution in weeks. 

‍