Key takeaways_

• The best AI agent use cases start with a real workflow problem, especially where tasks are high-volume, repetitive and time sensitive.
• Our agent success came from mapping the process before building anything and keeping the scope tight.
• The result was a practical operational gain: faster lead response, less manual effort and more time for sales teams to focus on qualified conversations.

Like generative AI before it, agentic AI is everywhere. As it gets hyped, more organisations are under pressure to test agents in their workflows and generate results.

But this is easier said than done. Often, the difficulty is finding the right use cases. You need something that is possible to turn into an automated workflow, while solving a genuine business challenge. This is where value is found.

And once you have a use case, you need to know how to move it from vision to reality.

This is a problem we faced ourselves. But by identifying pains in our business and experimenting with agentic AI, we ended up creating an agent that’s already bringing results. In this blog, discover how we came up the idea and brought it to life – and our top tips for doing the same in your organisation.

The problem: lead handling was slowing us down_

We never lacked leads when it came to our Business Central work, but this meant we often had to juggle many conversations at one time, largely for our sales and marketing teams.

A few issues compounded this. Lead data was fragmented across multiple systems, from HubSpot and Dynamics 365 to spreadsheets and partner tools, with no consistent way to manage or qualify them end-to-end. This created a workflow that was heavily manual, disconnected and difficult to scale.

On top of this, there was a sequence of effort that every lead required, covering research to qualification.

In practice, that meant 45-60 minutes of work per lead before a meaningful sales conversation even began. Multiply that across a high volume of inbound enquiries and the impact was clear: sales development representatives (SDRs) were spending more time on administration and triage than on actual selling.

This created three core issues:

• Slow and inconsistent response times: Follow-up depended on team capacity, meaning delays in the most critical early engagement window
• High manual effort for low-value tasks: Time was being spent on repetitive research and outreach rather than high-impact activity
• Scaling meant adding headcount: The only way to handle more leads was to add more resource, particularly at the junior level

As volume increased, so did the inefficiency. But agentic AI served a useful way to reduce the friction.

The idea: a lead qualification agent_

With a clear challenge, we sought to create an AI agent that would streamline qualification and save time without dropping quality. This was tied specifically to our BCQuick package, focusing specifically on leads with more standard requirements for their ERP.

Instead of trying to redesign the entire process, we focused on replicating the first stage of SDR work: the repetitive, admin-heavy tasks that happen before a real sales conversation begins.

That meant designing the agent to handle a very specific set of responsibilities, as per the sequence of activity a human SDR would do:

• Research the organisation
• Assess whether it fits our ideal customer profile
• Qualify the organisation fit against the BCQuick proposition
• Initiate outreach and start the conversation
• Move qualified leads into the next stage of the process

The agent would effectively act as a junior resource doing first-stage triage. This would remove the groundwork for our human sellers, allowing them to focus on the post-qualification stages.

This framing was important. It kept the scope focused, avoided over-automation and ensured the output was genuinely useful to the sales team rather than disruptive to it.

It also made the use case a strong fit for an agent approach. The work being automated was:

• High volume – a constant flow of inbound leads
• Repetitive – the same research and qualification steps every time
• Structured enough to model – clear criteria for fit, clear next actions
• Time-sensitive – speed of response directly impacts engagement

As it was aimed at BCQuick, a fixed scope project, there was also less nuance that would complicate the agent. This made it a worthy AI agent use case.

How we mapped the agent_

Many AI agent projects fail when the process behind them isn’t clearly defined. Because of this, we deliberately began by mapping the workflow end-to-end, based on how leads were actually handled in reality. The aim was to understand what needed to happen, in what order and where an agent could remove friction without breaking the experience.

At its simplest, the process breaks down into a series of stages:

• Trigger: a new lead is submitted
• Inputs: lead form data, source, context and any available intent signals
• Decisions:
  • Is this a good fit?
  • How urgent is it?
  • Where does it need to go next?
• Actions:
  • Draft an initial response
  • Route or assign the lead
  • Log activity and context in the CRM

Once this was defined, the agent had a clear structured sequence of steps to execute.

Design principles_

With a clear workflow, we also applied some key design principles that would keep the agent focused on the task at hand and performing well.

• Keep the scope tight: Making it clear this was one workflow, covering lead triage and early engagement, not cross-sales automation
• Define what the agent should not do: Avoiding edge cases and overreach was just as important as defining capability
• Introduce human checkpoints where it matters: The agent progresses the process but doesn’t replace final judgement
• Prioritise usable outputs: Everything it produces (research, responses, qualification) needs to be immediately usable, not something that requires rework

By clearly defining the process and setting up guardrails, we could build an AI agent that did the work to the right quality and aligned with our existing handoff processes for leads.

Turning the workflow into an operational agent_

Once the workflow was clearly defined, the focus shifted to making it operational: connecting the right systems, structuring the logic and ensuring it worked in a real environment rather than just on paper.

At a foundational level, this depended on connected systems. Lead capture and tracking sat within CRM, while communication (initial outreach and engagement) was executed through integrated channels. Because the data was already aligned across platforms, thanks to our integrated tech stack, the agent could operate with the context it needed rather than starting from scratch.

From there, we translated the mapped workflow into something executable. The agent logic was structured around:

• Clear qualification criteria, aligned to ideal customer profile definitions
• Predefined response frameworks, based on how SDRs would typically engage
• Routing rules to determine what happens next and where a lead should go

This ensured the agent was operating within a defined, repeatable structure. Responses were context-aware, drawing on company research and internal data rather than generating generic replies. At the same time, guardrails were built in to reduce the risk of incorrect or off-brand outputs, particularly in early interactions.

Equally important was the ownership model. While the agent handles the early-stage work, we made sure humans remained responsible for progressing and closing opportunities, ensuring quality and accountability aren’t lost. It is at this point, after initial qualification, our sales teams will intervene to ensure criteria have been met.

The results_

The impact was immediate because the problem was clear. By removing manual effort at the start of the process, the agent changed how quickly and consistently leads were handled, leading to:

• 1 hour saved per lead: Manual research, qualification and initial response drafting were significantly reduced, freeing up SDR time for higher-value work

• Guaranteed 15-minute response time: Every lead is engaged quickly and consistently, regardless of team availability or workload
• Improved sales focus: SDRs spend more time on qualified, high-intent conversations instead of early-stage administration
• Better lead experience: Faster, more relevant responses improve first impressions and early engagement
• More consistent messaging: Responses are aligned to defined positioning and qualification criteria, reducing variation in quality

This is consistency at scale. Every lead gets the same level of attention, without increasing the burden on the team.

How to successfully apply AI agents in real workflows_

By going through the process of creating our AI agent – as well as helping our own clients get to grips with agentic AI – we’ve learnt some crucial lessons. These are our best practice tips if you’re looking to experiment:

1. Start with the process, not the technology_

The most common failure pattern is jumping straight into tools: choosing a platform or testing a model. But first, you need to understand what the agent is actually supposed to do. This usually leads to something that works technically but doesn’t solve a meaningful problem.

Instead, start by breaking down the real-world workflow as it exists today. Look at what actually happens when a task is completed by a human:

• What triggers it?
• What information do they use?
• What decisions do they make, and why?

Once that process is visible, it becomes much easier to identify where AI adds value.

2. Focus on high-volume, repeatable workflows_

The strongest use cases for AI agents are processes that are:

• Repetitive but still require some judgement
• Time-sensitive (where speed directly impacts outcomes)
• Structured enough to guide decisions

Search for these within your business, gathering feedback from internal teams about their daily pains. This will give you a wealth of ideas to pursue further.

3. Define the workflow clearly before building anything_

Mapping the workflow first makes build much easier. At minimum, map:

• Trigger (what starts the process)
• Inputs (data, context, signals)
• Decisions (qualification, routing, prioritisation)
• Actions (responses, updates, handoffs)

4. Start small and keep scope tight_

One of the fastest ways for an AI agent project to stall is trying to solve too much at once. That’s why it’s far more effective to focus on a single, well-defined workflow: something narrow enough to fully understand end-to-end, but valuable enough to deliver measurable impact. For example, lead triage rather than full sales automation, or first response rather than the entire customer journey.

Starting small reduces complexity and makes success achievable. It allows you to test how the agent behaves in a controlled environment, validate that it integrates properly with existing systems and ensure the outputs are genuinely useful to the business. Most importantly, it gives you something tangible: a workflow that is faster, more consistent, and less reliant on manual effort.

5. Use the systems and data you already have_

AI agents rely on the quality, accessibility and structure of the data around them. So, the most effective place to start is by leveraging what already exists: your CRM, your sales and marketing content, your defined processes and your customer data. In most cases, organisations already have what they need; they just haven’t connected it in a way that’s usable.

Existing systems like CRM platforms already hold critical context. When fed into an agent, this becomes the foundation for consistent, context-aware outputs rather than generic responses.

6. Design for human involvement_

One of the biggest misconceptions around AI agents is that the goal should be full automation. In reality, the most effective implementations remove manual effort from the process while preserving the points where human judgement, context or relationship-building adds real value.

This starts with identifying where human input is essential. In most workflows, that’s at key decision points: final qualification, complex conversations or moments where nuance and experience matter. By designing clear handoff points, the agent can progress work to a certain stage (e.g. research, initial response, qualification) before passing it to a person with the right context already in place.

7. Build with guardrails from the start_

The most effective agents are constrained by design, with clear boundaries around what they should do and, just as importantly, what they should not attempt.

This means defining the scope in practical terms. What types of inputs should the agent respond to? What decisions is it allowed to make? Where should it stop and either escalate or hand off to a human? For example, an agent might be responsible for initial research and outreach, but not for handling complex objections or negotiating terms. These boundaries ensure the agent operates within a safe, predictable range.

Putting guardrails in place early reduces the likelihood of poor outputs, off-brand responses or incorrect actions.

8. Measure impact in operational terms_

One of the key reasons AI initiatives fail to gain traction is that their value isn’t clearly measured in ways the business actually cares about. Success needs to be framed in operational outcomes: how the agent improves speed, efficiency and consistency in real workflows.

The most effective approach is to anchor measurement to the process you’re improving. For example, if the agent is handling lead triage, the right questions become: how much time is saved per lead? How quickly are leads being responded to now compared to before? Has the quality or consistency of responses improved? These are tangible, observable changes that teams can immediately understand and validate.

From use case to scalable capability_

What this case study shows is that success doesn’t come from experimenting broadly or chasing use cases in isolation. It comes from identifying a clear operational bottleneck, mapping the process behind it and applying AI in a focused, controlled way. That’s what turns an idea into something that actually works.

By being able to do this once, you unlock a greater opportunity: a repeatable approach you can apply across the organisation to find friction, define the process and apply AI where it adds value. This brings scale.

If you want to see how this looks in practice across multiple real-world use cases, explore our Customer Zero hub. Inspired by our own digital transformation to an all-Microsoft tech stack, it contains useful guidance to improve your foundations, stay safe and make a pathway for value-adding AI.