Rethinking process automation with AI

Hannah Colley00:03

Hello everyone, and welcome to today's webinar. I'm Hannah, digital marketing and events specialist here at FlowForma, and I'll be your host today. Before we get started, a few housekeeping points while people join. This webinar is being recorded and your lines are muted, but we want to make it as interactive as possible, so pop your questions into the question box at any time. We'll also run a couple of polls during the session and a short survey at the end. If you have any trouble logging in, email info@flowforma.com and someone will help you.

Hannah Colley01:05

Let's meet today's speakers: Niamh and Gerard. Gerard is FlowForma's product manager and Niamh is FlowForma's head of marketing. With that, I'll hand over to you both.

Niamh Lordan01:35

Thanks, Hannah, and hi everyone. For anyone new to FlowForma, we've been helping organizations rapidly automate processes without code for over 12 years. We're here today because Gerard and the product team have completely transformed our platform with AI features and functionality, and we'll show some of that during the session. AI is changing how work gets done, and judging by the volume of registrations, many of you are asking the same questions: how can I digitize faster, how can I rethink or re-engineer my processes with AI, and, more importantly, how do we keep control of people, processes, and agents?

Niamh Lordan02:54

Our goal is for everyone to leave with new tips and insights. Over the next 40 to 50 minutes we'll look at how to pinpoint the ideal process candidates for AI, and we'll run four mini demonstrations showing how you can accelerate process mapping, get processes live, and deploy agents to support your employees. Then we'll come back to the big challenge everyone shares: how to govern those agents.

Niamh Lordan04:10

Let's start with our first question, Gerard. Everyone's using AI, but are they using it in their processes? For people who want AI to speed up processes, with AI agents working inside them, how do they figure out the best ones to start with?

Gerard Newman04:35

Good question. Process automation has always been about automating work, making it more efficient, and keeping it moving. So with AI, the question is how it can be built into processes to do more of that. Many of our customers already have their big line-of-business systems in place: banks have their banking and lending systems, insurers have their claims management systems. What we look at a lot of the time is the processes that sit around those systems. These are the characteristics we look for when we work with customers to see where AI could make an impact.

Gerard Newman05:33

The first is unstructured information. Workflow is great at capturing documents and sending them to the right people, but people still have to read, review, and digest them. That's a big issue in areas like supplier onboarding and contract management. Given the right instructions, AI is great at ingesting large volumes of information in lots of formats, distilling it into a clear summary, answering questions on it, and making decisions. There are huge efficiencies to be gained there.

Gerard Newman06:23

The second is working between systems. In the past, particularly in lower-volume processes, people often parked integration with other business systems because it was too technically complicated or they didn't have the resources. AI makes that far more feasible. Integration using AI agents has become much easier, so there's an opportunity to enhance processes and how they connect with other systems.

Gerard Newman07:01

The third is chasing. We all know processes where there's two or three hours of actual work, but it takes 10 days because the rest of the time is spent chasing people for updates. Agents are great at that. They can nudge people to respond and keep things moving, and they're also good at monitoring processes so you can see what's falling behind. We'll look at that in one of our demos.

Gerard Newman07:38

The fourth is judgment calls. Given the right instructions, AI is good at assimilating lots of information and making an initial judgment. In a health and safety process, for example, you might have lots of incidents logged, but what you're really looking for are the serious ones that need significant effort to review and address. AI can do the initial triage so you can decide which items need more attention. That increases productivity and throughput, because not everything gets the same level of attention when it doesn't need it.

Gerard Newman08:28

And last but not least, exceptions. Exceptions take up a lot of time. AI can help eliminate some of them: by analyzing your process data, it can spot where exceptions occur too often. Processes that have been in place for a long time often carry exceptions created by events in the past, and sometimes those can simply be stripped out. AI can also identify exceptions earlier and help you handle them within the main line of the process rather than sending them off to a separate exception process.

Gerard Newman09:12

If you can spot a couple of these characteristics in a process, that's a good place to start introducing AI. And a recurring theme today is keeping AI within our control rather than working autonomously. All of these changes can be made with a human in the loop overseeing what AI delivers.

Niamh Lordan09:36

Thanks, Gerard. The opportunities are endless, so people can be unsure where to start and what to prioritize. These are really helpful tips. If you spot a few of these across your processes, that's where to start.

Gerard Newman09:53

Absolutely. That's where the opportunities lie.

Niamh Lordan09:56

We've just talked about where AI can make an impact in processes, so Hannah is launching a poll: what does rethinking your processes with AI mean for you? The votes are coming in, and there's a clear winner: freeing our people up for higher-value work. Any comments on that, Gerard?

Gerard Newman11:09

It's a really good use of AI, and it's a recurring theme in today's session. We'll show a number of solutions that let AI do some of the more detailed work, which frees you and other people up to do a different type of work.

Niamh Lordan11:29

It's great to see votes on every answer. The runner-up is that work moves faster, with less waiting and chasing, which goes back to your earlier tip.

Gerard Newman11:46

Absolutely. We all experience processes with a lot of wait time and not a lot of work time. AI can help with the chasing and follow-up, and potentially let you change the process so there's less waiting.

Niamh Lordan12:07

Let's get into one of the key topics for today. People on the webinar may have processes that have never been automated, or processes they've automated that don't use any AI, which are effectively legacy processes now. Can you share how AI has changed how fast we can map and automate processes?

Gerard Newman12:43

Using AI, we can understand a process a lot quicker. Anyone who's worked as a business analyst or process designer knows that a huge part of automating a process is understanding it: engaging with the business, documenting it, getting agreement, and fleshing it out into a working process rather than just a nice diagram. I'll show two approaches. The first uses AI-based tools to understand a process and draw a map of it, which suits process experts and business analysts. The second is a chat-driven way to build a process.

Gerard Newman13:57

This is a process mapping capability we've created. The point is not just to draw a diagram, but to design a process from start to finish in a way that's interactive and uses chat. I explained that I was working with the finance team to implement a structured workflow for supplier invoices, a process most people understand. The team receives about 100 invoices a month, directly into finance but also through other people in the business, and invoices are paid within their terms unless they're in dispute. That description wasn't onerous to write, but it set out several characteristics of the process.

Gerard Newman15:02

My request was played back to me to confirm the agents understood it, and then they asked me questions: what are the value bands for invoice approval, are PO-backed invoices handled differently, and who owns resolution when an invoice is disputed? Each question carries a percentage, which is the AI's confidence that the question makes sense and hasn't already been answered. I answered, it played that back and asked more, and after a few rounds the agents said they had enough information and asked whether to build the map.

Gerard Newman16:13

Here's what it built. A supplier invoice payment process sounds simple, but anyone who has implemented one knows there's more to it than a few steps. Behind this map, 12 agents work together, using my instructions and their own knowledge, to create the process in detail, and then they check each other's work. The synthesizer agent created 25 steps and identified 10 roles, 13 conditions, and 42 edge cases. Imagine the work for a business analyst to pull all of that out by hand.

Gerard Newman17:45

One agent I like is the judge agent, which rates the process. It gave this one 18.8 out of 20, calling it an excellent, complete, and believable supplier invoice payment workflow with only minor overspecification and evidence hygiene issues. It flags inferred operational details and where it drew on standard practice rather than my description, so I know which areas to look at in more detail.

Gerard Newman18:33

If I open the approval decision, I can see my payment approval levels, who the invoice routes to, and that the AI is 99% confident the branching logic is correct, as it should be, because I provided that information. I can also chat directly with this part of the process to change things like SLAs. For CEO approval of invoices over €2,000, it set a three-business-day SLA. I didn't specify that, and the judge flagged it, but it also decided how to notify the CEO and what happens if the SLA is missed. What you end up with is a fully defined process, including how exceptions, communications, service levels, and escalations are handled.

Gerard Newman20:23

That's great for process and business analysts who want that level of detail. AI is doing a lot of the heavy lifting in the background, but it may not be everybody's starting point when they want to implement process automation.

Niamh Lordan20:42

Sometimes prospects with a big digital initiative slow it down because they want their processes mapped first. This completely changes how quickly you can turn those maps around, and the skills and resources you need.

Gerard Newman21:07

It really does. You can start with chat, and the agents will use what you provide plus what they find is best practice. You can also give it your own procedure manuals and it will extract processes from them. It creates the process, and then you refine and improve it over time. It saves a lot of time.

Niamh Lordan21:38

We often hear, “We'll kick off our digital initiative once everything's mapped.” This helps with the objective we set at the start: moving faster and letting people do more value-added work.

Gerard Newman21:59

The next example is another way of creating processes. People are shifting to LLM-based chat interfaces and using Claude to do more and more of their work, so we wanted people to be able to build a process within Claude, without using any FlowForma-specific tools. I started by saying: we use FlowForma, and I need to set up a new process to onboard new employees. Claude has a skill we've developed, so it knows all about FlowForma and starts asking me questions.

Gerard Newman22:46

It's much more conversational than the previous example. It tells me what an onboarding process normally looks like, maps out six steps, and asks whether steps three to five run in parallel or in sequence, who signs off, and so on. I don't need to be a business analyst to answer those questions; I just need to know the process. I explained where our process differs: how we collect information about new joiners, and that we expect people to read and confirm certain documents when they start.

Gerard Newman23:42

It keeps asking understandable questions: the links to the handbook and security policies so they can be included, how we handle contractors versus permanent employees, what happens when problems arise, what the line manager and HR do, which teams are involved, how we communicate, and what records we keep. They're all questions you'd need to ask someone, or yourself if you're the subject matter expert.

Gerard Newman24:43

Then it develops a specification, which summarizes the conversation and the decisions made, so it can be shared and reviewed, and so you can see later how you arrived at this process. When I provided the document links, it spotted that one wasn't a sharing link, so people probably wouldn't be able to open it. I didn't ask it to check that; Claude spotted the issue itself.

Gerard Newman25:31

Then it builds the flow, fully configured to work in FlowForma, and I bring it in. Here's my new starter onboarding flow running in FlowForma: a new starter request, HR checks, IT setup, facilities, and steps that change depending on whether someone is an employee or a contractor, such as payroll setup versus contractor payment setup. So that's two ways in: a detailed mapping tool for process experts and business analysts, and a simple conversational tool that people are increasingly used to.

Niamh Lordan26:38

Great examples. It's about letting people build from where they already work, and so many people now live in chat interfaces. It's almost like an interview with a business analyst.

Gerard Newman26:54

It is. A big difficulty on any automation journey is getting people to engage and figuring out what your processes actually are. This takes you much further down that road: you can go to the business and say, “This is the process I think we have. Is it correct?” You have it prototyped in minutes.

Niamh Lordan27:20

Now we're on to my favorite part. Once your process is digitized, you can consider putting agents to work in it to make it more efficient, faster, and more cost-effective. Gerard, what can agents actually do within processes?

Gerard Newman27:50

Once your processes are in place, you add AI agents to them. Agents are workers within your process that can do tasks themselves, just as people do. They take in information as the process moves along, assimilate it, make decisions, create output, and integrate with other applications. You can add agents to existing processes, where they take over the lower-value tasks so people can concentrate on higher-value work. We strongly believe in humans working alongside agents and checking their output. Agents are simple to create, they're built into a process, and people stay in control of them.

Gerard Newman28:54

Here's an example from a product we have called FlowAssure. It automates vendor onboarding, which in most organizations is a structured process, particularly the security assessment. You ask the vendor for a set of evidence that the security controls you expect are actually in place. Here, the vendor has provided pen tests, security policies, incident response plans, and so on, and these are being evaluated against NIST CSF controls.

Gerard Newman30:00

Take identity management. NIST CSF says an access control or identity management policy should cover the provisioning and lifecycle of identities and credentials. Previously, a security analyst would assess that from a questionnaire or emailed documents. Here, an agent picks up all of the documentation in the background and assesses whether it's adequate evidence that the control is implemented to the right level in the vendor's organization.

Gerard Newman31:04

The agent's verdict is a pass, with medium confidence, and this is a good example of an agent working with humans. It explains its rationale, for example that the access control policy grants access on a role-based, least-privilege basis, and it references the exact source: page one of the security policy. It has extracted five referenced pieces of evidence and, on that basis, passed the control.

Gerard Newman32:16

At this point I can approve it, but I can always override. If I don't think the agent has done enough, or it failed a control I'd pass, I can change the verdict. It's doing what AI is great at, assimilating lots of information and evaluating it under very strict guidelines, while providing evidence for everything so a person can decide whether it got it right.

Niamh Lordan33:04

I love those buttons. The agents speed up the assessments, but we're still in control by approving, overriding, or asking for more clarification.

Gerard Newman33:17

The idea came about because security analysts put a huge amount of effort into reviewing this documentation, and there's a lot of subjectivity from one analyst to another on whether to pass or fail a control. We're using AI intelligently to speed the process up and make sure analysts focus on the right things.

Niamh Lordan33:49

Let's jump into a quick poll. How much are you comfortable letting an agent do in your process today: nothing yet, recommend but a person decides, act with a person's approval, or act and then report back? Not surprisingly, most people want the AI to act with a person's approval.

Gerard Newman34:46

That's fairly common. Most people want to use AI because they believe it can deliver productivity, but they want to use it on their terms and remain in control of it, which I think is essential. Like other automation tools, it helps you do your job, but you need to monitor it.

Niamh Lordan35:23

And once you build up that trust, you'd be happier for your agents to do more of the acting without your approval.

Gerard Newman35:34

I think so, and that's about how you onboard agents and monitor them. It's not a one-off. Traditional software is deterministic: once it's in place, it does the same thing every time. With AI there are other variables. Models change all the time, and that changes how agents make decisions, so they need continuous monitoring.

Niamh Lordan36:09

On to our last demonstration. You've shown how rapidly people can map and automate their processes, and where agents can get work done within them. The next big thing is governance: how do we see what agents are doing, what they're costing, and optimize all of that over time? Can you show us how FlowForma's agent trust platform supports that?

Gerard Newman36:45

This is an application we're working on, built on our heritage in process automation. Anyone who has implemented a process platform has dashboards showing key things about their processes. Here we have processes, people, and agents. The process view shows work in flight by department, items at risk of missing their SLA, overdue items, first-time-right versus rework, and so on.

Gerard Newman37:42

But in the background, AI agents are assessing this information and making recommendations. This example is an insurance broker that renews many of its policies at year-end, which is also peak annual leave, so capacity is dropping. One interesting recommendation is around priority inversion in claims: the simplest claims are being done first and complex claims are left in the queue longer. It suggests changing the priority order on claims intake and applying the same ordering to the first notification of loss triage agent, so work is taken in order of arrival rather than cherry-picked.

Gerard Newman39:16

Everyone then asks: what's the impact? It shows that high-priority SLA adherence goes from 71% to 86%, wait time drops from 9.4 hours to 4.1, and claims at risk in the queue fall from 96 to 74. It also shows the trade-offs: low-priority claims will wait longer, rising from about 11 hours to 15. So AI is enhancing the analysis of the process information that every process platform collects.

Gerard Newman40:13

The people view shows where there's spare capacity, who's falling behind, and who has a backlog and is about to go on leave. And we've made agents first-class citizens in the process environment, as visible to an operations manager or team lead as any member of staff. For each agent we can see items handled, items completed automatically, and items handed over to people, which is a key measure of how effective an agent is. Agents may hand off by design or because they can't complete something, so we track why, for example a policy data mismatch, where the data an agent looks up doesn't tie to the information provided. We can also see run cost by LLM model and more detail on each agent.

Gerard Newman42:13

A great example is our first notification of loss triage agent in insurance. It has a bad problem for an agent: it's confident and wrong. As far as the agent is concerned, it's right a huge percentage of the time, but sampling of its results shows it isn't. One option is to route all of its output to human review, raising the sample rate to 100% so everything it does is checked by a person. That's a reasonably safe option until the issue is fixed; perhaps the model changed and the agent no longer performs the way we want. Another option is to pause the agent, so work queues up behind it until the issue is addressed.

Gerard Newman43:34

The key point is that through a single pane of glass, you can see every element that delivers your process: the process itself, the people, and the agents. And with AI recommendations, you can improve the process in both its structure and its operation. That's what we're doing on overall agent trust and oversight.

Niamh Lordan44:09

It gives brilliant visibility. In the end, it's a lot like governing and managing people, but for your agents.

Gerard Newman44:24

That's the way to think about it. How would you manage people? The same things often apply to agents.

Niamh Lordan44:30

That's our last demonstration. Before the live Q&A, we have an offer for everyone. We don't want this to be something you just watched; we want it to be something you can act on. We're offering a one-to-one AI automation assessment: a 45-minute session with one of our experts, where you take what you learned and saw in today's demos and see how it applies in your own organization. Our team will give you a priority list of your automation and AI opportunities for the months ahead. If you'd prefer a free trial to test the tools Gerard demoed, we can set that up too.

Hannah Colley46:28

We have a few questions. We won't get to all of them today, but we'll follow up. Gerard, can AI improve a process we've had running for years, or is it only for new processes?

Gerard Newman46:44

AI can absolutely be used to improve existing processes. The Claude-based chat approach I showed will also take in existing processes and let you improve them. Processes that show any of the characteristics we discussed at the start are especially good candidates. Given how accessible and affordable AI now is, it makes a lot of sense to use it to further improve the efficiency and automation of your processes.

Hannah Colley47:22

How quickly can we change an agent's behavior if a policy or role changes?

Gerard Newman47:31

A lot depends on how your agents are defined and built. There's a lot of talk about kill switches, but I think each individual agent needs its own kill switch so you can shut it down. As I showed in the agent trust application, you need to be able to act quickly when agents start performing in unexpected ways, and there are lots of reasons they can. There's also a discipline in how you implement agents and bring them on board.

Hannah Colley48:08

How do you test an agent before letting it loose on real work?

Gerard Newman48:14

As you've seen in our demos, we're big fans of human in the loop, and that's the core of testing. To me, it's about implementing agents in shadow mode. They work alongside people, you look at 100% of their output, and you tune them to make them more effective and predictable. Then you let them take on certain types of task, so over time you build confidence and the agent demonstrates it can deliver. Test them thoroughly, and a good way to do that is shadow mode, where people are still doing the work, AI is supporting them, and they can see and assess what the agent is doing.

Hannah Colley49:17

One more question: “I like how you've followed market leaders like Salesforce in embracing Claude and chat interfaces, and released your own FlowForma one. Is it the end of SaaS-style interfaces, with users driving everything from Claude, Copilot, or ChatGPT?”

Gerard Newman49:43

It's very much down to users, but I believe chat is becoming a far more important interface, and people will expect to use it more and more. If you spend much of your day in Claude, Copilot, ChatGPT, or another LLM, you want to do your work from there rather than keep switching applications. The more connected that environment is, the more you can do from it. It's a trend we'll see more of, and it does reduce the unique selling points of individual application interfaces.

Hannah Colley50:21

That's all for today. Thank you all for joining us; we hope you found the session valuable. Please take a moment to complete our short three-question survey. We'll send the recording to everyone, and if you selected a next step in the survey, a member of our team will be in touch. Thanks to Niamh and Gerard for sharing their expertise today.