Workflow Automation with AI: New Tech, Same Problems

Automating business workflows with AI Agents represents an exciting opportunity for companies to drive productivity.

Mark Benioff, CEO of Salesforce, talks about being in the business of digital labour. Leaning into this idea that you will hire a digital workforce so you won’t need to hire a human one with inevitable associated, very human, HR challenges.

It’s easy to get lost in the hyperbole around AI replacing humans, but the reality is much less binary. AI Agents, like any technology, will replace some things that people do and allow them to do other things that AI can’t do. Harnessed wisely, they represent an answer to the productivity malaise that’s affecting economies with aging and shrinking working populations.

The main challenge we’re facing when adopting these technologies is one of a mindset shift. Computers have always been predictable. They are deterministic. You provide the same inputs, you get the same outputs.

The current generation of AI technologies, based on Large Language Models (LLM) do not follow this mode of operation. Instead they essentially predict what comes next based on the inputs. The output represents something that matches what should be expected. They are probabilistic. They are probably right. They therefore make mistakes.

The good news for people designing workflows that include these probabilistic workers is that humans make mistakes to and have done for millennia. It’s part of our makeup. Making mistakes is what has allowed us to evolve our amazing minds and capabilities. Sometimes those mistakes turn out to be the new way of thinking, or way of doing things.

The way we operate businesses is with mistakes in mind. All of our processes have various checks and balances in place to catch the inevitable miss-steps, ideally before they have a negative impact on the business.

The scale of these checks and the levels of control and oversight they introduce are generally in proportion to the risk of mistakes. In a customer service workflow for a low-cost consumer item, the jeopardy is low. When you’re operating in a highly regulated industry like mortgage broking, mistakes can threaten the entire business.

Operations teams in different industries and with different use cases will be adopting these technologies at different speeds according to their risk appetite. Where mistakes can be accommodated in a cost effective way, there will be clear ROI to be taken advantage of.

Scheduling is part of so many business workflows. The reason Cronofy exists is because we saw how hard it was to allow system of record applications that run these workflows to get access to people’s schedules and make decisions about them.

The challenge a probabilistic model has with making a scheduling decision for an individual is understanding the context for that decision.

Prioritisation, ie. is meeting A sufficiently more important than meeting B to justify rescheduling the latter, is the hardest problem in scheduling.

The information required to make this decision almost certainly doesn’t exist in the calendar, it may exist in the associated system of record apps (CRM, Recruiting, etc), but often there is a large component that exists in the minds of the people involved. So it’s not available to a model, even if the model could make use of it.

I have a sales meeting in my calendar with a prospect. It’s a small value deal according to the CRM. I also have an interview with a candidate that the recruiting team have been struggling to find a time for. They want to book over the sales meeting so they can get the candidate through the process as quickly as possible. Who should win?

Context around the sales meeting is that whilst it is a small deal, it’s as a result of an introduction by a new partner who we want to impress with how well we service their network. Which meeting should take precedence now?

The candidate has been fantastic in the interview process and is in multiple processes so speed is of the essence. Does that change things?

In the absence of the ability to delegate decision making to a model, people need to be able to understand the behaviour of a system and thus predict the outcomes. Scheduling is one of those decision making processes that needs predictability. It needs to be deterministic.

It has to give people the capability to define the rules and heuristics for the system to use when operating in an automated context. But, it needs to give people the flexibility to take exceptional decisions when necessary.

What we’re talking about with AI Agents is a new type of automation. But, it’s just software that can be combined with other software. Business workflows of the future will be formed from a collection of probabilistic and deterministic operations. The two are absolutely not mutually exclusive.

An example for volume recruitment would be:

1. AI Agent processes a job application and decides whether a candidate should be contacted.
2. Another AI Agent contacts the candidate by text or voice to ask them a series of screening questions to assess fit.
3. If the candidate passes this screen, then this agent hands off to a Scheduling Engine to allow the candidate to choose a time for their first interview which is recorded in the recruiting system.
4. Scheduling Engine then hands back to the AI Agent to wrap up with candidate and prep them for the interview.

This represents a combination of probabilistic and deterministic decision making with checks and balances where necessary. Recruiters can periodically assess whether applications are being interpreted correctly. They can review screening transcripts to assess the efficacy of the scoring. They can tweak the scheduling rules to make sure candidates can book in as quickly as possible.

Now look back over the process and replace AI Agent with Person. Each will need a requisite skill level necessary to conduct the task. The same checks and balances would be in place. A person could do any one of those steps. What is different about AI Agents is that they are becoming capable enough.

AI Agents are also cheaper to run, which opens up additional possibilities. What if step 5 was an AI Agent contacting the candidate 24 hours before their interview to make sure they had everything they needed? If the candidate has concerns, then they could be connected with a Person straight away to discuss them.

Most of the time, the candidate will be fine and will likely be pleased to receive the call, whether it was obviously automated or not. Someone cared enough to program it in to the process. If the candidate isn’t fine, or has more questions, enter the human to empathise, discuss and do what they can to set the candidate up for success.

Business workflows will look very similar to how they look today, certainly in the short to medium term. They will still be composed of a series of tasks. Instead of being performed people and deterministic software as they are today, probabilistic software will be added to take on certain types of tasks.

Interpreting natural language responses, assessing fit to a profile - probabilistic.

Finding a time for a meeting, performing a commission calculation - deterministic.

Probabilistic software needs a management approach to its outputs that’s more akin to coping with humans and the randomness they introduce.

This new software also opens other opportunities to delight stakeholders in ways that would be uneconomic or impractical if a person was required to run the task at scale.

As we have always done with new software, we will revel in the efficiency and productivity gains of the new tools we’ve brought to bear on our workflows. Then we’ll discover all the other things we could be doing, and the cycle will start again.

I'll be speaking on stage on this topic at HR Tech Amsterdam - find the details of the session below.