Designing future business processes: shifting to an AI-driven approach
Copilot to Pilot: making AI a driver
Looking across the landscape of AI solutions, it’s readily apparent that ‘assistant’ or ‘copilot’ is the correct term. They play a supporting role in business processes. Take Copilot for instance. It shows up in chat windows in all the right areas; embedded in all the forms where tasks are completed and with the ability to just tell it what to do, such as to draft an email.
The challenge with this approach is adoption. Yes, everyone can access ChatGPT. Yes, everyone can leverage Copilot features. But will they? Are they even aware of them? The implementation becomes focussed on getting the knowledge out there and getting users to see what’s in it for them - how it’s helping them. This is of course, completely valid and part of any technology rollout. What I’m pointing out is that it’s a bit of a risk and presents a seemingly ad-hoc benefit.
I want to get closer to something AI-driven. Something where the AI is running the show, and we have crucial roles to play in it. But how do you turn it from Copilot to Pilot?
First, you need to understand and address key concerns.
Addressing concerns
Let’s start with the need to address how AI is perceived in the market. A year ago, my uncle called me and he wasn’t too impressed with ChatGPT. He said “I looked up what Qantas deals were running for flights to Fiji and it just gave me rubbish”. He was trying to apply a GPT to a search engine task - something they have improved at in the interim, but not their core skillset. We need to educate the market that AI is compute: raw intellect.
Second, lets consider how it is that people really work. We’re human, we have attention spans and finite energy reserves. We live in an external environment where new information is presented to us all the time. And, we live in a world of subtle subconscious cues and influences - e.g. getting that particular lunch with that particular person was important in the long run to get that particular result. There’s a seeming randomness to it - humans act as a kind of dynamic glue between systems and processes: one that we have to confront if we’re really looking to automate beyond the current state.
Even if we accept that all of this can be centralised, structured and automated, there are questions of consistency, oversight, and ultimate accountability.
Consistency
My dad is a doctor who has served tens of thousands of patients over his career. So, when you think about it, any of those diseases that are one in ten thousand, he inevitably came across them eventually. With AI, it’s the same; even if we hit 99.9% accuracy and reliability, it’s that 0.01% that will get us. But that’s why dad was trained for that 0.01%, and it’s why we will prepare for it with AI. If we think there’s a chance of that impact, we’ll design around it - self-review mechanisms, and human approval if required.
Oversight
But what does approval mean? What is oversight? It’s likely fair to say that we don’t pay as much attention to the things we don’t directly do, no matter how much we may care about them. The biggest micromanager still considers less than the ‘doer’ they are managing. Think of the absent-minded driver in the self-driving car. So, we are losing some of that attention in taking more of an oversight role. This is where we need to think of mechanisms to focus attention on the right thing, such as alerts or focussed user interfaces.
Accountability
Ultimately oversight serves to make people accountable. Nobody is really going to go straight to the vendor (e.g. Microsoft, Open AI) for the minutiae of accountable items, such as a hallucinated response. In our designs, we need to make sure that someone always owns a task, and owns it in a meaningful way, where they are not just a head to cut off, but someone that can make meaningful judgements and decisions within the process.
A structured approach to applying automation
Given the concerns and biases many have towards AI solutions, I think we need to take a rules-based approach.
Firstly, lets apply a grading system to the extent to which something can be automated, where 0 is impossible and 10 is certain.
| Grade | Description | Examples |
|---|---|---|
| 0 | It cannot ever possibly be automated no matter how intelligent or advanced the solution is. | Purely creative arts, philosophical debate |
| 1-3 | Very challenging to automate; most tasks need human intervention and decision-making. | Complex legal analysis, psychotherapy |
| 4-6 | Somewhat challenging to automate; automation is feasible for routine tasks, but significant human oversight is required. | Project management, medical diagnosis |
| 7-9 | Quite easy to automate; most tasks can be automated with minimal human oversight for rare exceptions. | Manufacturing processes, basic IT support |
| 10 | It absolutely can be completely automated with no need for any oversight. | Simple repetitive tasks, basic calculations |
When you think about it, with AI solutions, we’re not really interested in the 10s, or even that region. Those have already been solved, years ago. Really, we’re targeting the 3-4 range; that’s where it’s an unsolved problem. Above that, we can take it on with traditional solutions.
The model applied is a good start, but to really begin assessing this on a more granular level, we need to break it down more. The following are five criteria that we can assess business operations against to really understand if we can give AI a more driving role.
Task Complexity: The level of difficulty and intricacy involved in completing the task, ranging from highly complex to entirely repetitive.
Human interaction requirement: The extent to which human interaction and emotional intelligence are necessary for the task, from high empathy needs to fully transactional interactions.
Decision-making frequency: How often decisions need to be made and the complexity of those decisions, from frequent high-level judgment to entirely rule-based processes.
Standardisation of tasks: The degree to which tasks are standardised and follow predictable patterns, from highly variable tasks to fully repetitive and centralised/standardised ones.
Availability of technology: The current state and accessibility of technology to automate the task, considering both technological existence and practical access to it. Here in Australia, I think about rural businesses with spotty connections - even if the technology exists, they can’t use it moment-to-moment.
Beyond this, we could introduce further criteria depending on the specific industry/sector/business area. For instance, there are some areas where the level of impact for false data is too high, and there are some areas where sensitivity of specific data is critical. These are topics we will go into in a future post.
For now, going with these measures, lets break the scoring mechanisms down.
| Criteria \ Score Range | 0-1 | 2-3 | 4-5 | 6-7 | 8-9 | 10 |
|---|---|---|---|---|---|---|
| Task Complexity | Highly complex, requiring advanced problem-solving and creativity. | Moderately complex, involving varied tasks with occasional exceptions. | Routine complexity, with well-defined tasks and some variability. | Low complexity, mostly repetitive tasks with few exceptions. | Very low complexity, entirely repetitive and well-defined tasks. | Fully repetitive, well-defined tasks with no variability. |
| Human Interaction Requirement | High requirement for empathy and human touch. | Moderate need for human interaction and soft skills. | Some human interaction needed, but can be supplemented with automated responses. | Minimal human interaction required, mostly transactional. | No human interaction needed, fully transactional. | No human interaction needed, entirely automated. |
| Decision-Making Frequency | Frequent, complex decision-making requiring high-level judgment. | Regular decision-making with some complexity. | Occasional decision-making with straightforward options. | Rare decision-making, mostly rule-based. | No decision-making required, entirely rule-based. | No decision-making required, fully automated. |
| Standardisation of Tasks | Highly variable tasks with no standardisation. | Tasks with some standard procedures but variability exists. | Predominantly standardised tasks with some variations. | Mostly standardised tasks with occasional deviations. | Fully standardised, repetitive tasks. | Fully standardised, entirely repetitive tasks. |
| Availability of Technology | No existing technology capable of automation. | Limited technology with emerging solutions. | Existing technology can automate parts of the process. | Advanced technology available for most parts of the process. | Fully automated solutions are readily available. | Fully automated solutions are readily available and commonly implemented. |
These scores can be weighted based on the specific business as well. For instance, a company that primarily operates in regional areas may place a significant emphasis on availability of technology, where a business primarily based on human relationships may emphasise human interactions.
Process example
Let’s take this approach and apply it to a made-up business. GymCo Equipment create and sell gym equipment to gyms. Looking into their sales process they have the following high-level stages.
Lead creation
Sales Reps typically meet prospects at trade fairs. It’s very traditional - business cards are exchanged, future discussions are set up. They scan the business cards into the CRM afterwards creating a lead.
Scoring (higher is easier)
Task complexity: 4. It’s not rocket science. It’s more that these interactions are dynamic - the Sales Rep needs to think about what the prospect is looking for.
Human interaction requirement: 1. This kind of in-person sales is deeply rooted in human connections. As cost-pressures increase, there may be a change in the format of these interactions, such as automated procurement processes, though, given the current state, we need to acknowledge the importance of this format.
Decision-making frequency: 5. Decisions are made in terms of how certain features are presented to prospective clients, or even which prospects to approach.
Standardisation of tasks: 5. It’s one product line, but the way these discussions happen is dynamic.
Availability of technology: 3. Nobody wants phones out in these conversations (again, up for future challenge).
For us to get a high degree of automation here, there needs to be an industry-wide movement towards more integrated systems of selling and procurement. A tall order, sure, but not impossible. Over time we can expect a higher level of centralisation.
Technology solutions:
Quickly understand the prospect with AI-augmented search.
Build up a strong case for the sale quickly based on best practices and relevant findings about the prospects business.
Understand where the prospect sits within the stakeholder map of the organisation immediately so that you know how to play it.
Understand social connections, such as through LinkedIn, where you or someone in GymCo may have a connection.
Lead conversion
To convert a lead, Sales Reps must answer: is it real? Can I win? Is it worth it? These questions have specific precedent and metrics behind them, and drive whether further effort is required.
Scoring (higher is easier)
Task complexity: 4. There are three questions but really they come down to subjective assessments that have no raw data.
Human interaction requirement: 3. There’s a fair degree of human interaction needed to really determine the answer to these conversion questions.
Decision-making frequency: 5. This is a decision, though it is based on a broad set of rules.
Standardisation of tasks: 5. This is a standardised, centralised process in a subjective area.
Availability of technology: 4. This is usually less remote than a trade show, but in some cases may be immediate.
Here we can see more policies and rules coming into play. Centralising this may be beneficial, as it leaves it less up to the subjective and perhaps motivated perspective of the Sales Rep, and more up to agreed definitions.
AI solutions:
AI can convert the lead itself based on pre-defined subjective definitions. If required, this can have an approval step from the Sales Rep or their Sales Manager
Opportunity development
With the lead converted, the sales rep works with the team to understand the prospective client in further detail: their organisation structure, business need and budget. They also need to understand any potential competition, and any differentiators that may provide GymCo with a competitive advantage.
Scoring (higher is easier)
Task complexity: 5. There are standardised, templated elements to this
Human interaction requirement: 5.There are prospect interactions at this point and judgement based on human subjective experiences.
Decision-making frequency: 6. There is a degree of judgement in terms of approach, but this mostly follows trends set earlier in the process.
Standardisation of tasks: 6. This is generally a templated approach, but methods vary somewhat between Reps.
Availability of technology: 7. This is a more centrally managed process.
Now we have a structured process. There’s a good number of areas to automate here, from understanding the prospect to forming the perfect proposal.
AI solutions:
Understand the competition based on searches for historical sales news
Understand the prospect deeply, as if analysis has been performed by a group of consultants, with the use of directed AI assistants
Opportunity proposal
Typically, the Sales Rep will draft up a quote and share it with the prospective client, making adjustments to the quote lines based on feedback, and maybe even throwing in a discount if they think it will draw in future sales.
Scoring (higher is easier)
Task complexity: 7. We’re producing a quote based on specific negotiated outcomes and a set product list.
Human interaction requirement: 4. To get the best outcomes, there’s a fair degree of human interaction.
Decision-making frequency: 6. This is about how it is negotiated and worked out. There’s only so much to bend, but a fair width within that.
Standardisation of tasks: 8. It is always selling the same sort of thing. There’s ambiguity on a case-by-case process, but the overall process is fairly set.
Availability of technology: 8. This is typically a centrally performed process.
As we get later into the process, the prospect of automation becomes more certain. This is certainly centralised with specific rules and policies, and it’s more a question of implementing them to win the deal.
AI solutions:
‘Read’ the prospect with AI and understand the strategy they are susceptible to
Instantly create the right quote with the right adjustments, based on multi-modal information (e.g. the gym the client is running and space available, previous discussions notes).
Opportunity closure
At the end, the Sales Rep knows whether it is won or lost. If it is won, a sales order can be generated, the delivery can be made and an invoice can be sent. If it’s lost, it’s recorded and ideally noted why, so that future improvements can be made.
Scoring (higher is easier)
Task complexity: 8. This is just administration at the end of the day.
Human interaction requirement: 8. The information is already in, though some interaction may go into smoothing out delivery.
Decision-making frequency: 9. The call has already been made by this stage.
Standardisation of tasks: 9. This is just administration.
Availability of technology: 10. This is typically performed centrally and is very rules-based.
By this point, most of the process is very structured and repetitive. Traditional automations can handle this, with some AI enhancements.
AI solutions:
Take unstructured notes and convert these to structured inputs on the competition and the prospect, improving future insights
Automatically provide notation on products for future improvements, improving integration between sales and development
Warehouse pick and pack. This is where I see something like Boston Dynamics coming up with a great connector.
Run the billing process through your ERP, with any required adjustments based on semantic information, and within pre-defined rules.
Overall, the sales process becomes more open to wholesale automation the further we progress through it.
Summing up
We need to step back and ask why. It seems that, when we really dig into tasks on a granular level, there’s always a level of automation we can apply, Even when the answer is that it’s impossible, we can ask why again - we can reframe the structure and think about ways we can make it work, instead of reasons it can’t.
