[Originally posted 5/4/2023 on LinkedIn]
A Proof of Concept
Introduction
There is a lot of good work and conversation around the benefits generative AI can bring to the Customer Support space, particularly in knowledge, content, chat and adjacent uses. ChatGPT, Co-pilot and others will revolutionize support and customer facing functions.
But it shouldn’t stop there- data AI has as much potential to massively upend the ways customer support data can be changed from latent to highly proactive uses.
What follows is an illustrative example of Data AI connected to a customer support data feed, using mock data in a curated, fictional scenario but with a commercially available tool (in this case, the Microsoft Azure Applied AI Services) to demonstrate the very tangible capabilities – in terms of cost, resources, customer sentiment and overall impact mitigation- of highly proactive data in a customer support model .
About this proof of concept: I created a dataset containing support contact (CRM) data for 2,500 inbound contacts over a 3 day span, injected a volume spike event, and followed that data through the Azure Applied AI data visualization dashboards and tools to demonstrate the ways it can be used in real-time.
I based this exercise on “Gardenworks”, a fictional SaaS service providing a range of growing, farming and agricultural information with historical, current and forecast data for US climate zone-based environmental and growing conditions. They have around 225,000 subscribers across Basic (gardeners and landscapers) and Advanced (farmers, agricultural producers) customer segments.
A brewing storm
Friday, 1:00p- Something’s up with contact volume
It was a little after 1:00 on Friday when Skylar, a Sr Support Engineer at Gardenworks, noticed that customer contact volume was picking up. Gardenworks CRM classifies all customer support contacts, whether online, chat or phone, by the type (category) of the inquiry, the major product component involved, and what product version the customer is using. There are additional data fields available such as issue description (free text and CRM/support notes), customer segmentation, etc.
Skylar clicked into the hourly data feed* to see the detail of the volume coming in and saw that activity for the issue category of “Problem Report” had ticked up to 10 in the last hour, a bit higher than usual.
It could be just random noisy data; to see whether it continued to trend, she added a data anomaly detection threshold* at 15 per hour and continued reviewing the other dashboard items.
1:30p – Anomaly alerts around volume spikes
The alert Skylar had set fired, and she drilled into the support contact metrics detail to the incident view. With some quick data pivoting she saw that the Problem Report category volume was being driven heavily by Product Component of “Filters” (55%) and Product version of 23.4 (91%)*.
She also looked at service operations incidents on the Gardenworks service dashboard, but this component isn’t instrumented for operational monitoring, so there were no active incidents being reported around them.
Taking action based on the data
1:45p- Having such a clear read on the details of the contact volume drivers enables Skylar and the support team to dive directly into the specific web, chat and phone contacts that are coming in for this issue
At this point she knew there was an emerging issue with the filtering component in version 23.4. At 1:45 she engaged some Level 2 support engineers to dig into support contacts in the last hour with these two attributes. They used text analytics to find that the text strings “zone 9”, “zone 9a” and “zone 9b” were surfacing in many of the contact notes. (USDA Agricultural Grow Zones from 1a to 13b are data that Gardenworks provides, and in the Enterprise version, allows users to filter views on this). The team went directly to the production environment and see that the 9a/9b zones were not showing- somehow they had been removed from the UI filters.
2:05p- Support now has a read on the root cause of the customer issue – and it could affect thousands more users
Skylar reviewed the latest release notes and found that a USDA Zone configuration file had been updated as part of a production release earlier in the day. It was 2:05 – just 60 minutes from the first volume increases- and the support team knew the root cause of the spike, which was now trending at over 35 per hour, with 47 support contacts received across the web, chat and phone teams.
The size of the approaching storm
Skylar knew that Zone 9 was one of the 4 largest in the US, by population, covering much of the west coast, California’s valley farmland, and western Arizona.
In fact, a customer report showed that nearly 6% of Gardenworks 225,000 customers were located in Zone 9, and 70% of those were commercial or SMB farms that used the Enterprise version at the heart of these issues; any of them doing zone filtering in the product would run into this issue. Gardenworks support was looking at an impact to potentially 9,700 support contacts over the next days and weeks– a tsunami type impact that would swamp them and create huge and painful customer impacts. Skylars Friday had just taken a very bad turn.
2:30p- A support and product triage spins up
She alerted the product operations team, who by 2:30 joined an issue triage where she briefed them on the information. They reviewed the zone configuration file and found that Zones 9a and 9b had indeed been removed. They left the call with a plan to fix, test and re-release the file, and Skylar pulled all of the information she had collected to have ChatGPT quickly create an online customer facing support note about the issue, looking to stem at least some of the chat and phone volumes.
Avoiding impact
3:35p- a fix is deployed, and contact volume starts to dissipate
When the fixed file was deployed at 3:35, 112 support contacts had been made in just 2 hours, and they were climbing at a rate of 45 per hour. Skylar continued to monitor the data feeds, and saw them quickly dissipate over the next 45 minutes until all Problem Report contacts were back to less than 5 per hour, a normal trend.
125 total support contacts had been received due to this issue, but more than 9,000 potentially impacted customers had been spared from the issue, avoiding all of the associated support impact from them.
Averting future storms like this
Following the issue, Skylar and the Product Operations team collaborated to develop a QA test script and support test release gate for the zone configuration file going forward, which includes a support data alert threshold (the same one that Skylar had used earlier) to be set for any contact category and product component related to configuration file issues.
Proof of Concept outtakes
Data AI services can turn latent support data into powerful proactive tools used to create enormously impactful customer and business outcomes.
There are a variety of platforms that can be used- from Azure Applied AI, Salesforce Analytics with Tableau integrations, Oracle Service Cloud customizations, IBM Data AI, Amazon AI Data Extraction, and others.
Support CRM and service data is commonly captured via enterprise tech stacks in the form of incident/case categorization fields and reported daily, weekly or monthly- relegating customer impacting issues to after-the-fact reactive views and delaying effective support responses while perpetuating customer impacts.
Data AI enables highly pro-active and actionable views of data, slices and insights, which can drive customer impact avoidance, product satisfaction detractors, support cost mitigation, and tightly collaborative capabilities among product, support and success teams.
Support data can fill a critical gap within SaaS product instrumentation and service operations monitoring.
Instrumentation is limited in the range of customer issues that can drive support contacts. Where there are instrumented product features, a service or product operations team can quickly see and react to real-time incidents. But many customer impacting issues that drive support contacts are not related to scenarios that include any monitored instrumentation. I created this proof of concept specifically around such an example, where a confusing UI or malformed source configuration leads to missing UI elements or incorrect user selections within a common task workflow.
Instrumented operations monitoring can be blind to these types of issues, although they can drive significant support impacts with the associated resource cost and customer satisfaction impacts.
These types of capabilities aren’t difficult to build.
I created this working proof of concept with a SQL database, ingesting CSV mock data and connected as a live Azure Applied AI data feed within Metrics Advisor within a couple days, without being deeply familiar with Azure services. Of course within organizations there are IT, data governance and related considerations. Using proven, commercially available data AI platforms should integrate with these requirements well.
A more agile data platform built on live CRM and support contact data feeds (Web, text, chat, phone) can be an enormous and tangible asset in terms of:
- Providing rich data pattern visibility, allowing customer impacting issues to be surfaced, triaged and mitigated in near real-time.
- Increased customer satisfaction (realized through quickly mitigated issues- minimizing the number of customers who aren’t impacted by something).
- Decreased support costs through much shorter impact-to-resolution cycles.
- A customer support function with vastly more proactive capabilities, leading to higher value and more effective use of resources, and the ability to be a more synergistic part of the product ecosystem where customer support data and priorities are aligned tightly with agile product and service operations cycles.
Conclusion
There is a lot of good work and conversation around generative AI in support, particularly in the knowledge, content, chat and adjacent spaces. ChatGPT, Co-pilot and others will revolutionize support and customer facing functions. But it shouldn’t stop there- data AI has as much potential to massively upend the ways customer support data can be purposed.
Consider whether your customer support data is actionable and proactive? Are you able to use it to pivot to real-time actions that are beneficial to customer experience, the business, and support costs? Does it give you capabilities to support robust and impactful product, operations, and service partnerships? There are many solutions available today that are relatively low-drag to deploy, work with existing enterprise CRM’s and tech stacks, and will put your support data to work in powerful ways.