Putting the Cart in front of the Horse – Chatbots in Support


Share on LinkedIn

My recent rant on chatbots having the potential to kill user experience got some nice reactions. It brought me into some interesting discussions on support, mobile, the role, strengths and deficiencies of artificial intelligence (AI) and machine learning (ML) and so forth. Most of these discussions dealt with mobile support but also with the question where AI could benefit most. Particularly good one were with Abinash Tripathy, CEO of mobile support platform Helpshift and Srikrishnan Ganesan, founder of Konotor, now hotline.io after being acquired by Freshdesk at the end of 2015.

Both companies have a focus on in-app support, a solution category that basically got introduced by Helpshift, after Abinash identified a lack of good options or delivering support directly to and via mobile phones. One of the premises is that a lot of the technically necessary and relevant information can get collected directly and sent to the service back end transparently. They have some big customers, including Microsoft and a raft of gaming companies, including Zynga and Supercell. He, of course, has an opinion on bots in support, which he recently also expressed on Venturebeat.

Hotline.io has a customer base that is mainly made of transactional companies, which, too, leads to a high message load but also leads to different approaches, as the user context is often about past transactions. This means that regularly not that much information gets sent together with the support request. Sri, too, has a vision on how to incorporate AIs and bots into support.

Hotline.io is offering a browsing style of offering help using a shallow tree with icon-supported categories on top of a search interface as it is also offered by Helpshift. Of course both systems offer direct in-app chat to support, too; here again hotline.io offers context via the categories (called channels), which can be used for entering the chat session. Helpshift is more relying on system context here. What both companies are doing with this is to establish a focus and to initiate meaningful first reactions.

Why do I talk about this here and now? Because both companies, as well as others, are looking into adding bots into their infrastructures.

AIs and Chatbots still have a Problem

While my criticism to quite an extent was around the poor user interface that a chat application offers, as compared to richer environments, I acknowledge that many people are texting and messaging. In fact, the number is only increasing. This means that there is a viable user interface.

However, everybody has their own dialect, choice of words and, worse, abbreviations. Sometimes people even go to the stretch of asking their questions rap style or a veritable rhyming competition about a dead worm evolves. A lot of important context that is not immediately visible to a machine is needed by this type of communication. Add potentially overlapping messages between the communication partners to this.

All this makes it hard for machines to ‘understand’ the nature of a request and to answer correctly. It is already hard for humans.

It seems to be general consensus that the accuracy of natural language recognition is by far not yet where it needs to be in order to provide useful support; support being delivered in text based environments or, even more difficult, in speech. As good as a 90 per cent plus recognition rate sounds, this is still far too low to be really accepted and the remaining about 10 per cent will antagonize a lot of customers.

A lot of them!

On top of this, although specialized AIs often work surprisingly well, more generalized tasks still are difficult to cover by them. Yet, chatbot platforms are focusing in on helping customers who are already in distress – or are doing funny stuff like selling flowers, for which one wouldn’t really need an artificial intelligence … but then this is likely also the easier part, as the process is much more guided.

I think this phenomenon of applying AI everywhere is largely fueled by a technological hype that lets us forget that not everything that is possible needs to be done, let alone should be done.

A hype that seems to put the cart in front of the horse, as the outcome could potentially be disastrous for a company’s image.

After all bad news travels fast and far – faster and farther than good news.

On the other hand, if AI’s are not working well enough yet, they need to get trained. This works best by, well, using them.

A Way Forward

Of course this causes a classic chicken vs. egg problem, which could become a real problem for companies that need to keep their investments in check.

There seem to be three ways out of this dilemma:

  • Follow the KISS principle and increment the usage of AI’s and/or bots from a domain of structured data into unstructured data, essentially starting from the simple problems (although these do not need an AI nor machine learning)
  • Train the AIs in parallel to support sessions done by customer service agents or self-service sessions
  • Combining the above approaches

The first approach is pursued by both Helpshift and hotline.io, again using different approaches. An additional precondition to AIs successfully delivering support is that chat via mobiles will be recognized as an important channel, if not the primary channel for the delivery of customer service; this not only by customers and businesses, but also by software vendors. According to Abinash, e.g. Microsoft and Salesforce are ahead of SAP and Oracle with this understanding.

This way the bots can provide some value early, which then gradually and constantly can increase by supporting more difficult requests. How could this look like in real life?

  • Use a kind of first response bot that takes up essential missing user data, routes the request into the proper queue and sends an acknowledgement, thus buying some time for the support agent
  • Improve the quality of the retrieved knowledge base articles – learn using the time that a user spends reading an article and the users’ rating on helpfulness in correlation to the question asked as well as from the suggestions of the human operators
  • Forecast wait times and provide intelligent notifications, so that customers are not bound to ‘places’ when in chat based support

More sophisticated approaches include

  • Have a bot ask relevant questions about signs and symptoms that the user did observe or could have observed before calling support, while the human operators are busy. This helps in shortening the wait times for the customers, who already are in distress.
  • Narrow down the range of possible hits in the knowledge base and/or suggest next best steps; this then gets evaluated/used by the agent. The human operator takes over equipped with relevant information.
  • Have the bot in addition suggest solutions or next best steps for simpler problems directly to the customer. In essence this would model a tiered support system. The bot in the first level catches as much information as possible and also attempts at solutions, if the problem appears simple enough. Else the incident is handed over to a human agent with deeper knowledge.
  • Analysis of usage patterns for potential improvements of the application, to better help the service agent, or (if not too annoying) suggestions on how to do things more efficiently

Most of these approaches require a seamless handover to the human service agent. And, using these approaches, the covered scenarios can become increasingly complex, thus becoming more valuable for both, customers and service providers.

Additionally, communities can be used as a helpful vehicle, too. Not only are they possible training grounds for AI’s but they also serve as a valuable source of results. Further, it is possible to have artificially intelligent community managers or even –members that have the ability to provide other members with helpful answers.

In a highly advanced future state these AIs could then have the expertise to answer and solve problems on their own (thanks to Esteban Kolsky for planting this train of thought).

But that might be part of another post.

Republished with author's permission from original post.

Thomas Wieberneit

Thomas helps organisations of different industries and sizes to unlock their potential through digital transformation initiatives using a Think Big - Act Small approach. He is a long standing CRM practitioner, covering sales, marketing, service, collaboration, customer engagement and -experience. Coming from the technology side Thomas has the ability to translate business needs into technology solutions that add value. In his successful leadership positions and consulting engagements he has initiated, designed and implemented transformational change and delivered mission critical systems.


  1. Facebook: using an anti-consumer platform to promote customer advocacy, disavowing customer’s preferences and reporting it as record:
    1. user counts [driven by bot initiated actions]
    2. user engagement levels
    3. and growth rates

    (despite international bans and national news outlet research documenting declining metrics) resulting in….

    (Drumroll please) The bandwagon’s replacement by a bandwagon-train on which to put the masses of new [bot-driven] SMBs lured by Bonnie & Clyde’s (Sheryl & Mark) idiotic cause & effect reasoning (which devalues marketing attribution to non-target reach and total impressions served and regardless of all other business inputs). Bots represent the Wild West of advertising in which every B2C communication is an ad – and who will count Facebook’s self proclaimed (generated) 60 billion messages per day.

    Those types of claims are necessary to offset:
    1. Facebook’s pummeling at the hands of YouTube in video (users HATE Facebook’s auto play content and silent ads [which steal 50% of battery even when app is supposedly disabled – FB allegedly ‘fixed’ the bug but never formally acknowledged it]
    2. Google’s bludgeoning of Facebook in search, Facebook’s complete failure with Graph Search and Facebook’s asinine claim of a burgeoning search business that now generates a claimed 2 billion searches/day.*
    3. Facebook’s ‘Mock-Internet’ is like every village idiot going to a library resourced by and sourced from each users’ personal friends read each day and photographed and documented (tagged, detailed, etc) – not what is discoverable in the world, and influenced in part by user effort, user intent and the precise yet creatively enlightening and bountiful nature of the search.

    That library sounds like a pre-Y2K & AOL walled-garden shit show. And funny thing, at current values, like Madoff, Ebbers and Jain before, Facebook leaders will keep 1/2 of what they have, plus whatever they’ve shifted to family members & trusts.

  2. Hi Stacey,

    if you want to say that a lot is going wrong with FB and co I wouldn’t really disagree – while grudgingly being amazed at what Zuckerberg has achieved.

    But that is a deviation.

    Regardless of whatever is going on with AI/machine learning/bots in sales and marketing, and I think that much of what is going on is in the category of not so valuable, the same holds true for support scenarios.

    Which I think are more difficult to handle.

    I believe that we are still in a phase of trial and error (‘wild west’) in all areas. I also believe that viable approaches – for businesses as well as for customers – will emerge. Probably supported by some regulation. And then: Even giants are not infallible.


  3. Hi Thomas,

    I agreed with your comments back in August, notwithstanding the incentives for chatbot service providers who wield and deliver the user base (reach) and scale required for support scenario engagements. The recent barrage of erroneous or untrustworthy engagement, measurement and effectiveness metrics disclosed (so far) by Facebook underscores the point I was trying to make.

    Since Q4 ’15, numerous examples including Sparkle-n-Pink, Lighting Etcetera, and others cited on Facebook’s quarterly earnings calls heap full responsibility for successful campaigns on its advertising solutions despite a lack of objective, verifiable, quantitative criteria and/or include obvious errors or incorrect assumptions implicit in the reasoning, with no mention of the 85% markdown of website pricing, the company’s new location or any other business drivers.

    So, when Facebook claims of 3 million SMB advertisers and 60 billion messages sent across its app base amid slow adoption by enterprises suggests SMBs, it is easy to imagine a scenario in which those SMBs are again targeted. And it’s easy to anticipate similar abuses of both when to engage the proprietary measures themselves (scenarios) and to what extent (depth of engagement for each metric, in part because of the less-savvy client base whose specific needs were ‘underserved’ by traditional “3rd party measurement” providers. ;

    Regulation and standardization in defining, measuring, evaluating and reporting done in advance will never eliminate error. But it can head off their continuous recurrence.

    If “It’s easier to beg forgiveness than ask permission” is an acceptable offensive strategy, then “computers only lie when they are told to” might be an acceptable assumption going in for defending against it.

    I hope I am wrong; I would love nothing more than to learn that all publishers and advertisers agreed to a common set of definitions, KPI, etc. and all data were squeaky clean. But until then, intellectual curiosity and wonder are useful – and necessary – tools.

    Best regards,


Please use comments to add value to the discussion. Maximum one link to an educational blog post or article. We will NOT PUBLISH brief comments like "good post," comments that mainly promote links, or comments with links to companies, products, or services.

Please enter your comment!
Please enter your name here