machine learning Archives and Blog Posts - Entrepreneurs with Denis Oakley

Entrepreneurs with Denis Oakley

I HELP BOLD LEADERS TRANSFORM THEIR BUSINESSES AND THE INDUSTRIES THEY COMPETE IN

You are here: Home / Archives for machine learning

machine learning

By

PrecisionBit Business Model Analysis – Predictive Image Analytics for Advertisers

PrecisionBit Business Model Analysis - Predictive Image Analytics for Advertisers
PrecisionBit Business Model Canvas

This week we’re having a look at an early stage startup in Singapore. PrecisionBit is hoping to change the way that advertisers connect with their target markets with clever use of machine learning. They are really early stage though, so not a lot of information is available about them. However they are part of Telstra’s Muru-D incubator so they’ve got some funding and just as importantly people who believe in them. Let’s see what they are trying to do and how it works

Introduction

The basic premise is that PrecisionBit scans through all the metadata that it can find on billions of images on the web and then uses that to recommend visual themes that will appeal to target audiences. In their own words

 PrecisionBit analyzes information embedded in user-generated photos to

recommend visual themes that will generate the most audience interest

There are a number of ways of doing this but clearly the largest sources of user generated photos on the web are found on Pinterest, Facebook, Instagram and other social networks. There are plenty more sources of course but what is critical about these and similar sites is the association of user data with the images and signals about the power of the imagery.

Let me explain. Your presence on Facebook establishes you as part of a particular demographic. Age, sex, location  etc all allow advertisers to focus in on you. Now I’m fairly sure that social network analysis shows strong patterns in the makeup of people’s networks. So for example a 42 year old may have a multigenerational family component, a hobby component, a life long friends component and a work component to their network. Broadly each of these components has a range but across large numbers of users we can make some useful simplifying assumptions. That say 92% of a 19 year old American female in Dallas friend’s will be in the same demographic as her.

Then we can pretty quickly see what she posts and what she likes. If she posts something that’s a signal that the imagery, text of meaning of the image is popular. likes or upvotes are a further signal of the image’s strength

So if we know what this hypothetical user likes, and millions of other people like, then we can start analysing the data to see what very particular demographics like

That’s the background. The problem that PrecisionBit is trying to solve is that advertising click through rates (CTR) are declining as audiences get saturated with digital advertising. That’s a problem for advertisers. They have to pay more or reach fewer people meaning a worse ROI. At the same time Google and Facebook take 54% of the global digital advertising spend.

So by being able to look at the data and predict that a particular creative will result in a much higher CTR, engagement, conversion rate or whatever other metric is being used will directly push through to advertisers bottom lines. That’s a pretty good proposition. How does the business model stack up?

We’ll now take a bit of a safari through the different parts of the business model, make some educated guesses, and see how it could work

Customer Segments

PrecisionBit, I think the name comes from CNC drill bits doing very focused machining on high tech products – or perhaps a focus on bits of data with nanometer resolution, states that they are after 2 customers. Agencies and brands. We can pretty confidently say that in it’s current state PrecisionBit is not a tool for Mom and Pop advertisers. They haven’t got the digital or the creative skills to worried about the exact colour palette used and it’s impact on engagement. You need to be large to worry about that and having a large ad spend for something like this to move the needle. Remember Google’s blue period of analysis.

So we are looking at large corporates, FMCG and the agencies that serve them. Specifically within them we are looking at the people managing and designing digital campaigns and of course optimising them. That’s a relatively small population of tens of thousands of people who can be prioritised and you can target the most likely prospects fairly easily with lists of big advertisers. (That’s a bit old but there are plenty of industry ranking available)

An alternative customer approach would be to ignore the brands and agencies and develop the technology for integration into Facebook or Google’s ad technology. It’s a small advantage for one over the other.

Value Propositions

So that bring’s us on to the value proposition. I don’t really have enough data to do a proper value proposition design here but a simple summary of what PrecisionBit is saying on their website is that their value proposition is the:

Ability to predict the most engaging visual content to maximise ROI

For a digital marketer that’s taking away a bunch of optimisation pain and offering her the chance of higher performing campaigns and better ROI. Of course it’s just another weapon in the arms race. Once the manager’s competitors have it the net effect will be to solidify market share and increase their defences against incumbents.

That said the VP is powerful and attractive and there should be the tech behind it to deliver. There are a number of alternate VP’s lurking in the background for other product lines. The tech for example could easily be used to be identifying negative pictures and thus as an anti-spam mechanism (that would require real time analysis)

Customer Relationships

As we’ve seen there’s probably a relatively small number of target customers. This is not a mass B2C offering. It’s also worthwhile considering that, depending on the price point, many of the users will be technical marketers rather the deciders who will be targeted by marketing and sales. The aim will be to provide technical marketers with another tool to improve their conversion rates.  So pretty much I’d expect this to be a standard web based automated tool that spits out answers with a support forum and dedicated helpline for higher pricing tiers.

Channels

At the moment PrecisionBit has a ‘contact us for demo’ form and negligible SEO and social media presence. It’s early days but it means that we have to take some guesses here. The marketing and sales team is weak except for the CEO, Rachelle Lao, who has a FMCG marketing background and a track record as a hustler to put big deals together.

That suggests that the primary route to market is going to be a direct sales route into large corporations using an existing network of contacts. Once that sales process is established then the primary channel is likely to be a traditional enterprise software sales route. This really does depend on the price point that PrecisionBit can achieve. If the company is looking at $100/month subscription then it will have to fall back on and inside sales team. If it’s selling it on a seat basis then the economics are quite different.

That really depends on the economic value that can be delivered to customers and how much of that PrecisionBit can secure for itself. .

Revenue Streams

It’s Saas. There are a bunch of ways of doing this. It will probably be a company or per seat subscription basis with several tiers. Pay for performance is probably too difficult to do as the campaign data probably won’t be flowing through PrecisionBit servers. A % of ad spend could be one route but this would be difficult and if CTR’s continue their decline this would either be complex of unattractive. So subscripotion sems a fair bet.

Key Activities

There are two key activities here. The first is to gather the data required. The second is to analyse it and use it to predict outcomes

I said earlier that there is loads of data inside Facebook – but Facebook is not famously friendly to releasing it’s data. It’s how it makes it money after all. So however good the API’s are they are not going to release the really valuable data. On the other hand if it is a walled garden it’s quite hard to climb over the wall. FaceBook has reasonably good anti scraping defences as do many other social networks. The less good a social network the easier it’s data is to access but then it is smaller, less representative and there is more statistical jiggery pokery involved  to get statistical significance in the predictions. One alternative here is that the whole company is being designed for aquisition and everything that is being done is to demonstrate proof of concept by FB or Google. That would provide the raw data that would really make the proposition effective.

On the other side analysing the data is an easier problem. using photo metadata and social signals we already have relatively structured data and there are plenty of techniques out there for creating data sets and establishing predictions. The obvious question is how significantly does the image choice impact the campaign outcome, and lesserly, which of the key parameters have the most significant impact. Then you have to set that against the ability of designers and digital marketers.

My aesthetic talents are notoriously poor so even at a very early stage the PrecisionBit algorithm would be beating me. It doesn’t need to play at the same level as Lee-Seedol either. It does need to deliver a significant advantage to people who have been making creatives for decades, and have two centuries of creative history behind them and two millennia of art history behind that.

On a speed basis the algorithm would have little advantage. It and an experienced Creative Director would make a judgement in seconds. Creative Directors do get it wrong, especially in novel turbulent markets, so one advantage would be in the certainty of prediction. Does the algo get it right more consistently?

The other question and one which I have no answer is how does the machine learning play? By implication it should be looking at campaign as well as raw data and using that to refine it’s predictions. The more advertisers using the platform for moe campaigns the better the predictions would be but that would mean an aggregation of data at some level, even when anonymised, and that would lead to the transference of power between advertisers. In contrast running a different instance for each advertiser would lead to slower training  and worse results.

Key Resources

The key resources have to be the algorithm and the data aquisition mechanisms. As mentioned in the Key Activities section the way that data aggregation or sharing happens could also be a key resource, as could access to detailed data from inside a Social Network fence.

Key Partners

Any Social Networks that come on board are going to be key partners as would be agencies with multiple large accounts which could act as resellers.

Cost Structure

Engineers, Engineers and Engineers.

PrecisionBit Business Model Canvas

 

PrecisionBit Business Model Analysis - Predictive Image Analytics for Advertisers
PrecisionBit Business Model Canvas

Are you an Entrepreneur?

Are you copying a business model?

Are you creating a new one?

I can help you get it right

Review YOUR BUSINESS MODEL

By

Dynamic Workflows in Business Process Automation

IT-WorkflowNote: Dynamic workflows do not as far as I know yet exist in any business process automation software or solution

A dynamic workflow is, very simply, a workflow that is automatically created and destroyed by the business automation software.

In traditional business process automation software solutions, such as Salesforce, you automate a business process by analysis the process and then coding it into the software. It is launched when one or more conditions are met and completed when all the actions associated with it are completed. Conditions can be based on any structured data that the software has access to and actions can be anything that the software can do itself or trigger in external applications through connectors.

An example would be to send a customer an email automatically on his birthday.

Problems with Traditional Static Workflows

This is great and it has enabled businesses to massively increase their productivity. There are a number of problems with this approach.

  • It suits businesses with well known and well mapped processes
  • It is time consuming and often tedious, difficult and expensive to implement (or all three)
  • If the business changes a lot of the automation has to be reviewed for relevance and effectiveness
  • It can increase the costs of organisational change

Creating a Static Workflow

An example is a small change that I made to a Salesforce instance recently. In this case there was a certain category of opportunity that was automatically generated on the system which needed to be passed to a partner. You simply had to tell the customer what was happening and pass the details to the partner.

  • Clone an existing email template to the customer and update wording
  • Clone an existing email template to the partner and update wording.
  • Create two new email alerts
  • Create the workflow
  • Test
  • Deploy.

All in it took about 90 minutes and it will save about 2 minutes every time that it is triggered. So you get the pay back pretty quickly. The trouble is as you work through a sales process and increase the granularity you find that there are hundreds and thousands of these small requirements. At some point the opportunity cost becomes too high and the Salesforce admin revolts.

When you use dynamic workflows in contrast you are able to see patterns of emerging demand for workflows through patterns of user behaviour and implement them automatically.

workflow_processCreating a Dynamic Workflow

So to use the above example again.

Think of the sales person working through her email in box. She replies to an email and forwards it to a partner. A few minutes later she repeats the pattern of behaviour. The dynamic automation software sees this and looks at the emails, the recipients, the opportunities that they relate to and the text of the original email and the salesperson’s message. Critically it then goes back through much of the salesperson’s email history to validate the patterns it finds

It then suggests to the salesperson that a new workflow rule should be created. The response of the salesperson then helps train the system.

Depending on the scale of the organisation you apply this on a general basis or an individual basis. People have different styles of working so in much the same way that you may have different page layouts for different profiles you could well have different workflows for different users. These would be constrained by a general automation architecture that would place limits on both what is acceptable behaviour by the user and on their changes to the way the CRM is automated.

The key benefit is that, depending on how you set your  system up, you are able to capture and automate a very significant part of user’s daily administrative work. It’s likely to achieve 80% over time with the remainder being edge and emergent cases that require more thought.

You are able to do this across the whole organisation, at once, continually, and as the organisation changes the workflows will adapt. So, if you get the architecture right, that’s not a small if, you will have a much more robust business process automation system that reacts quickly to the changing market with far fewer requirements for expensive integration consultants

Scaling at the Organisational Level

An alternate approach is to use dynamic workflow creation to generalise best practice across the organisation. So for example when you have 5 or 10 sales people in a large team repeating patterns of behaviour – even something as trivial as editing data or selecting products – you can identify that behaviour and then modify fields, layouts and workflows for the organisation as a whole to enforce the best practice again automatically.

A final example; consider a team, of sales people sending out customised proposals by email. By looking at the information that is sent to the prospect and how it is presented and then data on the prospects behaviour  you can extract best practices that are then fed back automatically into the template for a process of continuous improvement.

Now there is no way that I am suggesting that any of this is easy. Ignoring for the moment the engineering problems in developing systems flexible enough to do this you also have a problem with the business process automation software which suddenly has to have another meta layer of complexity added to it. No fun at all for system admins.

The promise though is that using machine learning and other algorithmic tools you can start to reclaim much of the 60% of a salesperson’s time that is not spent selling. You are essentially shifting complexity from the human arena to the machine. By doing so you have many more tools to manage it effectively.

Some Thoughts on Creation of Dynamic Workflows

You need to have a confidence level. This can actually be set fairly high because a lot of the time the pain has to be fairly high for users to push for a positive change (as opposed to complaining about bugs). I find that when a user has to do a trivial task dozens of times a month they start agitating for a workflow.

Activation: Could this used unwisely trash an organisation? Sure. One trash case would be having emails hacked and used to send malicious messages. This could then trigger a dynamic workflow and thus perpetuate the hacking. So you need checks and controls in place. Mostly though normal management controls should be sufficient to control negative dynamic workflows.  After all if staff use of the email and CRM system is non compliant with your procedures you probably have pretty big problems to worry about.

Senescence. People change. Organisations change. Certainly with Salesforce instances bloat over time as obsolete code is deactivated but not removed. It requires constant admin focus to keep it lean. If you are dynamically creating workflows based on user demand then you build in graceful senescence. If a workflow is not used over time it is deactivated and eventually archived.

Data size. Now this is interesting. One of the big trends of the last 5 years has been the rise of big data as Google, Facebook et al struggle to extract value from the torrents of data that they gather. Most businesses are relatively small, and even for larger businesses relatively small numbers of people use the same processes. Large multinationals may have hundreds or thousands of Salesforce seats. Rarely tens of thousands and never hundreds.

Further as you move from the traditional computer domain to the human one you see an increasing use of unstructured data. It can be analysed but more complex problems on smaller data sets are harder to solve with confidence. This is why the combination of email with the CRM system is so powerful.

You can look at the unstructured email data and match it to the structured data in the CRM. Depending on the organisation training will take a shorter on longer period of time. but it can be done.

Conclusion

I’ll say again. As far as I know know one is doing this at the moment. We’ve started experimenting at the edges seeing what’s possible. There’s definitely an opportunity there. Some of the bigger speculations? You’ll have to start reworking a lot of Salesforce architecture from the ground up to get this to work for example. Enabling user specific workflows as a subset of organisational workflows is a big change in itself and probably not on the SFDC roadmap.

But we are starting to get the tools and the experience in using machine learning that now makes this a plausible approach and one that could if successfully implemented provide massive value to corporations and the job satisfaction of individuals

By

Why Technology Will Change the World, but Hasn’t Yet

In some respects computers have been disappointing. Whilst the internet, smartphones and spreadsheets have had huge impacts on many peoples lives and have made billionaires out of many entrepreneurs the impact has been small.

U.S. Total Factor Business Productivity
From The Economist

Data from The Economist and the FT shows this clearly. Since 1970 the growth rate of the US and many of the developed nations has slowed significantly which seems strange given the huge changes that advances in computers and technology have brought.

Put simply, if computers, and technology more broadly, were having a large impact on national productivity, one would have expected faster economic growth

This isn’t the case. One explanation is that whilst technology is responsible for some productivity growth broadly most people do most jobs in a similar way that they did in 1950 or even 1900.

The Steam Engine and the Electric Light

Let us step back a bit further in time for a minute and consider two key inventions: the steam engine and electricity. James Watt patented the steam engine in 1781 but it was only in 1850 that the UK started to the massive growth in productivity that is seen as the start of the industrial revolution.  Similarly whilst Edison got his patent for the electric light in 1880 it was only in the 1930’s that the US and UK created national grids and offered cheap, universal and reliable electrical power to their entire populations.

Both of these inventions took approximately half a century of improvement and massive innovation before they had a real impact on the economy. My argument is that the impact of technology has been similar to that of the steam engine and electricity in these gestation periods.

There are, I think, three aspects to consider when looking at technological change. Ubiquity, substitution, and enhancement.

Ubiquity, Substitution and Enhancement

 

By the start of the industrial revolution steam engines were ubiquitous, just as technology is today.  Without ubiquity there simply are not enough people spending enough of their lives thinking about how to use the technology to make money. There needs to be a critical mass to generate the innovations and improvements that create the ‘killer apps’.

For much of an innovation’s early life it is a substitute for another way of doing things. The classic expression of this is the 10hp rating of Watts first steam engine. One steam engine substituted for the power of 10 horses. Or to use a more modern example. One ENIAC substituted for the power of 9,000 computers (mostly women) at Bletchley Park.

But for it to break through to make a massive change in society it must go significantly beyond substitution to enhance productivity. The steam engines and its successors have gone unimaginably far beyond the limits of what could be provided by any number of horses. The same is true for electricity and it’s perceived ‘killer app’ the electric chair.

Most People – Including You – are Glorified Clerks

So where does this leave us with computers and technology now?

One test is for change is the impact on people’s lives. At it’s simplest the steam engine meant that you didn’t have to work as a farmer but could find more rewarding work in the cities. Electricity meant that you had 24 hours of daylight each day.

Computers will change the world by removing clerking. Clerking the administrative duties that need to be done in order to enable the real work to be undertaken. One study showed sales reps only spending 22% of their work week on actually selling but almost 65% of  their time on administrative tasks.

These tasks exists across almost every job and in almost all cases the educational, experiential and skill requirements to complete them are far below the person’s pay grade. Using the figures from above it is as you you are paying a salesman $100,000 but 65% of it could be done by someone on a salary of $20,000. That’s $52,000 of unnecessary expense.

And there’s No Escape

But, as is well known, these tasks require personal involvement. They can’t be done by someone else. Far less can they be done by a computer. Many of the activities require detailed background knowledge of the job and the tasks undertaken as well as an organisations SOP and procedures. It often takes months to onboard a new starter to being fully productive for just this reason – even if they are technically capable of the role

And this is one of the fundamental reasons why technology has not impacted productivity growth.

As an aside, one of the other reasons is that whilst some professions have improved productivity massively – think of the 20,000 Google engineers at Mountain View – others have barely changed; for example the 20,000 cleaners in Birmingham and every other city in the world. The impacts of the Googlers is swamped.

Programmatic Automation is Fragile

So we are at a cusp now. Companies such as Salesforce and Nintex have products that enable some of the admin work to be automated. However it is slow, it is expensive and it is fragile. If the business changes, if the way that you work changes many of these laboriously code workflows (and it is laborious even if it is point and click) have to be rewritten as they are now a cost rather than a benefit to your productivity. They make an organisation more efficient at the cost of sclerosis; a difficult tradeoff in a world where companies must be increasingly agile.

Machine Learning will be as Ubiquitous as  Sellotape

The answer is going to be in how we apply machine learning and all the associated tools of artificial intelligence, deep learning, big data etc etc.  For now these are being applied to data on a massive scale – think Hadoop – to handle things that are pretty trivial to economic productivity – information search and social popularity.

Where we will see a genuine revolution is when they can be applied simply and easily by individuals and small companies, with no technical skills, on many small unstructured data sets.

How can we redesign and apply these tools so that individuals can take them and apply them to their own jobs. Not all jobs are office jobs but all require administrative tasks that add little value to the individual (but great coordination and control value). How do you take that 65% of time wasted on administrative tasks and start automating it, taking it all away?

Incremental Improvements

There is not yet a solution. Nor will there be one for decades; but we can now start working at the problem at the edges. There is no reason why filling out expense forms should be anything more involved than taking photos of receipts; maybe not even that.

Managing and responding to email, arranging meetings, planning a day or a trip, keeping your boss informed, coaching new staff, responding to customers, reception duties, updating records and making notes. None of these things – and thousands more – should we really need to do

And in the decades to come we won’t

I see a world where businesses and individuals buy automation software. It plugs in to their lives, watches, learns and slowly takes the load away.

Dynamic Creation and Destruction of Automation Workflows

We see a world where business automation dynamically creates and destroys workflows in response to the way that people work within a culture. A world where based on your own writing style, your customer base, their historical engagement behaviour across multiple channels and the best performers within your organisation, the software suggests the emails that you write to your customers. Dynamically. Real time. Always improving.

 

Why Technology Will Change the World, but Hasn't Yet