How to Prioritize AI Use Cases in B2B Marketing

By Tom Swanson, Senior Engagement Manager at Heinz Marketing

Final month I wrote about using SWOTs to identify use cases for AI in B2B marketing workflows.  In the present day, I’m writing about evaluating these use instances.

I’ve labored with quite a few groups on bettering their advertising orchestration.  For the final 1.5 years that has meant “how do we integrate AI?”.  Clearly step one is how do you determine use instances that will be significant, after which need to prioritize them.

This framework is how I construction my pondering for groups of all sizes to deal with doing only a few issues at a time, very well.

The framework is easy: leverage – threat – chainability.

So first, slightly little bit of definition, and in line with the easy theme:

Leverage = how typically is the duty repeated?

Threat = how unhealthy will it’s if this will get tousled?

Chainability = does this activity produce an enter for one more activity?

It’s fairly simple how these 3 match collectively.  If the duty will save plenty of time/repetition, isn’t tough to repair if issues go awry, and produces an enter for one more related activity, then it’s a nice candidate for AI.

If, like me, you take pleasure in a easy visible, then lay these out on a cartesian aircraft.  Like this:

Chainability is extra of a sure/no on whether or not or not the output will go AS-IS into one other activity.  If the output will have to be reviewed/modified then it isn’t chainable.  This may differ by crew, so if you need people to overview/approve your marketing campaign technique earlier than the machine generates a tactical plan, then it isn’t chainable.  Different groups may not care a lot.

Working it this fashion offers you some flexibility in the way you rank issues alongside the x and y axes, with chainability as a Boolean gate or no less than as a big consideration.  Every little thing is a spectrum.

There are plenty of nuances in doing this.  A lot of it’s cultural.  The danger of a phrase being improper is likely to be big for one agency or business and small for one more.  You’ll have to do that subjectively for most of the use-cases.

Anyway, let’s get into some examples.

Leverage: How typically is that this activity repeated?

Rote, repeated duties are the very best match for AI, significantly early on if you are nonetheless studying to construct.  Extra complicated duties are simpler to interrupt down later, after you perceive how the machine thinks.

When evaluating leverage, I like to take a look at the scenario (learn: use case), the frequency that the duty is repeated, the price of doing the duty (normally: time), how the workflow differs from run to run, and eventually how uniform the output is every time in comparison with earlier runs.

Here’s a actual consumer instance:

Native Knowledge Miner

State of affairs: Regional reps wanted to remain on high of recent native companies as potential prospects.

Frequency: Weekly.

Value: ½ day.

Workflow: uniform, the identical each time.

Output construction: uniform, the identical each time.

Finally, this was a terrific candidate.  It has a daily frequency that may be mechanically triggered, and it prices a good period of time that could possibly be spent promoting.  The automation requires little upkeep as a result of the workflow and the outputs are the identical each time.  Modifications ought to be rare and additive.

Because it occurred, the danger profile and chainability had been additionally stable for this instrument.  A knockout!

Threat: What occurs whether it is tousled?

There are numerous, some ways issues can go improper.  Dangers are task-specific, have excessive variance between duties, and are weighted otherwise by organizations, groups, and people.  A crew with a excessive threat tolerance resulting from AI-automation mandates goes to simply accept a monetary threat quicker than a crew whose govt management is skeptical and measured about AI adoption.

I discover it most helpful to categorize dangers by sort after which basically do a low/mid/excessive to outline the danger inherent to the duty.

Right here is one other actual instance:

Automated Marketing campaign Technique with Human Assessment

Use case context: the consumer wished to automate their marketing campaign technique drafting with a crew of brokers that will pull CRM knowledge, net analysis, and buyer knowledge to develop marketing campaign methods that will expedite time to market.  You will need to notice that this is able to produce a deck for strategic overview by leaders, not push something into market itself.

Monetary threat: Low. Token-cost solely.

Rework threat: Excessive. Strategic drift doubtless.

TCO threat: Mid. Documentation must be up-to-date.

Model threat: Low. Offset by management overview.

Adoption threat: Mid. Crew would require coaxing to make use of it.

Accountability threat: Low. Offset by management overview.

I’ve but to codify a particular set of dangers that have to be assessed on a task-by-task foundation.  I feel monetary, model, and TCO are in all probability the large 3, however there are others.  Content material bots could have a distinct profile of dangers than a win/loss evaluation bot.

Chainability: How does this hook up with different duties?

For a fan of workflows, similar to myself, that is essentially the most attention-grabbing consideration.  How a bot matches right into a workflow is fairly simple.  It’s triggered both by a person or by one other bot, supplied and enter, after which it runs its activity.

The large query to ask is: when would you like a human overview?  Maximally chained bots take an enter from the earlier bot within the chain, do their factor, then cross the output to the subsequent bot.  This comes with some tradeoffs.

On the professionals:

• You get an output down the road quicker

On the cons:

• Issues require extra sleuthing to search out the damaged hyperlink
• Changes to early hyperlinks could require reworks to downstream hyperlinks

Personally, I prefer to chain as a lot as I can, however am cognizant of how tough it will likely be to make systemic changes ought to one thing change.

Right here is an instance from the bots I constructed for Heinz Advertising and marketing:

Use case context: We have to analyze web sites and mix that with different strategic inputs throughout our discovery course of.

Inputs: Shopper web site URL and key pages

Enter sources: Shopper information analyzer

Set off: Enter obtained

Output: Web site evaluation

Output vacation spot: Content material analyzer, messaging developer, marketing campaign strategist

This agent takes in a URL supplied from an analyst agent that critiques an information dump from purchasers.  It may also be kicked off manually, however that could be a ache.  The principle enter is the URL, so it’s straightforward to set off.

It additionally outputs into 3 totally different brokers.  That’s stable chaining proper there.  Not each agent can act immediately (for instance the marketing campaign strategist solely triggers as soon as a couple of different inputs are supplied), however it’s nonetheless a part of that chain.

Anyway, the higher the chainability, the higher the use case.

Conclusion

On the finish of the day, what you prioritize is as much as you.  This framework is a information for a way we attempt to maximize objectivity and work out the suitable locations to begin.  Nonetheless, in case you are extra superior, you may not want this framework as a lot.  Dangers that appear daunting at present is likely to be much less scary while you higher perceive these instruments.  Your particular setup could require extra human intervention and thus cut back the load of chainability.

The one manner to do that successfully is to make sure you begin by understanding your personal crew (so go learn the SWOT submit for those who haven’t already).  Start with this, after which work in direction of a roadmap.

Advertising and marketing, I feel moreso than different fields, has real opportunity from AI that can generate ROI.  This, in fact, requires it’s performed proper.

One very last thing: know you can actually solely construct 1 factor per crew at a time.  There may be different work to be performed, and this takes some trial and error.  That is much more true while you want inputs/outputs to chain collectively.  Construct one chain at a time, add parallelism modularly, and preserve a visible workflow so you recognize what matches the place.

Wish to discuss it?  Email us.

Wish to do some gap-finding?  Take our Orchestration Self-Audit.