How to Use AI - A Guide for Marketers
Everyone is talking about how transformative AI is, but the information is still very piecemeal.
There are tons of app focused listicles or cool tricks for ChatGPT, but if you're like me, you don't have time to go looking for ways to put apps to work.
So I set out to take a comprehensive look at the business of marketing, breaking it down into the fundamental jobs to be done and creating practical guidance for marketers on how they might incorporate AI into their daily work in each area.
This episode covers four main applications that have me excited, but if you visit the link below, you can access an Airtable base with over 35 use cases and over 40 apps.
Thanks to Our Sponsor
Many thanks to the sponsor of this episode - Knak.
If you don't know them (you should), Knak is an amazing email and landing page builder that integrates directly with your marketing automation platform.
You set the brand guidelines and then give your users a building experience that’s slick, modern and beautiful. When they’re done, everything goes to your MAP at the push of a button.
What's more, it supports global teams, approval workflows, and it’s got your integrations. Click the link below to get a special offer just for my listeners.
Key Topics
- [00:00] - Introduction
- [01:30] - Why make this episode
- [03:17] - Unstructured text analysis
- [05:27] - Structured data analysis
- [09:47] - Visual media generation
- [11:37] - AI and workflow automation
- [15:09] - AI for creative writing
- [17:15] - Two visions of AI usage
Resource Links
- A Very Gentle Introduction to Large Language Models without the Hype - A useful primer to AI and large language models.
Transcript
If you're in marketing, AI made 2023 a lot of things.
2
:Exciting, scary, overwhelming,
maybe all of the above.
3
:If I'm being honest, I've had some
mixed feelings about AI myself, and I'll
4
:talk about why later in the episode.
5
:But I can tell you now what I don't think.
6
:I don't think AI is another one of our
Buzzword y, flash in the pan marketing
7
:trends that we can safely ignore
because everyone will be talking about
8
:something different in six months.
9
:Even though, yes, it's overhyped,
yes, it's buzzword y, and yes, we are
10
:ruining it by slapping AI on every
app's homepage, AI is still important.
11
:Let's say, for a second, you
were a textile weaver on the eve
12
:of the industrial revolution.
13
:Your job is going to be
fundamentally changed.
14
:What the rise of mechanization did
for physical labor and what the rise
15
:of computers did for computational
labor, AI is doing it now to
16
:creative and analytical labor.
17
:Some tasks will disappear,
some jobs will disappear, and
18
:this will all take some time.
19
:But I think it's important for all
marketers to become AI literate
20
:and start understanding it today.
21
:Now, there is no shortage of AI
content out there, and I don't
22
:like to just add noise to the room.
23
:So why am I making this episode?
24
:Well, everyone is talking about
how transformative AI is, but the
25
:information is still very piecemeal.
26
:There are tons of app focused listicles
or cool tricks for ChatGPT, but if
27
:you're like me, you don't have time to
go looking for ways to put apps to work.
28
:You have business problems to solve,
and I haven't yet seen anyone take a
29
:comprehensive look at the business of
marketing, breaking it down into the
30
:fundamental jobs to be done and creating
guidance for marketers on how they might
31
:incorporate AI into their daily work
in each area, if that's even possible.
32
:So that's what I set out to
do, and if you go to revops.
33
:fm forward slash AI, you can
find an Airtable base I made.
34
:The main thing that makes it special,
in my point of view, is that it
35
:starts with the marketing use cases,
rather than with tools, and identifies
36
:where AI can help with each one.
37
:It then looks at some apps that can
support each use case and provides
38
:an analysis of both the potential
feasibility and the potential impact.
39
:To start with, there are 35
use cases and over 40 apps.
40
:And to be honest, I could have kept
going, but I do have a day job.
41
:It's a V1, and I hope to
keep it updated over time.
42
:I encourage you to go to revops.
43
:fm forward slash AI once the
episode is done to look at it.
44
:But in this show, I'm going
to give you the highlights.
45
:I'll summarize some key
learnings, my points of view on
46
:where you should be using AI.
47
:And also where maybe not to use it.
48
:And talk about the use cases
that have me most excited.
49
:By the way, I am assuming here that
you have basic familiarity with what
50
:generative AI and large language
models are and how they work.
51
:If you need a total 101 primer on
these concepts, that's no problem.
52
:Just pause here and go
check out the show notes.
53
:I'll include a link for you there.
54
:Okay, let's dive in.
55
:Now this first AI application is
one I've seen relatively few people
56
:talk about, but it's actually
the one that has me most excited.
57
:Using AI to analyze unstructured text.
58
:I'm excited about this because
of the potential impact.
59
:As marketers we've become increasingly
quantitative in our orientation over
60
:the last 10 years, but there is gold
in the form of customer insights
61
:located inside our call transcripts,
our interviews, social media posts.
62
:It's incredibly time consuming,
however, to analyze this data manually.
63
:Trust me.
64
:I've done it.
65
:It's repetitive work, it relies on
pattern recognition, it's pretty boring.
66
:This is a perfect application for AI.
67
:Machines are just better at
this, and they're less biased by
68
:proximity or by small samples.
69
:So as an example of a specific use
case in this area, how about processing
70
:your gong calls to extract things like
common pain points, specific competitors
71
:mentioned, or answers to questions
like, how did you hear about us?
72
:You could do this in a few ways.
73
:One way might be to export all the call
transcripts from Gong or whatever other
74
:call recording software you're using,
using that tool's API, and then having a
75
:developer run those transcripts through
an LLM, and summarizing the key insights.
76
:Alternatively, and probably the way
I do it, is to create a workflow
77
:in a tool like Zapier or Workato to
process those new calls as they occur.
78
:So these workflows could be
triggered any time a call matching
79
:your parameters is logged.
80
:The workflow would extract the transcript,
send it to an AI model for processing,
81
:and then the resulting output can be
fed into CRM or other database fields.
82
:This is something that's
totally achievable today.
83
:I like this because it keeps the
data accessible to CRM users,
84
:but it can also be aggregated and
analyzed quantitatively in reports.
85
:Another use case for unstructured
text analysis is document
86
:or video summarization.
87
:So, for example, there's
a tool called PandaChat.
88
:It lets you submit PDFs, documents,
websites, even video or audio files, and
89
:then you can just ask it to summarize
those assets for you, using whatever
90
:prompts make sense for that purpose.
91
:Now, I don't think this replaces
detailed reading or actually
92
:consuming the asset yourself.
93
:But it's a super useful research tool
for when you want to quickly identify
94
:whether that resource has what you need
and if it's worth exploring further.
95
:That's just a few of the
potential use cases here.
96
:Moving on from the analysis of
structured data, the second big
97
:bucket of applications for AI
I'm interested in is analysis of
98
:structured or quantitative data.
99
:The kind of data we typically report on.
100
:It excites me for kind of similar reasons.
101
:Here's a few specific use cases,
and a lot of these already have
102
:relatively mature solutions.
103
:You might even be doing
some of them already.
104
:The first one is creating
forecast projections.
105
:So I'm talking about business
or sales or marketing forecasts.
106
:Humans are generally
pretty terrible at this.
107
:Most companies, the forecast is almost
always wrong, it's just a question of by
108
:how much, and often that number is a lot.
109
:But if we summarize all the inputs
that go into a forecast, a machine can
110
:actually forecast pretty accurately.
111
:And a great example of an app
that does this is Groblox.
112
:Side note, I actually interviewed
their CEO Tony Holbein in episode 9,
113
:so you can learn more about it there.
114
:But one of the things that Groblox
can do is continually track all your
115
:metrics down to the tiniest KPIs.
116
:And continually calibrate a forecast
based on your actual business performance.
117
:So even the most disciplined and rigorous
human is going to have trouble managing
118
:all those inputs as efficiently, but
for a machine that's something it
119
:can totally do and produce a forecast
that is much more reliable and much
120
:more consistently close to reality.
121
:A system like this can even act as an
early warning system by detecting, for
122
:example, that the MQL to SQL conversion
rate for hand raisers in your German
123
:market It's 10 percent lower than
expected this week, and then indicating
124
:the projected impact on revenue two
to three quarters down the road.
125
:That sort of thing is magic, and again
is a use case where machines are just
126
:going to do a better job, or will do
a job that humans couldn't really do
127
:at all, to that level of granularity.
128
:Let's talk about recording
and analytics more generally.
129
:We've long employed analysts to help
turn data into insights, and I don't
130
:think those jobs are going away.
131
:But a lot of that work is simultaneously
easy for a machine, because it involves
132
:some relatively straightforward
trend or pattern recognition, and
133
:at the same time it can be kind of
intimidating for some business users.
134
:So this is a great place to bring in
AI and present users with insights
135
:summarized in plain language.
136
:For example, rather than giving the
user a report and requiring them to
137
:do 100 percent of the interpretation,
AI can provide a text based analysis
138
:or a data story to go along with it.
139
:Another use case is to actually generate
the report based on a text prompt
140
:rather than requiring a user to know
SQL or manipulate a reporting interface.
141
:So you can just say, show me the sales
trend in this geography for the past 12
142
:months and segment it by lead source.
143
:And it can just produce that for you.
144
:ThoughtSpot and Tableau are
two BI tools that already
145
:incorporate these features today.
146
:And my friend Grant even has
a startup called Moji that's
147
:creating these data stories with AI
specifically for your Marketo emails.
148
:So something that's already
happening and I think this is pretty
149
:obviously the way of the future
when it comes to data analysis.
150
:And what about paid media?
151
:If you work in Dimension, you
probably already use AI in some
152
:way to optimize your ad spend.
153
:Most platforms have some sort of
machine learning based goal setting
154
:that can allow you to optimize for
specific outcomes in your bidding.
155
:There are also third party platforms
like BrightBid or Albert, and there's
156
:a few others that I've put in the
Airtable base that do the same thing.
157
:Now assuming these work as
intended, this makes total sense.
158
:A computer can understand the millions
of potential variables that affect
159
:an auction better than a person.
160
:So why not let AI do that work?
161
:And what you may not be doing
is using these tools to actually
162
:generate new ad variants by AI.
163
:So, new headline, new copy, and playing
with those things to optimize your
164
:overall performance on that ad platform.
165
:That's a separate use case from the
data analysis, but it is something
166
:that you can do in some of these
platforms, even natively in Google
167
:Ads using their automatically
created assets feature, for example.
168
:Last but not least in the area
of data analysis, let's think
169
:about good old fashioned lead
scoring and account scoring.
170
:This is actually one of the
original AI use cases for marketing,
171
:going all the way back to when
it was called predictive scoring.
172
:There were tons of new startups
doing it in the mid to late:
173
:And unlike traditional rules based
scoring, which uses static, predefined
174
:values, predictive scoring uses
more sophisticated machine learning
175
:models to, uh, well, predict which
leads are most likely to buy.
176
:And there are vendors like SixthSense
or MadKudu doing it today, and you
177
:might already be using one of them.
178
:So let's move on to the third big
bucket of AI applications I want to talk
179
:about, which is visual media generation.
180
:So this is an interesting one for me, as
I'm actually not a huge fan of generative
181
:AI when it comes to creative writing
and more on this in a few minutes.
182
:So, why do I think imagery is different?
183
:Possibly because I can't draw.
184
:No, really, I think the main
reason is that most imagery in B2B
185
:marketing is already pretty bad.
186
:Stock imagery is usually kind of lame,
and it's hard to find something that's
187
:super relevant or that's personalized
to a subject that you're working on.
188
:Now, with AI, you can create imagery
that is way more personalized and
189
:specific to your subject matter with
a lot of fine grained control and
190
:just continually revise that image
to get almost exactly what you want.
191
:And it generally looks
just as professional.
192
:I'm thinking here about feature images for
blog posts, social media, even websites.
193
:The advances in AI image generators like
Midjourney or Dolly over the past year
194
:have been pretty astounding and they're
only going to continue to get better.
195
:Now I don't think these tools are ever
going to create like truly original
196
:masterpieces of art, but that's not really
what we're looking for in marketing.
197
:We just need images that look
good, that are relevant, that
198
:are effective for the purpose.
199
:And I think AI can provide that
in most cases, and usually at a
200
:lower cost than human illustrators.
201
:The far out thing is that going
forward, the same advancements are
202
:going to apply to video as well.
203
:If you look at most of the, you
know, what's ahead for:
204
:newsletters and lists, AI generated
video is usually the top of them all,
205
:at least the ones that I've read.
206
:I still think there's going to be a
place for human produced feature videos
207
:and marquee pieces of content, but my
expectation is that AI tools are going
208
:to eat up The commodity bottom end
of this, and it's going to make video
209
:more accessible to use in places where
we may not have used it at all before
210
:for cost reasons or for time reasons.
211
:So if you can spend 20k with an agency
or a few hours with AI and get relatively
212
:similar results, it's going to be
harder and harder to justify that spend.
213
:The last bucket of use cases
I'm excited about is in the
214
:area of workflow automation.
215
:And this is in part because I'm just
an automation geek in general, but when
216
:you add AI to automation, you see a huge
expansion of what's possible versus just
217
:using predefined, rules based automation.
218
:For example, you may use an automation
platform like Zapier, Mercado, or Trey,
219
:and within those workflows you can now
include action steps that reference
220
:large language models like ChatGPT.
221
:You can even call ChatGPT from a
marketing automation platform that
222
:supports outbound webhooks like Marketo.
223
:So this opens up a wide range of potential
use cases to leverage AI in a completely
224
:automated way, rather than needing
a human to interact with a chatbot
225
:and type prompts into that interface.
226
:So for example, Within those workflows,
you could generate customized email
227
:text based on a lead's attributes.
228
:You could cleanse lead data and
feed it back to a source system.
229
:You could send an AI generated
custom alert to your internal team
230
:based on certain sales triggers.
231
:There's a lot of different use
cases you could think about here.
232
:In addition to using AI inside
your workflows, all the main
233
:workflow builders also offer the
ability to generate those workflows
234
:themselves using AI via text prompts.
235
:So for example, you could log into
Zapier and you can use a prompt like,
236
:create a workflow that's triggered
whenever I have a new lead in Marketo
237
:with a source of partner and add
it to such and such a Google Sheet.
238
:Now the challenge I have with these
natural language workflow builders is that
239
:there does seem to be a last mile problem.
240
:So in my experiments using Zapier,
the workflows were on point.
241
:for simple requests, but often a
bit wonky for more complicated ones.
242
:So it's not like you can just hand this
over to a user that doesn't know how
243
:to use the system and let them at it.
244
:You still need a sophisticated builder
to understand and debug the output, and
245
:it may or may not save you a ton of time.
246
:That being said, even if this is still
half baked today, it won't be long, I
247
:suspect, until they're more mature and
are able to produce more reliable outputs.
248
:Here's another completely different
example of workflow automation applied
249
:to the domain of media production.
250
:So you can use AI to do things like
generate transcripts of audio and
251
:video, automatically clean up filler
words like um and uh, shorten word
252
:gaps, process a longer file to produce
shorter clips for social media that
253
:are automatically formatted for
different platforms, and so on.
254
:This is the kind of non value
added work that takes a lot of
255
:time and doesn't necessarily
require a human's creative input.
256
:Again, though, the challenge I
found in practice is that a lot of
257
:the tech is still not fully baked.
258
:For example, in Descript, which I
use as my podcast editing software,
259
:removing the filler words often
results in clipping an adjacent word
260
:or producing an ugly transition.
261
:So it still requires manual cleanup.
262
:In practice, I still often
find it's easier to just do
263
:the whole thing manually.
264
:Now this really sucks.
265
:I hope we get to a place where it's
reliable 99 percent of the time because
266
:it will save me personally a lot of work.
267
:I think inevitably we will get there.
268
:Last workflow use case and one that's
near to my heart is documentation.
269
:It can be a real pain to document a
system by hand, screenshotting each
270
:step, typing out instructions, and so on.
271
:With tools like Scribe, you can simply
record your screen while executing a
272
:process, and it's going to automatically
capture that process and then summarize
273
:it with both text and screenshot.
274
:I've used Scribe myself,
and it was pretty magical.
275
:I still edited the output and
cleaned it up, but what it did was
276
:get me 90 percent of the way there.
277
:And the result is something
that's way clearer, and that's
278
:a delight for people to use.
279
:So that's it for the use cases
I'm really excited about.
280
:Obviously there's so many more
examples and apps I could have covered.
281
:Those are just some of the main
ones that are top of mind for me.
282
:Something that you might want
to start thinking about as well.
283
:I now want to just touch lightly
on an area where I haven't actually
284
:been that impressed with AI.
285
:Where I'm not even sure, ultimately, how
appropriate or beneficial it is to use.
286
:It relates to creative writing, and I'm
being somewhat contrary in here, as this
287
:is where most people seem to get really
excited about AI, and there are lots of
288
:popular apps focused just on this use
case that people seem to really like.
289
:So maybe I'm missing the boat, but where
I've experimented with it, and it is
290
:where I've experimented the most, I've
really found that almost 100 percent
291
:of the output was not fit for purpose.
292
:Now to be clear on what I mean, I
think AI can do a fine job at producing
293
:transactional text, things like Summaries.
294
:Or anything where the clarity of
communication is the main objective.
295
:But when we think of text that needs
to persuade, needs to entertain,
296
:to educate, to build affinity.
297
:Things like content marketing, longer
form copywriting, social media posts.
298
:Anything that really requires the
spark and imagination of a human
299
:being, I just don't think it works.
300
:What I have found is that AI does a
great job creating middle of the road,
301
:conventional, generic sounding content.
302
:So, yeah, if you need to produce
commodity SEO articles and don't care
303
:about originality or how valuable the
content actually is, Maybe that's fine.
304
:What it can't do, by its nature,
is produce novel, original, or
305
:unconventional points of view.
306
:Something that goes beyond the status
quo, and actually advances the state
307
:of knowledge in that particular area.
308
:However, that ability, the
ability to do that, is what makes
309
:content actually worth reading.
310
:Now, I suppose you could provide
your original points of view in
311
:bullet form, and have AI do a
first draft that you then finesse.
312
:And maybe this is going
to save you some time.
313
:But for me as a writer, so much of
the communication is also in how
314
:you choose to present your thoughts.
315
:The order, the word choice.
316
:That's what makes it you.
317
:That's what makes it who it is.
318
:And I don't want to outsource
those decisions to AI.
319
:Ultimately, I feel like creative writing
is a fundamentally human activity.
320
:Like, would you want to listen
to this podcast if you were just
321
:hearing me read out an AI script that
represented the average of the average
322
:of what everyone else was saying?
323
:What would be the point of that?
324
:So I have strong feelings here,
and if you think I'm missing the
325
:boat, feel free to reach out and
let me know why you disagree.
326
:Maybe I'll evolve.
327
:In conclusion, I want to leave
you with one parting thought.
328
:I think in some ways we stand at
the crossroads of two potential
329
:ways of relating to AI as a society,
and I'll express them in terms
330
:of two popular works of fiction.
331
:For the positive vision, we have Star
Trek, and like so many developments
332
:in technology, the Star Trek
universe has really anticipated the
333
:rise of generative AI decades ago.
334
:If you're familiar with any of the
series, Star Trek ships have massively
335
:powerful computers and they can
accept natural language prompts and
336
:produce almost anything, including
lifelike 3D replicas of people and
337
:places via holographic generation.
338
:Now there's still risks to this
technology, but for the most part in the
339
:shows, humans remain firmly in control.
340
:They treat the computer like
what it is, a powerful tool that
341
:can amplify human capabilities.
342
:And assist them in their tasks.
343
:Ultimately, it's something that
makes them stronger, not weaker.
344
:I would argue.
345
:On the flip side of this, we have
my personal favorite Pixar movie,
346
:Wally, and in this world we have the
autopilot of a spaceship on which
347
:a group of humans have escaped an
environmentally ravaged planet earth.
348
:The autopilot and the ship ecosystem
handles their every need for that
349
:group of humans, leading them to
become essentially adult babies.
350
:Completely dependent consumers
who no longer play an active role
351
:in managing their own destinies.
352
:So obviously it's pretty clear which
vision of AI use I'd prefer, and I'm
353
:sure almost everyone listening would
prefer the same in principle, but to
354
:me the WALL E example is a cautionary
tale about outsourcing too much
355
:decision making, too much creativity
to the machine, and that's how I aim
356
:to apply it in my own day to day use.
357
:I'm going to wrap it up here, but
just a reminder to go to revops.
358
:fm.
359
:ai to see the full
Airtable base of use cases.
360
:I would honestly love your thoughts
and feedback, so please comment on
361
:the post and share if you agree, if
you disagree, or anything in between.
362
:Thanks for listening.