Why SMEs Are The Real Bottleneck (Not Resources. Not AI)

WEBVTT
00:00:00.080 --> 00:00:04.480
Hello, I'm Stephanie with Argos, back here for another episode with Field Notes.
00:00:04.480 --> 00:00:08.240
And today I'm with Erik, who's an independent advisor with Vogt Strategy.
00:00:08.800 --> 00:00:11.279
And he's an expert in a lot of things in the field.
00:00:11.279 --> 00:00:15.199
So we love having him on to talk about all these cool concepts.
00:00:15.199 --> 00:00:19.120
And today's concept, we're going to be talking about SMEs.
00:00:19.120 --> 00:00:30.160
So, Erik, you've said that the real bottleneck in localization isn't AI or resources, it's subject matter expert feedback loops.
00:00:30.160 --> 00:00:32.479
What is really the core issue here?
00:00:33.200 --> 00:00:37.200
Yeah, the challenge is that AI is now ubiquitous.
00:00:37.200 --> 00:00:39.920
Translation is no longer a constraint.
00:00:39.920 --> 00:00:43.280
So it's relatively easy to make a translation happen.
00:00:43.280 --> 00:00:49.359
The challenge is in knowing what is truth, what is intent, and the accountability.
00:00:49.359 --> 00:00:55.520
We've talked about that before, that accountability in the risk aspects of quality are an important part of it.
00:00:55.520 --> 00:01:05.519
But without somebody who actually knows what the product does in the market and in the context, localization, any kind of automatic solution is largely guesswork.
00:01:05.519 --> 00:01:08.480
And that can open up some significant risk for organizations.
00:01:08.719 --> 00:01:12.319
Yeah, and I'd imagine especially in the enterprise space.
00:01:12.319 --> 00:01:14.560
So maybe let's focus there a little bit.
00:01:14.560 --> 00:01:18.400
Why are SMEs such a bottleneck, especially in that space?
00:01:18.719 --> 00:01:25.280
Yeah, and this is probably a good idea to differentiate between the different kinds of subject matter experts, but there's internal ones and there's partner SMEs.
00:01:25.280 --> 00:01:29.599
So the internal ones could be within an organization who really deeply understand the product.
00:01:29.599 --> 00:01:36.000
They have a constraint in that they're a full-time resource often, and that they often have other jobs to do.
00:01:36.000 --> 00:01:40.159
So reviewing things might not be their primary responsibility.
00:01:40.159 --> 00:01:43.920
But even translation might not even be their formal responsibility.
00:01:43.920 --> 00:01:49.439
Then there's partner SMEs, and I'm thinking about these as sort of the extension of the sales engine sometimes.
00:01:49.439 --> 00:01:58.400
You might have in-country sales reps or partner entities that know about their product, what is available in their country, and what those services offer.
00:01:58.400 --> 00:02:07.920
But there's another branch of SMEs, and that is a subject, a linguist who has invested in the time to really understand deeply a product or domain space.
00:02:07.920 --> 00:02:17.360
And they are restrained in with regards to time and capacity, largely because they're being plugged in via an external business ecosystem.
00:02:17.360 --> 00:02:25.680
So I think in general, either localization is not their priority or they are not fully plugged into the full expert ecosystem.
00:02:25.680 --> 00:02:29.199
There's just not a good ownership model for this.
00:02:29.199 --> 00:02:35.039
I think that we don't buy experts, we buy words and we buy hours.
00:02:35.039 --> 00:02:46.639
And that transaction tends to bury what the value of that subject matter is and how to inject that expertise into a into the workflow where it's needed.
00:02:47.120 --> 00:02:47.439
Okay.
00:02:47.439 --> 00:02:52.000
So then when we're looking at this, of course, AI comes into the picture.
00:02:52.000 --> 00:02:56.800
And AI, it's made that that translation bit or at least the words part very fast.
00:02:56.800 --> 00:03:01.759
It's helped out a lot with maybe versus traditional machine translation.
00:03:01.759 --> 00:03:03.199
The fluency is better.
00:03:03.199 --> 00:03:08.960
Shouldn't it also kind of help out somehow to reduce the burden on these SMEs?
00:03:09.360 --> 00:03:22.560
It could, if the data is available for the model to produce accurate results, which is and it's interesting in the sense that now you could go to AI to get answers to become an expert in a topic that the LLM has some information about.
00:03:22.560 --> 00:03:33.680
But the challenge here is that you're still going back to the trust and the context, both of which are generally not something that you can outsource to a large language model.
00:03:33.680 --> 00:03:46.400
So you have the hallucination, you have subject meaning drift, you have product misrepresentation, and you'll see that AI, when it's doing translations, generally can't really do fuzzy matches very well.
00:03:46.400 --> 00:03:54.080
And it tends to focus more on kind of the most probable outcome, the most probable unit, the entire unit.
00:03:54.080 --> 00:03:58.719
And that can sometimes create tonal errors and that can lead to brain damage.
00:03:58.719 --> 00:04:01.599
So there's a variety of different uh risks that come up here.
00:04:01.599 --> 00:04:12.240
But here's what's really interesting because AI makes fewer ugly and easily identifiable errors, it's making more believable mistakes that are harder to detect.
00:04:12.240 --> 00:04:15.280
So you have this kind of two different forces going on.
00:04:15.280 --> 00:04:21.920
One is yes, you can easily produce a lot of content to help subject matter experts to be better informed.
00:04:21.920 --> 00:04:29.920
And it becomes even more intense and more difficult to find the actual mistakes because the way that it's being presented is so believable.
00:04:30.240 --> 00:04:33.839
So then let's like zoom out into the other end of the spectrum, right?
00:04:33.839 --> 00:04:42.959
So we have translators who are one of these SMEs, not necessarily the in-country reviewer type of SMEs, but that second type of SME that you've been talking about.
00:04:42.959 --> 00:04:54.240
And they're saying that work is evaporating, or um we've seen many like educational programs for translators, interpreters, et cetera, going under these days and the rates keep dropping.
00:04:54.240 --> 00:04:55.680
Where's the disconnect here?
00:04:55.680 --> 00:05:02.959
We're saying these are very valuable, they have that kind of expertise, that knowledge, and it's a rare resource, but then we're not seeing that reflected in the market.
00:05:02.959 --> 00:05:04.319
Where's the disconnect?
00:05:04.639 --> 00:05:07.920
Yeah, I think this is where the system is differentiating human contributions.
00:05:07.920 --> 00:05:11.680
So it's really, really easy to produce AI deployments.
00:05:11.680 --> 00:05:16.399
You can spin up an AI tool in a matter of days, it can produce an enormous amount of outcome.
00:05:16.399 --> 00:05:18.399
But how do we manage that output?
00:05:18.399 --> 00:05:23.199
With in with SMEs, it's hard to focus their time in a meaningful way.
00:05:23.199 --> 00:05:26.480
And all industries say that they want expertise.
00:05:26.480 --> 00:05:29.920
You know, they want to have the translations done by somebody who knows what they're doing.
00:05:29.920 --> 00:05:30.800
It's obvious.
00:05:30.800 --> 00:05:38.160
And yet, translators are being commoditized at the exact moment we need them to be acting as domain experts.
00:05:38.160 --> 00:05:46.240
And so we have SMEs are missing from the equation because they're either too busy or they're not engaged enough, as we talked about before.
00:05:46.240 --> 00:05:48.000
So that increases risk.
00:05:48.000 --> 00:05:58.240
The curement engines and the tendency, and that happens at multiple layers within the organization, tends to drive unit prices down, and then that pushes out experts, right?
00:05:58.240 --> 00:06:07.360
So you end up buying the cheapest of a type of a resource, and then less people are motivated to get into the industry as translators or subject matter experts because of this.
00:06:07.360 --> 00:06:14.480
Then quality becomes harder to manage because you have fewer experts there looking at the process.
00:06:14.480 --> 00:06:24.720
And that generally will, if you think of the U-shaped cost quality curve, that tends to increase rework and the consequences of bad, like missing these translations.
00:06:24.720 --> 00:06:30.399
So enterprise complains about inconsistency, blames a vendor, and the cycle continues.
00:06:30.399 --> 00:06:36.000
And so you end up with a kind of a treadmill of diminishing returns for the entire system.
00:06:36.000 --> 00:06:47.519
You're isolating the vendor of the best, most expert eyeballs, the extra brains that we want to have associated with the review of this.
00:06:47.519 --> 00:06:51.600
It's harder and harder to focus our time in the most meaningful way.
00:06:51.600 --> 00:06:56.560
So we end up devaluing the exact people who are the most critical to reduce risk.
00:06:56.800 --> 00:07:00.720
So what, or I guess I guess is procurement not seeing this issue?
00:07:00.720 --> 00:07:02.480
Or how come we aren't seeing them yet?
00:07:02.480 --> 00:07:07.040
Trying to like fix this issue, looking in into the future, it could get much worse.
00:07:07.360 --> 00:07:08.240
Yeah, for sure.
00:07:08.240 --> 00:07:10.160
So I think there's several different things.
00:07:10.160 --> 00:07:23.040
Some procurement folks are realizing that you need to make sure that you've got the individuals who know what they're doing associated with your work, and they're protecting the erosion of those costs.
00:07:23.040 --> 00:07:29.439
So there's a group of this industry where procurement is, in fact, I want to acknowledge that because it's an important part of the equation.
00:07:29.439 --> 00:07:47.759
However, when you're driving for cost reduction, then the problem becomes people are shopping around for lower price, and they tend to believe that AI plus humans can sh can and should deliver a savings that lowers the top line expense and that creates some distortions within the supply chain.
00:07:47.759 --> 00:07:50.959
But the essence of it is what you can measure.
00:07:50.959 --> 00:07:58.560
How many conversations are there about scoring or quantifying domain familiarity or some credentialing system?
00:07:58.560 --> 00:08:03.600
Like we know, okay, it's a translator that has a degree, that's enough, or they do a test, that's enough.
00:08:03.600 --> 00:08:13.360
But we really don't have a good mechanism within our organization as a whole to measure performance as tied to business risk.
00:08:13.360 --> 00:08:16.399
And the pricing model for decision making is not there.
00:08:16.399 --> 00:08:20.480
We tend to think in terms of productivity efficiency how many words per hour can you do?
00:08:20.480 --> 00:08:22.000
How fast is this?
00:08:22.000 --> 00:08:29.759
And yet, when you talk to real translators who know what they're doing, they're like, well, this is not making my life easier.
00:08:29.759 --> 00:08:33.200
This is, you know, expertise isn't manageable.
00:08:33.200 --> 00:08:39.200
If expertise isn't measurable and it and it's hard to measure, then it becomes unpers purchasable.
00:08:39.200 --> 00:08:45.759
Like just trying to put a procurement line item in there of like, please rate the skill level of your resources.
00:08:45.759 --> 00:08:59.200
Now, there are some ways around this, and we can talk about that in a minute, but we really are constrained by 20 years of buying habits oriented around a unit rate, specifically the word rate.
00:08:59.440 --> 00:09:02.159
So then maybe let's pull it back down to okay, what can we do?
00:09:02.159 --> 00:09:04.080
The actionable insights here, hopefully.
00:09:04.080 --> 00:09:13.200
So, as localization leaders, experts, if they're in their own company, what can we do differently in order to try to help the situation?
00:09:13.679 --> 00:09:15.679
The theory of constraints is an interesting.
00:09:15.679 --> 00:09:17.919
There's a book about this, you can look it up.
00:09:17.919 --> 00:09:19.279
It's an interesting theory.
00:09:19.279 --> 00:09:26.879
It basically treats the most valuable entity and most expensive entity in the loop as in a different kind of a way, right?
00:09:26.879 --> 00:09:32.480
You're not trying to optimize down for unit price for that most expensive component.
00:09:32.480 --> 00:09:36.799
You're trying to make the utilization of that component as valuable as possible.
00:09:36.799 --> 00:09:41.519
So you tend to, for example, create a backlog so they can manage their efficiency.
00:09:41.519 --> 00:09:50.240
And they they are continually busy, so they're optimized around keeping them busy, and you focus on making them as uh effective as possible.
00:09:50.240 --> 00:09:53.200
So let's just park that for a second of a mindset.
00:09:53.200 --> 00:09:59.679
Then we have something that a lot of organizations are doing, which is differentiating by content type.
00:09:59.679 --> 00:10:07.440
And I think this is really helping to accordion out the different risk profiles of the you of the work that they're doing.
00:10:07.440 --> 00:10:12.559
So you have high-risk and low risk types of activities and you create processes that are appropriate for that.
00:10:12.559 --> 00:10:24.480
But I think at a more substantive level, we really need to rethink how we're using AI, how you're building workflows to optimize around the skills that you're trying to look for.
00:10:24.480 --> 00:10:35.600
So, how how, for example, could you look for consequences of failure and then tie that into what you're paying to make sure that you're mitigating those costs for sale of failure?
00:10:35.600 --> 00:10:39.759
You think about that value of that step of that human in there.
00:10:39.759 --> 00:10:43.039
What are they actually delivering and what are the consequences of not doing it?
00:10:43.039 --> 00:10:45.759
Don't just commoditize everything.
00:10:45.759 --> 00:10:46.960
It tends not to work.
00:10:46.960 --> 00:10:54.399
So now you have uh the serious players in the room, they're doing things like routing things on different workflows based on those risk profiles.
00:10:54.399 --> 00:11:00.080
And that really helps to make sure that you're putting the human effort in the most valuable place.
00:11:00.080 --> 00:11:05.440
I think also it means maybe thinking about what expertise is.
00:11:05.440 --> 00:11:06.480
You know, what is it?
00:11:06.480 --> 00:11:17.600
It's knowledge about the product that they're supporting, it's understanding the style guide intuitively, understanding the tools, workflows, the preferences of a particular workflow, and thinking about how valuable that is.
00:11:17.600 --> 00:11:30.879
Now, it when you do that, you start making different trade-offs and you start seeing instead of optimizing around work price, you start optimizing around the scheduling of these valuable resources.
00:11:30.879 --> 00:11:34.639
So you say, I would rather have this person who does know what they're doing.
00:11:34.639 --> 00:11:41.919
I want them in the loop, as opposed to I want it when I deliver it because it's more convenient for me to fit that into my sky cycle.
00:11:41.919 --> 00:11:48.000
That makes the most valuable resource essentially not available because they are valuable on other projects or other things.
00:11:48.000 --> 00:11:56.000
So you plan your workflow with plenty of heads up, plenty of managing the prioritization that you're really putting into these specialists.
00:11:56.000 --> 00:12:10.399
So I think also we need to think about, you know, if we're over-indexing on price and turnaround time, uh, and this isn't just the buyers, it's the LCPs, it's also the linguists, like the whole system is biased around a certain way of thinking about this.
00:12:10.399 --> 00:12:14.480
That's confidence is what you're selling, trust is what you're selling.
00:12:14.480 --> 00:12:25.200
So the degree to which you can amplify reasons to trust or accountability for individuals, we're almost getting back to the world of trust your translator, like know who they are.
00:12:25.200 --> 00:12:36.399
That individual person knows what they're doing for your product and are the best fit as opposed to just rapidly deploying N number of random translators as quickly as possible.
00:12:36.399 --> 00:12:46.960
So I think in general, when we think about what's happening with this market, where the translation is plentiful, there's tons of it, it's just saturating our market.
00:12:46.960 --> 00:12:53.120
Yeah, we do, and that's eroding, it's causing margin compression for everybody in the loop.
00:12:53.120 --> 00:12:59.519
And not to mention macroeconomic factors like weaker dollar, which is affecting our whole industry as well, is another thing.
00:12:59.519 --> 00:13:23.360
I if I had a one less one thing I'd want to say to everybody in this system is we need to not forget that SME scarcity and the limited number of people who actually know your product and actually know how to fix it and the investment in them and the management of their critical time is really the new localization tax.
00:13:23.360 --> 00:13:37.200
And it's going to be even the closer and closer we get to AI delivering more and more close enough content, the ability of those subject matter experts to validate truthiness or truthfulness is more important than ever.
00:13:37.200 --> 00:13:40.559
That's kind of the main takeaway, I think, of this whole this whole topic.
00:13:40.879 --> 00:13:43.440
We're out of time for today, but thank you so much, Erik.
00:13:43.440 --> 00:13:45.519
And yeah, this was an interesting one.
00:13:46.240 --> 00:13:49.039
I I hope this spawns some conversation.
00:13:49.039 --> 00:13:53.840
I think this is an important topic for it's a it's affecting the entire industry at the same time right now.
00:13:53.840 --> 00:13:55.759
So it's an important one for us to be talking about.
00:13:55.759 --> 00:13:57.440
Steph, thanks so much for the time.
00:13:57.440 --> 00:13:59.200
It's a pleasure, as always.
00:13:59.840 --> 00:14:00.399
Pleasure.