In this episode of Future Proof HR, Jim Kanichirayil sits down with Robert St-Jacques, VP of People at Apera AI, to talk about what it takes for HR teams to scale AI without losing control. Robert shares a practical framework for thinking about AI governance in different stages of growth, with a focus on how HR leaders can move quickly without opening the door to unnecessary risk.

The conversation centers on what safe AI adoption actually looks like in practice. Robert breaks early AI governance into three buckets: use cases, data, and controls. He explains what makes a true green-win use case, why internally controlled documentation is the safest place to start, and where human review has to remain in the loop before HR hands more work to AI systems.

Jim and Robert also get into what changes as organizations mature. They cover how governance has to deepen as scale increases, where chatbots need stronger fail-safes, and why AI should never be left to make subjective decisions about people, pay, or candidate comparisons. The result is a grounded conversation for HR leaders who want AI efficiency without losing accountability, trust, or operational control.

Topics Discussed:

• Balancing AI speed with governance, accountability, and control
• What a green-win AI use case looks like in HR
• Why internal documentation is the safest starting point
• How human review should work in early AI workflows
• Where candidate scoring and comparison create risk
• How lawsuits are shaping HR AI caution
• What changes when AI programs mature across larger organizations
• How to build chatbot fail-safes with citations and source links
• Why manager enablement matters in AI-powered learning and development
• A five-part checklist for AI governance in HR

If you are an HR leader trying to scale AI adoption with more clarity and less risk, this conversation offers a practical framework for deciding where AI can help, where human judgment still matters most, and how to put guardrails in place before small mistakes turn into bigger operational problems.

Additional Resources:

• Cleary's AI-powered HR Chatbot
• Future Proof HR Community
• Connect with Robert St-Jacques on LinkedIn

where do things go sideways?

2

What are the things that always

get screwed up when they're

3

looking to make that leap?

4

Bob St-Jacgues: AI drift, which

means I've got Claude at home, I've

5

got ChatGPT I use Grok, I use this.

6

They go to work and then start dumping

documents, from their workplace.

7

IBM said, Watson has figured out

when somebody's about to quit.

8

This is where the maturity comes in,

You could, everybody in HR just gasped.

9

counterintuitively, it's actually

giving more time is worse.

10

Parkinson's law, So people

ascribe a level of complexity

11

to a request based on the

time you're given to do it,

12

Speaker: Whether you're starting out

with AI or you're mature in AI, one of

13

the things that's critical for every

organization to have clearly defined

14

is what are the rules of the road?

15

What are the minimum viable guardrails

that we need to adhere to before we

16

launch headfirst into this initiative?

17

And whenever you're thinking about the

world of AI, one of the things that

18

you often notice is that there seems

to be a mad dash to just get something

19

in play and when you're often operating

with a level of urgency and not having

20

planned the step-by-step process,

that can often get you in trouble.

21

Now, what trouble looks like for a startup

is gonna be different than what it looks

22

like for a more mature organization.

23

But regardless of where you are in

your AI journey, understanding what the

24

guardrails should look like for your stage

of development is a critical part of the

25

process in successfully implementing AI,

and that's what we're gonna cover today.

26

In today's conversation, we're gonna

take a look at what the governance

27

model should consider when you're

early in your AI journey and what the

28

governance model should include when

you're more mature as an AI organization.

29

So the person that's gonna guide

us through that conversation is

30

joining us today, Bob Saint-Jacques

is a people, culture, and

31

digital transformation executive.

32

He's an international employment lawyer,

university professor, and keynote

33

speaker with 30 years of experience

scaling and modernizing organizations

34

across 29 countries and 11 industries.

35

He's built, optimized, and digitally

transformed 11 people and culture

36

functions and led 42 HR technology

implementations with deep expertise in

37

distributed and hybrid operating models.

38

He currently serves as the VP of

People at Apera AI, and he's supporting

39

rapid growth while strengthening

global people systems governance, and

40

cross-border employment frameworks.

41

Bob is a fellow with the Center for

Evidence-Based Management and brings

42

an evidence-based, metrics-driven

approach to hiring, performance,

43

and organizational design.

44

Jim Kanichirayil: Bob,

welcome to the show.

45

Bob St-Jacgues: All right.

46

thank you, Jim.

47

Jim Kanichirayil: I'm looking forward to

this conversation and that's gonna be a

48

kind of an odd thing to say, considering

that we're gonna be talking governance

49

model now before anybody clicks off,

this is gonna be a good conversation.

50

So you wanna stick around, especially

in the context of AI and HR and

51

building an AI initiative that

adheres to governance models and

52

how you actually do that, right?

53

So don't click away.

54

So with that being said, I

want to set the stage here.

55

And get a little bit more detail,

behind your story and some of the

56

things that you've been involved

in throughout your career that

57

sets the stage for this governance

conversation that we're gonna have.

58

Bob, why don't you share with the

class, some of the ins and outs of

59

your career and story and how that

leads into the governance conversation

60

that we're gonna have today.

61

Bob St-Jacgues: All right,

thanks for that, Jim.

62

So my career in terms of as it relates

to HR started, a couple years ago and

63

so we've gotta go back to the early

nineties where I worked in Parliament

64

of Canada drafting labor legislation.

65

much of which is still in, in place today.

66

I worked with a bunch of folks

called lawyers and decided,

67

hey, they got stuff done.

68

So went to law school and practiced

international labor and employment

69

law as happens to many of us.

70

I was hired by one of my clients,

which was LensCrafters at the

71

time, and thus began my career.

72

So I've worked in Fortune 150 companies

then, relocated overseas to Dubai and

73

worked for, as you could probably imagine,

74

a hypergrowth area, so got to test there.

75

And since then I've been mostly

working with startups and tech

76

companies in Southeast Asia and

77

In that time, what I've done is

combined my passions, so to speak.

78

So first of all, it's the

employment law aspect.

79

Secondly, it's the scale up aspect.

80

we're looking at digital transformations.

81

And lastly, what brings a lot of

this together is creating systems.

82

Not just a one-off project, not an

email, but you know, to put together

83

systems so that anything that is

improved or optimized is sustainable

84

and continues to, be in place long

after I leave in, in, in those areas.

85

And what I've learned, especially

starting in Dubai, you can imagine the

86

change management aspect of trying to

tell a company that's had 300% growth

87

for the past three years that they're

quote unquote leaving money on the table.

88

So I had to change the way I

did things, and pitch things.

89

And so that's why I learned

about the, importance of systems

90

and putting structure together.

91

Jim Kanichirayil: So one of the

interesting things about what

92

you described is that you're

typically working with, scale up

93

organizations or organizations going

through a digital transformation.

94

One of the things that I would imagine

that would be challenging, especially in

95

a scale up environment, is that scale ups

are a little bit different than startups.

96

I usually work in the early stage

startup phase where it's run fast

97

and break things, and then when you

enter into a scale up, you have to

98

put in a lot more operational rigor.

99

And the thing that I'm curious about is.

100

From your perspective as somebody

that's putting in systems and processes,

101

how do you get around or what's

your process for getting around some

102

of the resistance to those things?

103

Especially when you're talking about

a culture that's on the cusp of

104

transitioning from a true startup

environment to a scale up environment,

105

which are two completely different

things, in my opinion, as far as

106

the wiring and the makeup of the

people that work there and maybe

107

even some of the cultural things.

108

Bob St-Jacgues: Yeah, and in

terms of visualizing that.

109

In the beginning, right?

110

So pretty much, when you're a startup

is usually linear growth, right?

111

you've got a trajectory that's really

nice when you get to scale up, it starts

112

to hockey stick to a certain extent.

113

And thankfully, I went on Fiverr and I

got a quick, animation video done of.

114

Basically, when you're growing

linearly, you can build scaffolding

115

underneath and support that linear

growth, and you can run fast and

116

break things when you're scaling up.

117

That's typically exponential growth.

118

And in order to have, I explained

to people it's almost like a

119

half pipe in skateboarding.

120

If you're going to go up like

this, you're going to need

121

strong scaffolding behind it.

122

In other words, you've gotta

make sure that it's sustainable.

123

You can maybe get away

with it for a little while.

124

So that's why I'm the kind of a

little bit of the Debbie Downer coming

125

and going, trust me on this one.

126

It's like the old adage, right?

127

If I had one hour to cut down a

tree, I'd spend the first 45 minutes

128

sharpening the ax, putting in place

the underlying systems and structures.

129

And understanding how communication

really happens is how I get the

130

buy-in initially to say, okay.

131

Build a system so we can run

fast minimizing friction.

132

Jim Kanichirayil: So it

makes sense building in the

133

scaffolding that you described.

134

I'm curious how you get around

the people aspect of it, because

135

there's gonna be people that are not

136

Big on coloring inside the lines.

137

When I hear governance,

that's what I think.

138

here's the sheet of paper that you

can color within, and here are the

139

lines that you need to stay within.

140

So from a change management perspective

or a behavior change perspective,

141

what, are the conversations that you

need to have to shift behavior and

142

thinking from wide open green fields

with free to freedom within fences?

143

Bob St-Jacgues: Yeah, so the way I

tend to describe I use the car analogy

144

and again, in startups it's typically

you've just got the gas pedal.

145

And so basically what I'm saying

is that things are about to get

146

a little bit more complicated.

147

There's gonna be turns in the road.

148

We might have to go off road, so

we're gonna need a steering wheel,

149

gonna need brakes in those areas.

150

And maybe we are gonna need a little

bit of enhancement and automation.

151

We'll get into that in a few

minutes, it's just understanding

152

in order to go real fast.

153

In challenging situations, you do need

a bit of structure and guidelines.

154

And so I use various analogy depending

on the culture in which part of the

155

world I'm in at the time, and, the

type of organization I try to lock

156

it into the industry that they're in.

157

But typically the car

example is a good one.

158

that it's, you're much more in control.

159

If you have, a few things, and I'm

not talking about airbags and things

160

like that, don't go overboard.

161

I'm literally talking steering

wheel brakes and a few others.

162

Jim Kanichirayil: So that sets up the

stage really well for the conversation

163

that we're gonna have because gonna talk

about two different lifecycle stages

164

within organizations when it comes to ai.

165

So let's take maybe the easier one to

tackle, which would be an organization

166

that's more immature in their AI journey.

167

So if an organization and an HR

leader is looking to just start out

168

and implementing an AI strategy.

169

What are the things that they should be

thinking about at that startup stage of

170

the AI journey within their organization?

171

Bob St-Jacgues: Yeah, so I would put,

the answer into three buckets, right?

172

So you've got use cases,

data, and controls.

173

So let's just go for a fairly

simple use case, right?

174

You're solo HR person.

175

You've got a hundred person company.

176

Maybe you're operating in two or three

countries like mine and so you've

177

done some of the hard yards, right?

178

You've got employee guidelines, you've got

manager, playbooks, we've got performance

179

management playbooks and so on, so forth.

180

You have a lot of information.

181

even if you think about something

as simple as benefits, right?

182

I've got Canadian benefits,

US benefits, Mexican benefits,

183

Italian, German benefits, right?

184

And all these are, and then I've got

a hundred employees and we're growing.

185

So one of the easy use cases would be,

and what I call a green win, and I'll

186

go into that in a second, which is,

this is basically one of the easier

187

things you can do in the sense of,

you create an internal HR answer bot.

188

That only uses approved sources, right?

189

So this has gone through the approval

process and it, for organizations it

190

depends, especially with just scale ups.

191

You're probably looking at, SOC

two with Europeans, with GDPR, and

192

some of those tie it all together.

193

so I said use case.

194

Okay, so we've got these, and

then you don't wanna be answering

195

these questions one at a time.

196

That could help you.

197

So what happens, right?

198

So they go in, give an answer.

199

So what are the data that you look at?

200

Median time to the first response.

201

Median time to resolution, right?

202

Because it's one thing for

an AI to spit out a response.

203

It's another thing for the person

to say, awesome, I have the answer.

204

we've got deflection rates,

we've got reopen rates, right?

205

Where it wasn't, the answer wasn't clear.

206

So you get the data.

207

And then lastly, it's the controls, right?

208

So then you look at it weekly or monthly,

depending on how many requests you get.

209

How well is this working

and you iterate it through.

210

that would be one of the very simple, and

what I say, relatively as close as you get

211

to risk-free usage of, AI in a startup.

212

Jim Kanichirayil: So I get that.

213

You and I tend to be having these sorts

of conversations all the time, but I want

214

to go back to something that you mentioned

you wanna identify your green win.

215

if you don't know what you don't know, how

do you know what a green win looks like?

216

So give, share some of the

characteristics or identifiers of

217

what a green win looks like, that

low risk implementation win using ai.

218

What, you mentioned one use case.

219

What are some of those other use cases

that you need to think about, or maybe

220

even, what are some of the other problems

that an HR leader needs to be examining

221

to identify what's the greenest of

the green win within the environment?

222

Bob St-Jacgues: Correct.

223

So this is where you're

looking for two things.

224

It is based on data that you

internally and intrinsically control.

225

I'm talking policies, processes,

procedures, and they can be in Confluence,

226

they could be in Jira, they can be

in your service, but it's documents

227

you've created and you're limiting the

AI to that space and the responses.

228

The second piece is oversight.

229

In other words, it doesn't

spit out answers automatically.

230

It preps an answer and then

you click agree and send, and

231

you're able to edit those.

232

If you have those two elements, I would

say that tends to put it in the green.

233

would also add in a third piece there

where it doesn't make any decisions.

234

In other words, it doesn't say,

Hey, somebody says, Hey, is.

235

Orthodontics covered under our dental

benefits, and it spits out a no answer.

236

Sometimes you want a bit more nuance

than that, and saying, typically no.

237

However, in the case where

there's a underlying medical

238

condition, blah, blah, blah, right?

239

This is where, a lot of AI tends

to miss the nuance and, you make

240

sure that it doesn't have any

final decision making say on that.

241

So again, internal documentation review.

242

you wanna stay away from

it, making final decisions.

243

Jim Kanichirayil: So that's

a really good framework.

244

Just recapping, data and internal

control oversight over the output

245

and then decision control where the

platform isn't making decisions.

246

The way that translates to my head

when I'm thinking about this is human

247

in the lead versus, full automation

is how I consider, this falling in.

248

When you look at those three things that

you just mentioned, what is it about those

249

three things that make it a safe play?

250

If you're early in your AI journey.

251

Bob St-Jacgues: This is where

the external piece comes in.

252

Your listeners may have

heard at this point, right?

253

There are several ongoing lawsuits

against very big players, Workday,

254

Eightfold, HireVue, and so on.

255

if you look at the crux of

those complaints and they

256

dive deep into one case.

257

It's about California consumer laws.

258

So it's very legalistic.

259

However, let's pull back,

and I'm a pragmatic person.

260

I learned early on in my legal

career, you don't say no.

261

You say, where do you want to go and

help the client find a way to get there.

262

So I'm gonna help the listeners

find a way to get there.

263

Yes, it's a big scary world out

there and people are getting sued.

264

However, we've gotta do business.

265

And so by putting those buckets

together, what I call the green area,

266

that those items are not susceptible

to regulation and compliance regimes.

267

And so that's why, with those, you

wanna stay very narrow within those.

268

It's internal, I'm reviewing it,

and it never has a final say.

269

Jim Kanichirayil: I think when I hear all

that, the TLDR that comes to mind is set

270

it up in a way where you're mitigating

your risk of getting sued, especially

271

when you're talking about the workday

example and the eightfold example.

272

now that's a, that's solid stuff.

273

when you think about those three

criteria to set up sort of some

274

basic guardrails as you're setting

those things up, What are additional

275

considerations that they need to factor

in as they're building this model?

276

Bob St-Jacgues: Yeah.

277

And so this is where, you

mentioned it earlier, right?

278

About humans taking the lead or I've

talked about collaborative intelligence,

279

which is imagine having a great, PhD.

280

Chief of staff or

personal assistant, right?

281

They're brilliant, they just don't

know a lot about the real world,

282

And so you need to lead it, you

know your organization better than

283

it does y and so on and so forth.

284

And so that's why it's very important

to have that level of control.

285

And so when you structure it this way,

you can start moving down the line, right?

286

So I gave the example as HR service.

287

Lemme give you another specific example.

288

What about a questionnaire on a

job, in a job application process.

289

And so typically people ask

questions, and then, by the way,

290

this would be a yellow or a red.

291

Don't do this.

292

take people's answers and just ask

AI to evaluate people and rank them.

293

Do not do that.

294

What you can do, remember we're talking

about control and documentation.

295

If you have a rubric that says

you, you've got, there's four

296

answers that you can choose from.

297

This answer is the best one,

And so it can correct for you.

298

But remember, let's go back to basics all.

299

I'm trying to simplify this as much as

possible because you created the question.

300

You created the answers, you

identified the correct answer.

301

All it's doing is grading for you.

302

It's like a Scantron from,

back in the nineties.

303

on exams, that's the level you're going

at and staying in the beginning, right?

304

So if you look at almost every part.

305

Of the employee lifecycle.

306

So whether it's workforce planning

or even, employer branding,

307

recruitment, onboarding, all the

way through, if you focus on those

308

three buckets, make sure it's you.

309

and I wanna highlight that difference,

right between you writing questions, you

310

writing the answer, you identifying the

right answer, and they're telling you

311

if it matches, versus, Hey, I've got.

312

10 people typing in answers or

people answering 10 questions and

313

typing in the answers and just rank

these people against each other.

314

That is a definite yellow, red, and

we can get into those in a second.

315

by focusing on the three areas, you

could stay solidly in the green.

316

Jim Kanichirayil: Got it.

317

Now I want you to think through, the

process of putting, of building this out.

318

So when you're looking at that

early stage of the AI journey and

319

they're putting in a governor's mo

governance model of sorts, where do

320

organizations get themselves into?

321

What are the pitfalls that they need to

watch out for that, that is gonna set

322

them up for fail as they become bigger?

323

Bob St-Jacgues: Yeah, is

where you are asking anything.

324

You're asking AI to make a

decision on compare and contrast.

325

Not against a solid set rubric, right?

326

'Cause you don't need

compare and contrast here.

327

You're scoring people right

against like Scantron, fill

328

in the bubble sheets, right?

329

There's a right and wrong answer.

330

Anytime you are asking for levels

of, analysis and nuance and so on,

331

that's where you get into trouble.

332

Here's the reason, you get a

little bit legalistic, you know

333

that what's the big deal, right?

334

I'm just comparing people.

335

Actually, it's a big deal because.

336

Both in Europe and in North America,

the US and Canada and even Mexico.

337

Anytime you have technology, digital

technology spitting out a decision

338

and here's the law that they're using

to sue a, a Eightfold and workday,

339

it's the Consumer Protection Laws.

340

So typically if you get a credit score and

it's bad and they turn you down and the

341

information is false, you can sue them.

342

That's what these lawsuits are relying on.

343

So that's why I'm saying you want to

back up as far as possible and make

344

sure there's almost two levels there

where you review the work and you

345

don't let it make a decision, right?

346

You have to a bit manually, or you

have to click on some level where

347

the human is making the decision.

348

you have technology spitting

out a decision, think about it.

349

Somebody doesn't get a job.

350

that's a financial loss, right?

351

And this is how these

lawsuits are coming about.

352

So again, don't worry about that

'cause it's a bit complex and dry.

353

consumer protection laws, you

can imagine in the US there's

354

thousands of pages of it.

355

again, focusing on your own data from

yourself and making sure that you're

356

reviewing, making sure that you're making

the final decision and can defend it.

357

Jim Kanichirayil: So I'm gonna throw out

a hypothetical, and I want to go back to

358

a version of what you described earlier.

359

So let's say that you're going through,

the hiring process and part of the

360

hiring process includes, writing

sample and you're answering a question.

361

Everybody's answering the same

question to a writing sample.

362

Bob St-Jacgues: Correct.

363

Jim Kanichirayil: You've built a style

guide internally that says, this is

364

what the ideal answer looks like.

365

This is what a good

answer looks like this is.

366

Okay, so you have all of these inbound

candidates that are applying, going

367

through the application process,

submitting a writing sample, and then you

368

take those writing samples and you apply

it against the style guide and have the

369

AI do an analysis and stack rank the top

four candidates against the style guide.

370

Am I thinking about having that as a

green process the right way, and if not,

371

what are the holes in how I'm thinking

about it so that it can be a tight

372

process that doesn't set a candidate,

that doesn't exclude a candidate that

373

should be included in the process.

374

I.

375

Bob St-Jacgues: Yeah, this, these

use cases, are where the rubber

376

meets the road, so to speak.

377

So in your case, you're doing well.

378

So you've got the style guide.

379

It's yours.

380

You've anchored it, you've already

decided what excellent looks like,

381

good, looks like fair, looks like

poor, looks like say it's, four levels.

382

You can put that into ai.

383

I know as a university professor, right?

384

we do it, it, the first

screening is done through ai.

385

However, that's where it needs to stop.

386

the part where you said stack rank them.

387

No.

388

Okay, you can do that.

389

It will give you a score, right?

390

So you score it.

391

So you know section one,

they need to mention 1, 2, 3.

392

If they mention two outta

three, will they get 66%?

393

Section two, section three is

section four, and so on all the

394

way through, you will get a score.

395

You personally, individual,

as a human being.

396

look at that, right?

397

You can sort it in Excel, right?

398

If you could export it to Excel, okay,

this person scored out of 40 a 36.

399

just sort it out.

400

36 is first, 32 is second, and so on.

401

You cannot ask AI to compare and contrast.

402

Okay?

403

You want it one-to-one with ai, right?

404

This AI evaluated this person.

405

Forget about everybody else.

406

Our AI that we use evaluated this person.

407

everybody else, right?

408

And then you get a score, and

then you look at the score and

409

say, okay, I'm reading this.

410

I validate it.

411

Yes, I agree on the score.

412

then you, so you're very good

up until the last piece where

413

you asked at the stack rank.

414

'cause then you're asking it

to make a subjective decisions,

415

unless you say just rank them.

416

But it, that would be the same

as, doing a sort in Excel.

417

Jim Kanichirayil: Got it.

418

Okay.

419

That's helpful.

420

now when you think about the

governance decisions and building out

421

those frameworks, what are the key.

422

Or what are the critical components

that every HR leader should be

423

thinking about when they're early

in that start in the AI journey?

424

Most important things that

they need to keep in mind as

425

they're building this model.

426

Bob St-Jacgues: Yeah.

427

So the first piece is, I would

say company confidentiality.

428

So right now what we have is, what we

call in the industry, AI drift, which

429

means I've got Claude at home, I've

got ChatGPT, I use Grok, I use this.

430

They go to work and then start dumping

documents, from their workplace.

431

yes, I'm sure it'll make you

analyze things more quickly.

432

However, again, depending on your

settings, and this is where people are

433

not always aware of the settings and

how it's important in terms of privacy,

434

Some of the information that gets out.

435

Yeah.

436

Also, you've gotta be aware of

where you are, in the world.

437

There are, for example, in New

York City, there are specific

438

New York City AI use laws.

439

Colorado has a very

specific AI use laws, right?

440

So we've got AI drift, we've got

where are we located, and then the

441

next piece is looking at what is

the maturity level of individuals.

442

I'll give you an example of the last one.

443

I don't know if you remember.

444

It was about three, four years ago,

and there was a huge, press release,

445

and IBM said, Watson has figured

out when somebody's about to quit.

446

This is where the

maturity comes in, right?

447

You could, everybody in HR just gasped.

448

and by the way, that one directly,

it didn't go to hr, it went

449

directly to the person's manager.

450

Not everybody's going to be able

to handle that with the nuance

451

and finesse that it requires.

452

So it's understanding your

organization, the maturity level, right?

453

If you're working in, for example,

consulting, engineering, these people are

454

masters in doctorates, A little bit more

mindset, a little bit more structure.

455

you are dealing with more of a

retail startup environment, where

456

things are, young people, it's

fun, it's chaotic, or fast moving

457

consumer goods, red Bull, right?

458

and their startup phase.

459

You heard a lot of great stories

out of Austria back in the day.

460

those folks, I would say.

461

being ready to really clamp down.

462

So look at those three, first pieces

especially in terms of where you are,

463

because that could really constrict what

type of AI you use and for what reasons.

464

Thomas Kunjappu: This has been

a fantastic conversation so far.

465

If you haven't already done so,

make sure to join our community.

466

We are building a network of the

most forward-thinking, HR and

467

people, operational professionals

who are defining the future.

468

I will personally be sharing

news and ideas around how we

469

can all thrive in the age of AI.

470

You can find it at go cleary.com/cleary

471

community.

472

Now back to the show.

473

Jim Kanichirayil: I'm gonna resist the

urge being a retention and turnover guy to

474

go down in the turnover intention rabbit

hole that you just opened the door for.

475

So I'm gonna, not take the bait, but that

would be an interesting conversation.

476

So now I want you to fast forward.

477

Bob St-Jacgues: Yeah.

478

Jim Kanichirayil: So right now

we've been talking about building

479

governance models for organizations

that are early in their AI journey.

480

Now, let's think about something,

let's think about those organizations

481

that are maturing, as far as their AI

initiatives and implementations are going.

482

When you think about that maturing

organization, what needs to shift

483

from that early stage model to

be relevant in that accelerating

484

growth or maturing model?

485

Bob St-Jacgues: I guess you would think

about it almost like an employee, right?

486

So in the past when we used to develop

employees, we used to call up, turn

487

people into T-shaped employees, right?

488

So the vertical part of the T is your

depth of knowledge and your craft.

489

And the top of the t the horizontal

part is the breadth, right?

490

And how it goes in.

491

Think about it that way.

492

You're creating a more T-shaped

AI governance environment.

493

Which means it goes much deeper in the

sense that what are the requirements, for

494

example, with utilities, in Western Canada

and western United States, you have.

495

WEC, Western Electric Cooperative, which

has very strong opinions in terms of

496

cybersecurity in some of those areas.

497

you know how it relates to ai.

498

So it's not only where you are, whether

you're partly owned by the government.

499

And you also have these

overarching, sometimes multinational

500

regulatory frameworks to come in.

501

So this is what I mean by depth, right?

502

You start really going deep, into your

organization because you have many

503

more stakeholders and touch points.

504

In terms of the top level of

the T, this is where you get

505

into some of the more nuanced.

506

Examples and where data

collection optimization is

507

much more important, right?

508

'Cause at a small company it's

like, yeah, my bad, sorry.

509

It's gonna take a day

longer to get you an answer.

510

When you're talking about 5,000

employees in 10 countries.

511

And answers need to come

at people more quickly.

512

You need a much more robust, we

call like an HR knowledge base

513

triage and casework summaries.

514

So then you're building service levels

with review after review and so on.

515

So it's a bit more bureaucratic, right?

516

Because there needs to be more

redundancies and fail safes in these

517

areas because giving one employee

the wrong answer, it could cascade

518

right and turn into a precedent.

519

Oh, Bill got $500, why

am I not getting 500?

520

Well, that was an error.

521

Meanwhile, you've got 200

employees with their hand up

522

asking for that $500 as well.

523

So you see it's the scale of that.

524

So I'm trying to simplify it

and be pragmatic, for HR folks.

525

And yeah, just think about it, of just

dramatically deepening your vertical T and

526

dramatically widening the top of your T.

527

Jim Kanichirayil: So I wanna apply

what you said specifically when you're

528

talking about redundancies and fail safes

and apply a specific use case to it.

529

So in a more mature organization,

a lot more people, there's a lower

530

fault tolerance when it comes to

errors because one error can be

531

duplicated across an organization and

really incur some pretty big costs.

532

Let's apply it to the use case of a chat

bot because it seems like a lot of HR

533

organizations look at that as low hanging

fruit to implement to clear off some

534

of the load on their individual plates.

535

And you have a chat bot in place, you

have employees asking it questions.

536

The chat bot might be producing answers

or creating tickets for HR to deal with.

537

And when it's doing that, we all

know that no matter how well your

538

prompt is designed, your AI has a

tendency to go off the rails here

539

and there and just invent things.

540

So how do you structure

redundancies and fail safes?

541

In something that seems fairly

low level, but the implication

542

of getting something wrong can be

massive in a larger organization.

543

So how do you, build a

fail safe around that?

544

Bob St-Jacgues: Yeah, a great question.

545

And this is where honestly,

I spend a lot of my time.

546

So it comes down to training

the AI that you use, right?

547

So if you're in a Microsoft environment,

use Copilot, Google environment, Gemini.

548

Very specifically to that, for example,

and when I'm doing training, people

549

always say, hey, I just typed in

the same prompt you did and I got a

550

completely different answer than you.

551

I was like, yeah, I've been training

552

I got the 25 year in review and

I'm a top 3% ChatGPT user, right?

553

So I've had a lot of conversations

over 2,900 conversations, and so

554

I've tweaked and dialed it down.

555

Now, the biggest one for me

is I am a fellow in the center

556

for evidence-based management.

557

And as part of my commitment to that

organization is, can't say what I think.

558

I can only say what I can prove.

559

And that means there

needs to be a citation.

560

Let's go back to the chat bot example.

561

What you will have to do is say,

you cannot just pop up a sentence.

562

It needs to be quoted verbatim

from the test text I uploaded.

563

And there needs to be a link

to give to the employee.

564

So the employee clicks on it and

expands the actual document itself.

565

They read it and they're responsible

for, their own interpretation.

566

So I don't allow my AI's to freelance

guess moonlight, hallucinate,

567

whatever you want to call it.

568

I force it to be like a professor, right?

569

Where is your source?

570

Where is the citation?

571

You cannot give any answers unless

there's a citation to the document

572

I gave you in this instance only.

573

Jim Kanichirayil: So I'm grinning

because, I think all of us have had the

574

experience when working with various,

LLMs, where they'll produce a source, but

575

the source is just completely made up.

576

And the added layer of forcing the master

prompt to produce, a link that goes to

577

the source that's being cited, I think

that's a really good, simple way of

578

like pressure testing or validating the

actual evidence that's being provided.

579

At the same time, I was thinking through

that I had a flashback to my my doctoral

580

program where the instructor was like,

in God, we trust everyone else must

581

produce evidence with multiple citations.

582

Nobody cares what you think.

583

Only care about what you can

prove or cite back that to

584

research that's already been done.

585

So thanks for that flashback.

586

Continuing on this thread, we've

talked about building in redundancies

587

and fail safes, which becomes

critical in larger organizations.

588

And as you mature your AI practice.

589

What are the things that you

should be retiring as you

590

mature as an AI organization?

591

There are a certain set of rules

that apply when you're in that

592

startup stage or in that early stage.

593

What should be the things that you get

rid of as you become more mature in

594

the AI practice within your enterprise?

595

Bob St-Jacgues: The way I've

worked it, I've tried to bottom

596

up and I've tried it top down.

597

And then let me be a bit more specific.

598

So in terms of bottom up, you tend

to open up information, right?

599

and it's very similar to

normal policies and procedures.

600

back in the day we used to

hand people a handbook, right?

601

With all the nice pretty pictures and they

had to sign the back page and rip it out

602

saying they've read it and understood it.

603

it's very similar to new processes

using Once people develop a certain

604

level of maturity, you've then.

605

We moved it back, right?

606

So that everybody onboarding

gets that level of training.

607

Okay?

608

So you've taken care of everybody's

here, you locked it in, several cycles,

609

and you'll be able to go through

and analyze some of the information.

610

Okay?

611

How many errors are going?

612

That happens, you start

de layering it right?

613

In that area and giving people

direct access to information.

614

as you go through.

615

I've tried it top down, so as

I've trained managers in that area

616

and allowing, empowering them.

617

they tend to use it right as a tool.

618

So as you grow your managers, as

you train them, as you develop them,

619

because you thinking back to the

startup, you're getting some in there.

620

Yeah, they may have had experience

somewhere else, but once you get 'em

621

up to a certain level, start taking

off the training wheels and saying,

622

actually employees are gonna come to

you and ask you some hard questions.

623

You leader will still have access to.

624

The information and then it will help you

hear our standard answers coming through.

625

This is the way you

evolve it for the leaders.

626

So instead of everything open to

everyone and all these employees,

627

you start empowering manager, which

leads to sustainability, right?

628

Because if you're the crutch forever,

oh yeah, employee, don't bother

629

me with those kind of questions.

630

Go to, the chat bot versus hey,

have a question about your career.

631

Cool.

632

Don't go to the HR site, talk

to me about it and help it out.

633

So you can see what you're trying

to do is as your leadership matures,

634

as your employee base matures with

that, you start peeling away some

635

of the layers and then moving up.

636

Another good example is

learning and development.

637

In the beginning.

638

it will do is you set some algorithms,

it'll say, Hey, you are an iOS developer,

639

and Bill's also an iOS developer, so

and he took this course, so you should

640

take this course too, five other people

in your department co took this course.

641

You should too.

642

That's level one.

643

Once you start getting that information

and then you can start to automate

644

it, then you move up the AI food

chain, so to speak, where it suggests.

645

certain courses based on the

set of competencies that are

646

required for the next level.

647

Hey, it'd be a good time right for

you to spend the next year learning

648

these four competencies, which will

set you up well for this ne next level.

649

So you can see how you just keep

making things more, challenging

650

and mature with the organization.

651

Jim Kanichirayil: I like the example that

you just gave from an employee development

652

perspective where you have plaque.

653

Platforms that are integrated into,

especially if you're mature as an

654

organization and you have some sort of

criteria that's established in terms

655

of, for you to be a senior developer,

you need to be able to do these,

656

measurable things and these are the

soft skills that you need to develop.

657

And then you lump that in with your

learning management platform that has a

658

library that helps them shore that up.

659

I really like that as far as a.

660

Taking some of the guesswork, out

of the employee development plan.

661

And, I think one of the things

that a lot of organizations do

662

wrong is that they say, you're in

charge of your own development.

663

if you don't know how to go from

point A to point B or it's murky.

664

How are you supposed to take

charge of something that's

665

not clear in the first place?

666

So having bit of an assist where

it suggests learning paths for that

667

individual, that actually helps

the individual be more intentional.

668

And if it's fed into sort of a

one-on-one process with the manager

669

that actually gets the manager.

670

More actively developing their employees

instead of leaving it up to the

671

employees to actually develop themselves.

672

I think that's one of the biggest

crutches that exist in corporate

673

America is that you leave development

to the employee when oftentimes

674

you're not arming the employee with.

675

What does the career path even look like?

676

What's the hard criteria that I need

to attain to get to the next level?

677

so I really like, that you

brought, Looks like you have

678

something else to add to this.

679

So I'll be quiet for a second.

680

I.

681

Bob St-Jacgues: Yeah, and, you

touched on it a little bit, right?

682

Because again, there's that pressure,

hey, you've gotta develop yourself.

683

or even.

684

sometimes it's like,

Hey, you're the manager.

685

I'm holding you responsible.

686

20% of your bonus is based on, herding

cats and learning and development.

687

Both of those are patently unfair.

688

What AI can do is start

bridging that gap, right?

689

And nudging people along.

690

And so the employee, it's oh, okay.

691

I'm good at these four competencies,

these ones, oh, I've never tried before.

692

So lemme take courses and I'm click.

693

Whatever they are, live but then after

they do that, then the manager gets

694

a message, Hey, did you know that

Jim signed up for these four courses?

695

He's gonna do one per

quarter over the next year.

696

gives you something to talk about.

697

That's the end.

698

How's the course scoring?

699

can I help you with anything?

700

You wanna try it out?

701

I've got a stretch project for you

that will allow you to try skill

702

competency one and two on that side.

703

So you see how the lot of these AI

pieces are trying to push both ends

704

against the middle, but it's almost,

saying is don't get too clever.

705

Yeah.

706

Agentic is amazing and we can

touch on that later, but right now.

707

Putting in the basics right to

what you're trying to do is, use

708

AI to nudge human behavior, right?

709

It's, again, let's go back to

collaborative intelligence.

710

you're using it as a, I think

with notion you've got second

711

brains and things like that.

712

Use it across the whole, Hey,

you've got candidates to interview.

713

Hey, two of your employees

are taking courses.

714

Hint, they're thinking about growth.

715

Great.

716

That doesn't mean they've got

a foot out the door, right?

717

Typically.

718

and then you go through, hey,

they're looking at the pay scales.

719

Okay?

720

So they, they're growth minded.

721

They wanna see how they

can make more money.

722

So it's about, again, pushing both ends

against the middle and helping nudge.

723

Human, or try to minimize

human shyness, right?

724

People are like, oh, I

don't wanna talk to my boss.

725

I don't want be seen like

the problem employee.

726

what if it's AI going, Hey,

manager, employee over here

727

is looking at these things.

728

You may want to pay attention.

729

Jim Kanichirayil: Yeah, those

nudges and prompts are actually

730

a really good call out.

731

And I like the other piece that you

mentioned and I think, and I wanna

732

highlight it, which is bridging the

gap between theory and execution.

733

So if we're talking about that

learning management use case where.

734

It's informing the ecosystem

about, Hey, your employees here,

735

and here are taking these courses.

736

That as a manager and especially at

the line level manager, space, where

737

you're, you probably got a battlefield

promotion, you have no coaching or

738

training on how to actually be a manager.

739

Those sort of prompts that say.

740

Hey, you have a certain percentage

of your employee population

741

that are taking these courses.

742

What assignments or initiatives do

you have that you can delegate, which

743

would be a stretch goal for them

help them move what they're learning

744

into actual practice and give you an

opportunity to coach in real time and

745

build your skills that way as well.

746

I think that's a really important

call out that gets missed, especially

747

at that line level manager, tier.

748

In a lot of organizations.

749

So really good stuff here.

750

I want you to put your, put

your Debbie Downer hat back on.

751

Bob St-Jacgues: Yeah.

752

Jim Kanichirayil: So when you're looking

at organizations that are at the more

753

mature stage of their AI journey, and

they're looking at getting the ne getting

754

to the next level from an AI initiative

perspective, where do things go sideways?

755

What are the things that always

get screwed up or more of, most

756

often get screwed up when they're

looking to make that leap?

757

Bob St-Jacgues: It.

758

It's in the area of ai,

autonomous decision making.

759

Here's what happens.

760

You've got 2000 trying to

do performance management.

761

You've got either quarterly

reviews, annual reviews, right?

762

You're gathering all this information,

so you've asked employees to

763

write their self-evaluation.

764

You've asked them to get one or two peers.

765

You've asked their manager, so you've got

766

Come on.

767

We're only human, right?

768

So then all of a sudden, you

get into that gray area, oh,

769

I'll just compare them, right?

770

And it's that slippery slope.

771

Where it comes down, right?

772

So this is where, folks come

in, they're overwhelmed.

773

I have spent many, two, three ams in

the morning with my HR team globally

774

trying to finish performance reviews and

do it the hard way in Excel and compare

775

and contrast review each managers,

making sure there's equity within the

776

department, within the country, within

the group, and so on, so forth, right?

777

And it's the human.

778

sometimes a lot of HR folks and managers

are overwhelmed and they just click the.

779

Easy way.

780

Hey, ChatGPT, hey Claude, compare

and contrast my five employees.

781

And that's where gets

'em into trouble, right?

782

So it's then taking that nudge and

pushing that nudge too far, right?

783

And saying, oh, becoming codependent

on the AI piece and having AI make a

784

decision I, my budget for increases

is $112,000 for four employees.

785

How should I, distribute it?

786

Based on these four reviews.

787

Okay.

788

That you're in the red.

789

Okay.

790

'cause you're asking AI to

make paid decisions based on

791

subjective, review of four reviews.

792

Jim Kanichirayil: So

when I hear that example.

793

It almost strikes me as a root cause of

that sort of scenario is poor planning.

794

because if you're leaving everything

to the last minute, that's when

795

you're forced into those corners.

796

Now, some of that stuff, because we

know HR can't be helped because there's

797

always a fire to be fought somewhere.

798

So there's that aspect of it too.

799

I would imagine a better comms plan.

800

More accountability through the manager

tiers to get these things in done on

801

time or delivered on time would probably

mitigate I might be in an ideal world

802

scenario in this, but I think that's

how you get yourself in trouble, is

803

leaving everything in the last minute.

804

Bob St-Jacgues: Yeah, counterintuitively,

it's actually giving more time is worse.

805

Parkinson's law, right?

806

So people ascribe a level of

complexity to a request based on the

807

time you're given to do it, right?

808

for an employee we're doing quarterly,

it should take me write a review, right?

809

Especially if you're doing it

quarterly, you have one-on-ones,

810

whatever on that side.

811

and if you're giving two weeks to do it.

812

it's due.

813

And then HR is working all

over the weekend trying to you,

814

collate everything together.

815

And that's again, where mistakes get made.

816

Or the manager, oh, I've

reviews on the Friday.

817

So typically what I do is say, no,

you've got three days to do it.

818

Monday to Wednesday, and it could slip

into that first Friday, you play catch

819

up and then I've got the next week.

820

So exactly to your point.

821

But the one piece I wanted to

add in was, Parkinson's Law.

822

Giving more time is

actually counterproductive.

823

Jim Kanichirayil: Great

conversation so far.

824

And I think we learned a lot,

especially when we're covering sort

825

of two different business maturity

stages when it comes to this.

826

When you think about both

sets of scenarios that we

827

described, what are the biggest.

828

Items that people need to pay

attention to when they're looking

829

at rolling out any sort of AI

initiative, independent of stage of

830

business development that they're in.

831

Bob St-Jacgues: Yep.

832

So I would do a five item checklist,

which we've touched on before.

833

So let's go number one, AI use inventory.

834

So every tool, every use case, and just

make sure there's clarity on those, right?

835

So it's Hey, ChatGPT for everyone.

836

No, absolutely not.

837

Copilot for everyone.

838

No, absolutely not.

839

There needs to be guidance.

840

Number two, data boundaries.

841

What can or cannot be put into that model.

842

Okay.

843

tiers.

844

So the green, yellow, red, green.

845

Okay.

846

Pull HR information from our own playbook.

847

Awesome.

848

yellow, getting into maybe some evaluative

pieces in the recruitment process.

849

Yeah, it's not great red

making decisions on pay.

850

Okay, so let's understand the risk tiers.

851

The next one is, you can't

ignore this human accountability.

852

Who is the owner of that system

And use case and process.

853

And then the last piece, and I touched

on this in the beginning 'cause I'm a

854

big data nerd, is measurement, right?

855

So how long does it take to get an answer?

856

How often are the answers

correct or incorrect?

857

How you know?

858

And so on and so forth.

859

So that you're constantly looking at.

860

So use inventory, boundaries, risk

tiers, the human accountability

861

and basic measurement.

862

Jim Kanichirayil: Great stuff.

863

I know that we're scratching the

surface on both these fronts, in terms

864

of building governance models for

organizations that are early in their AI

865

journey and building governance models

for organizations that are more mature.

866

and I'm sure people wanna reach

out to you, what's the best way

867

for them to get in touch with you?

868

Bob St-Jacgues: best ways to

link up with me on LinkedIn.

869

In terms of direct messages

as well, I love nerding out.

870

As you could probably tell.

871

I enjoy talking about this subject and

having chats with like-minded people.

872

'Cause I'm on my own learning journey.

873

Just this year, I started a doctorate

in AI and LLM because, it became clear

874

what I didn't know was quite massive.

875

And so in order to be more helpful right,

to the organizations I work for my peer

876

community to folks, listening to podcasts

like this, it was, I took it as a duty.

877

to be better informed and

to be more evidence-based.

878

And rather than I think now I

could prove, some of these ideas.

879

And so on that side, happy to

continue the discussion and to learn.

880

Jim Kanichirayil: Awesome stuff.

881

Speaker: Thanks, Bob,

for hanging out with us.

882

really good conversation, and

I think a lot of our listeners

883

are going to take away a lot of

different things from the discussion.

884

When I think about this conversation,

I think the most foundational thing

885

that everyone needs to keep in mind

is defining what green looks like.

886

And when we say green, it means a

high impact, low risk, element of

887

implementing AI within your environment.

888

And I think that's pretty important

because when you think about any

889

initiative, you need to understand

what success looks like, and you

890

also need to understand what the

path to success looks like as well.

891

And when you define a process as safe

to implement, that gets you-- that, that

892

puts you in a position to be successful.

893

And I think defining that as a first

step, whether you're a small organization

894

or a large organization, is an

important step for many organizations

895

as they move in their AI journey.

896

So when you're looking at where to get

started and how to move things forward,

897

defining what safe, predictable, and

successful looks like is an important

898

first step that every organization

should consider before they get started

899

in launching any major initiative.

900

So I appreciate you sharing that with us.

901

For those of you who've been

listening to the conversation,

902

we appreciate you hanging out.

903

If you liked the discussion, make sure

you subscribe and follow the show, as

904

well as leave us a five-star review on

your favorite podcast player, and then

905

tune in next time where we'll have another

leader hanging out with us and sharing

906

with us the insights that they picked

up as they work on future-proofing HR.