\n From Days to Seconds\n Riyadh · 25 slides\n

\n \n \n

\n\n

Rehearsal script · Alteryx on Tour KSA 2026

From Days to Seconds

Turning retail analytics into store-level action. Every line you say, every beat you pause on, slide by slide.

25 slides

15:00 target slot

17:20 full script

3 key slides

\n\n

Fix these before you present

Your demo screenshots and your CSV are different datasets. Slides 12–20 show Store 3239, AED, 2023–2025 from Retail_Demo_Data with 50 stores. Your glasses CSV is a single London optician, GBP, 2022–2026 — and it grows instead of falling. Present the screenshots as they are, or rebuild the demo on Store 3239 data. Do not demo live on the glasses file behind these slides.
Currency is inconsistent. Slide 17 says AED. Slide 20 says $3.03M and $734.7k. Fix one — or just say “net sales fell three million” and never name the currency.
Slide 6 says “4 of 5 prompts passing.” Don’t skate past it. Scripted below as a credibility line — telling a room you test your agent beats implying it is perfect.

\n\n

Open

\n\n

0:00

0:30

Title

Say\n

Good morning. My name is Guneet Kaur, and I’m a Data and Analytics Consultant at Billigence.

The title of this talk is From Days to Seconds. Let me tell you what that means.

If your sales dropped last month, your dashboard tells you that immediately. But if you ask why they dropped — in most retail businesses, that answer takes about a week.

Today I’ll show you how that week becomes seconds.

Land

Seeing the drop is instant. Explaining it is slow.

Delivery

Look up. Do not read. This is the only slide you should have fully memorised.

\n\n

0:30

✂ trim to 0:15

Billigence & about you

Say\n

Very quickly — Billigence is a global data consultancy. We help organisations work better with data: governance, AI readiness, advanced analytics. We have eleven offices, including here in Riyadh.

My own background is retail and customer analytics — segmentation, loyalty, pricing, and trend analysis. So the problem I’m about to describe is one I’ve lived.

Land

I’ve done this job. I’m not just selling software.

Delivery

Move fast. Nobody came for your CV. Thirty seconds maximum.

\n\n

1:00

The Situation: Too Much Data, Too Many Questions

Say\n

Here’s the situation in almost every retail business.

On the left, you have data. Lots of it. Sales and transactions. Stock. Customer behaviour. Pricing history. Product mix. None of this is missing. Companies are rich in data.

On the right, you have people who need to make decisions today.

The District Manager asks: why is store 3239 down this week, and what do I do about it?

The Category Manager asks: which brands need a pricing change right now?

The Commercial Lead asks: if I change price by 5%, what happens to margin and volume?

Three different people. Three different questions. The same underlying data.

And every one of them waits.

Land

The data isn’t the problem. Getting the right answer to the right person is.

Delivery

Read the three questions as those people would say them — slightly frustrated. Make them real.

\n\n

2:00

0:45

Why dashboards don’t close the gap

Say\n

So why doesn’t the dashboard solve this?

Traditional BI dashboards show aggregates. Totals. Trends. And they do that well.

But they don’t break the number down by driver. They can’t answer a question that’s specific to one person’s job. And they will never recommend a commercial action.

So every new question becomes a new request to the analytics team. And more days of waiting.

Land

A dashboard shows you the number. It doesn’t explain it, and it never tells you what to do.

\n\n

2:45

1:00

What that delay actually costs

Say\n

Let me put numbers on that cost.

Five to seven days — that’s the typical turnaround for a store-level diagnosis.

Around seventy percent of that time is not analysis at all. It’s data preparation. Cleaning, joining, reshaping. Skilled analysts, doing plumbing.

And at the end of it — you get diagnosis only. BI stops at retrieval. You still have to decide what to do.

Now here’s what makes this expensive. This isn’t one hard question asked once. It’s the same question, asked every week, about a different store. Multiply that across a portfolio, and you’re spending a very large amount of skilled time on a question that has a repeatable method.

Land

This is the same question on repeat. Anything repeatable can be automated.

Delivery

Say “five to seven days” and “seventy percent” slowly. Pause after each.

\n\n

3:45

1:30

Our Approach — drivers into actionable insight

Say\n

So here’s what we built.

The scenario: Store 3239 is down year on year. The District Manager wants to know the cause, and wants a corrective action.

The data is real retail data — one store, two fiscal years, covering transactions, products, customers and pricing history.

We open with a single brief: “Diagnose this store’s year-on-year performance the way you would brief a District Manager. Show the working, and recommend a next action.”

Then it runs five steps, in order, every time.

One — the headline. What is the net sales change?
\n Two — the revenue bridge. Was it transactions, basket size, price, or product mix? In actual currency, not percentages.
\n Three — brand winners and losers. Top five brands, flagging where revenue is up but units are down — because that’s a price story, not a demand story.
\n Four — customer cohorts. Split into retained, lapsed, and new.
\n Five — synthesis and action. Three ranked hypotheses, each with a confidence level, the data that would confirm it, and one commercial action.

Say — the “4 of 5” line\n

One more thing on this slide, and I want to be straight with you. Our standard for calling this credible is four out of five prompts passing. We test it. We don’t assume it works. I’d rather tell you our bar than pretend there isn’t one.

Land

Five fixed steps. And we measure whether it works.

Delivery

Say the “four out of five” line confidently, not apologetically. Honesty about testing is what separates you from vendors claiming perfection.

\n\n

Architecture — four layers

\n\n

5:15

0:30

✂ trim to 0:10

Layer 1 — Data Sources

Say\n

Four layers. First, the sources. Transactions — sales, units, customers. Pricing history — list price, promotions, markdowns. Product and stock — hierarchy, on-hand, brand.

Nothing unusual here. This is data you already have.

\n\n

5:45

0:35

Layer 2 — Data Curation

Say\n

Second layer, and this is where the real work happens.

Alteryx Designer cleanses and joins the data, calculates elasticity, and — importantly — gives every field a human-readable name. The tables are structured specifically so an agent can consume them.

Alteryx One Auto Insights then acts as the agent’s data-access layer, and holds the semantic context across queries.

And underneath — governance. PII handling, lineage, and role-based access, applied here at the curation layer, before data ever reaches the agent. That ordering matters. Governance is not something we bolt on afterwards.

Land

Governance happens before the agent, not after. The agent never touches ungoverned data.

Never cut

If you’re running long, compress slides 7 and 10 — never this one.

\n\n

6:20

0:35

Layer 3 — Agent Studio

Say\n

Third layer — Alteryx One Agent Studio. Four parts.

Custom Instructions — its role, what data to expect, the analytical framework, the output style, and the guard rails.

The Reasoning Loop — it plans, calls tools, checks its own results, and keeps them consistent across a long conversation.

Tool Use — it queries Auto Insights, runs the decomposition, and produces the visuals.

Guard Rails — scope limits, confidence flagging, and escalation when the data isn’t good enough.

\n\n

6:55

0:20

✂ trim to 0:10

Layer 4 — Outputs & Users

Say\n

And the fourth layer — what people actually get. Charts generated from the agent’s own working. And actionable insights: ranked hypotheses, with confidence levels, and a recommended action.

\n\n

Demo

\n\n

7:15

0:15

“Output Demo”

Say\n

Let me show you what that actually looks like.

Delivery

Take a breath. Change your energy — you’re switching from explaining to showing.

\n\n

7:30

0:30

Agent Studio home

Say\n

This is Agent Studio inside Alteryx One. You can see the platform down the left — Designer, Live Query, Auto Insights, Orchestrator, and Agent Studio.

Agents live here as assets — named, owned, shared. Not scripts on someone’s laptop. This workspace has five.

The one we built reads transaction, pricing and stock data, and produces analysis and visuals from prompts.

Land

An agent is a managed, owned, shareable asset.

\n\n

8:00

0:30

Agent setup — the dataset

Say\n

Setting one up is straightforward. It has a name, a description, and — the part in the red box — the datasets it is allowed to use.

This agent can see one dataset: Store 3239 sales data. That’s it.

This is your first control. The agent cannot wander into data you haven’t approved, because it simply cannot see it.

Land

You choose exactly what it can look at.

\n\n

8:30

1:15

Custom instructions & guard rails

Say\n

This is the slide I most want you to remember.

At the top — output style. We tell it: lead with the headline number, then the driver, then the so what. Quantify everything — no adjectives without a number behind them. Call out data quality problems explicitly. Give ranked answers. And surface findings that are counter-intuitive, rather than just confirming what we already believed.

Then — the red box — guard rails.

If a comparison spans an incomplete period, flag it.
\n Returns are netted out by default, and flagged separately when they’re material.
\n Do not extrapolate from fewer than thirty transactions.

Think about that last one. We are explicitly telling this agent: below thirty transactions, you are not allowed to draw a conclusion. Because below that, it’s noise, not a pattern.

A general AI tool would give you an answer anyway. Confidently. This one is built to say: I don’t have enough evidence.

That is the difference between a tool that sounds clever and a tool you can run a business on.

Land

An AI that admits what it doesn’t know is worth more than one that always has an answer.

Delivery

Slow right down. Pause before the last line. This is the slide that wins the room — give it space.

\n\n

9:45

0:20

The chat interface

Say\n

And this is what the business user sees. A box. They type a question in plain English. No SQL, no report request, no waiting.

We’ve given them three starter prompts so they know where to begin.

\n\n

10:05

0:30

“Called tools — 7 of 7 completed”

Say\n

Now watch what happens when someone asks the first question.

It doesn’t just answer. It shows its working. Seven tool calls, all completed. It listed the datasets. Listed the measures. Listed the segments. Resolved the date period. Then calculated the change, three times.

You can open each one of these and see exactly what it did.

This is the auditable trail. When someone asks “where did this number come from?” — you can show them.

Land

It shows its working. You can audit every step.

\n\n

10:35

1:00

The headline answer

Say\n

And here’s the answer.

2023 to 2024 — net sales were essentially flat. Down less than one percent.

Then 2025. Net sales fell to twelve million, a drop of three million — that’s minus twenty percent.

But look at the second observation. The agent tells you what drove it: a twenty-two and a half percent reduction in unit volume — not price erosion. So this is a demand problem, not a pricing problem. Those need completely different responses.

And now look at the red box. Unprompted, the agent adds: 2026 data is currently partial — available through April — and has been excluded from this comparison to ensure accuracy.

Nobody asked it to check that. That’s the guard rail from the previous slide doing its job. Without it, this comparison would have looked catastrophic — and it would have been wrong.

Land

It caught the partial-period trap on its own. That’s the guard rail earning its money.

Delivery

The red box is your best moment in the whole demo. Point at it. Pause. Then deliver “Nobody asked it to check that.”

\n\n

11:35

0:30

✂ trim to 0:15

Traced back to Auto Insights

Say\n

And every number links back. See the source line at the bottom — that’s a live link into Alteryx Auto Insights, showing exactly where the figure came from.

The agent is not producing numbers from somewhere you can’t inspect. It reads from your governed data layer, and it tells you where it looked.

Land

Every claim traces back to governed data.

\n\n

12:05

0:30

✂ trim to 0:15

Inside Auto Insights

Say\n

This is that link opened up. The same finding — net amount for store 3239 down three million, minus twenty percent from 2024.

And from here a user can break it down themselves — by brand, category, city. Notice this dataset covers fifty stores; we filtered to one.

So the agent and the analyst are working from exactly the same numbers. There is no second version of the truth.

Land

Agent and human read from the same governed dataset.

\n\n

12:35

1:00

Brand winners and losers

Say\n

Last part of the demo, and this is where it stops being a report and starts being analysis.

The question was: top five brands by year-on-year change — and flag where revenue is up but units are down, because that’s a price story, not a demand story.

Twenty-three tool calls this time. It sets out the position: net sales down twenty percent, units down twenty-two and a half percent — that’s five thousand seven hundred fewer units. But despite that, several brands grew.

And then it separates the two stories.

Ray-Ban Meta brought in seven hundred and thirty-four thousand in new revenue — and the agent notes this brand had zero sales the previous year. So this is pure new demand, from a product launch, offsetting part of the decline in the older brands.

Chanel grew revenue nearly four percent — and units grew almost exactly the same amount. So that’s stable pricing with slightly more customer preference. Healthy growth, not a discount.

That distinction — is this a price effect or a demand effect — is what an experienced analyst gives you. And it’s what a dashboard never will.

Land

Price story versus demand story. That’s the analyst’s judgement — and it’s now automatic.

\n\n

13:35

0:15

“Demo End”

Say\n

That whole sequence took seconds.

In most retail businesses, that is five to seven days of work.

Delivery

Say it flatly. No excitement needed — the contrast does the work.

\n\n

Close

\n\n

13:50

1:00

✂ trim to 0:45

Set up by analysts, used by executives

Say\n

Now, I showed you one role. The same agent serves several.

The District Manager asks why a store is down and what to do — and gets an answer in minutes instead of days.

The Category Manager asks where mix is hurting them, and which brands need a pricing change — and gets pricing effects separated from demand effects, without manual analysis.

The Pricing Manager asks which products are overpriced or underpriced by store — supported by elasticity evidence.

And the Commercial Lead asks what happens if they raise price five percent — modelling the scenario before committing to it.

Notice the pattern: analysts build it once. Executives use it every day.

Land

Build once, serve the whole commercial team.

\n\n

14:50

1:00

Key considerations

Say\n

Four things make this work, and I want to name them plainly.

Custom instructions. Role, context, framework, output style, escalation rules — written once, applied identically in every session. Consistency comes from here.

Guard rails. Scope limits keep it inside curated data. Confidence flagging surfaces uncertainty instead of hiding it. And escalation triggers when the data isn’t sufficient.

A semantic layer. Human-readable field names and tables built for the agent, so context holds across a long conversation without re-briefing it every time.

And human in the loop. The District Manager reviews and approves every recommendation before any commercial action is taken. The agent advises. People decide.

That last point isn’t a limitation we’re apologising for. It’s the design.

Land

The agent advises. People decide.

Delivery

“The agent advises. People decide.” — two short sentences. It’s the line people repeat afterwards.

\n\n

15:50

1:00

✂ trim to 0:45

Comparison — BI, generic LLM, Agent Studio

Say\n

Let me place this against the two things you already have or have considered.

Traditional BI. Good for what happened — if you know which view to open. But it’s retrieval only. It doesn’t decompose, hypothesise or recommend. Logic like “revenue up but units down” needs development work. Every new question is a new request. Built for reporting, not diagnosis.

A generic LLM over your data — ChatGPT, a copilot, RAG over a warehouse. Genuinely good conversation. But no analytical framework, so it guesses at your fiscal year and your methodology. No native connection to curated retail data, which raises hallucination risk. And no governance, no guard rails, no auditable trail. Capable of talking — not suitable for governed commercial decisions.

Agent Studio plans across multi-step problems, calls tools, and reconciles its own results. The framework is encoded once and applied consistently. And it connects natively to Auto Insights, so access is governed and auditable.

To be fair — you still need curated data and someone to define the logic. That’s the investment. But it’s an investment most of you have already started.

Land

Dashboards can’t reason. Generic AI doesn’t know your business. This does both — with governance.

\n\n

16:50

0:30

Close

Say\n

Let me finish with one thought.

Everyone in this room already has the data. You have dashboards. You have skilled analysts.

The gap isn’t information. The gap is the time between seeing a number and understanding it.

And that gap is where margin quietly disappears.

What I’ve shown you today closes it — using the logic your team has already built, with guard rails you control, and a person still making the final call.

We have an office here in Riyadh, at KAFD. If you’d like to see what this looks like on your own data, come and find me afterwards.

Thank you.

Delivery

Then stop. Don’t add anything. Let it sit.

\n\n

Timing

Full script runs 17:20. Here’s how to land 15:00.

\n \n \n \n \n \n \n \n \n \n \n

Full script, every line as written	17:20
Trim slide 2 — cut the Billigence office list	−0:15
Trim slide 7 — one line, then move on	−0:20
Trim slide 10 — one line, then move on	−0:10
Trim slide 18 — drop the second paragraph	−0:15
Trim slide 19 — drop the fifty-stores aside	−0:15
Trim slide 22 — cover two roles, not four	−0:15
Trim slide 24 — drop the “to be fair” paragraph	−0:15
Trimmed running time	15:15

Protect

Never compress slides 14, 17 or 23. They carry the guard rail, the partial-period catch, and “the agent advises, people decide” — the three things the room will remember.

\n\n

Likely questions

Isn’t this just a chatbot on our dashboard?: A dashboard tells you what happened. This decomposes it, separates price from demand, and recommends an action — using logic your team already approved.

How do you stop it hallucinating?: Three ways. It can only see datasets you approve. The analytical steps are fixed, so it can’t wander. And explicit guard rails — like no conclusions from fewer than thirty transactions.

Will it replace our analysts?: No. Someone has to design the framework, define the logic and set the guard rails. That’s your analyst — and it’s better work than pulling the same report every week.

How long does it take to implement?: The agent is the fast part. The real work is curated data and agreed definitions. If you’re already using Alteryx, a lot of that exists.Do not invent a timeline.

Which AI model? Where does our data sit?: Alteryx One works with different models and runs in cloud, on-premises, or hybrid — so model choice and data residency are your decisions.If pressed: “Let me get you the precise answer rather than guess.”

Can we use it in Teams or Slack?: Yes — Agent Studio can extend agents into the tools your people already use.

What if a recommendation is wrong?: Then the logic or a guard rail was wrong — and both are visible and fixable. That’s exactly why grounding it in your business logic beats a general AI tool.

Does this only work for retail?: No. The pattern is: a number moved, find out why, recommend an action. That works for churn, margin, claims, supply chain — anything measured over time.

Is this real data?: It’s synthetic data built on a real retail structure — so the method is real, and the numbers are safe to show.

\n\n

Delivery

Do

Pause after questions. Silence feels long to you and thoughtful to them.
Say numbers slowly: twenty percent. Thirty transactions. Five to seven days.
Say “you” and “your business” — never “the client” or “the user”.
Screenshot everything before you present. You’re on someone else’s wifi.

Don’t

Don’t say “AI-powered”, “leverage”, “synergy”, “game-changing”. Plain words sound more confident.
Don’t read the slide aloud. They can read.
Don’t apologise more than once if something breaks.
Don’t name a currency on slides 17 and 20 until you’ve fixed the AED/$ mismatch.
Don’t invent a number under pressure. “Let me get you the exact figure” is a strong answer.

\n\n

Practice plan

Three passes

First — read it aloud sitting down, for the words.
Second — read it aloud standing, with the deck open beside you.
Third — use only the panic card. If you can do it from those eight lines, you’re ready.

\n\n

\n Sources — Alteryx on Tour KSA 2026, Billigence presentation (25 slides) · Billigence blog, “From Store Performance Questions to Actionable Retail Recommendations with Alteryx One AI” · Figures on slides 17 and 20 are read from the demo screenshots in the deck.
\n Press R for the panic card · B for big text · Esc to close.\n

\n\n

Panic card

\n \n

“If your sales dropped, how long to find out why?”
Open
“Five to seven days. Seventy percent of it is data prep.”
The cost
“Dashboards show what happened. They never tell you what to do.”
The gap
“Five steps. Same order. Every time.”
The method
“Don’t extrapolate from fewer than thirty transactions.”
Slide 14 — your best line
“Nobody asked it to check that.”
Slide 17 — the partial-period catch
“Price story, or demand story? That’s the analyst’s judgement.”
Slide 20
“The agent advises. People decide.”
Slide 23 — the line they repeat

\n\n\n\n\n",[],[168],["D",1787764861058],{"_11":12,"_120":69},{"_11":12,"_115":116,"_68":69,"_30":67,"_117":118,"_119":24},{"_38":106,"_13":16,"_32":35,"_107":71,"_108":-5,"_109":-5,"_110":-5,"_111":-5,"_112":-5,"_113":14,"_114":12},"ctx",{"_52":-5,"_139":211,"_212":213,"_214":215,"_216":24,"_217":24,"_218":24,"_219":83,"_220":-5,"_221":-7,"_222":-7,"_223":26,"_224":24,"_225":24,"_226":-5,"_227":24,"_228":24,"_229":24,"_230":-5},"ENV",{"_182":183,"_184":185,"_186":16,"_187":188,"_189":190,"_191":192,"_193":194,"_195":196,"_197":198,"_199":200,"_201":202,"_203":204,"_205":206,"_207":208,"_209":210},"REACT_APP_SENTRY_DSN","https://586020d86f05489db63a554690acf1e9@o1398359.ingest.sentry.io/6724499","RENDER_GIT_COMMIT","e87eb04e707cf64e7e532c8e8ec395f209e17f15","RENDER_GIT_BRANCH","POSTHOG_API_TOKEN","phc_nmoyuAkStEZuUp69rYlg1DZgrIt19Gx0Y4Hwf8tGgJP","POSTHOG_API_HOST","https://wart.val.town","WEB_API_TEMPLATE","https://{val}.web.val.run","APP_URL","https://www.val.town","MODULE_URL","https://esm.town","VANITY_WEB_API_TEMPLATE","https://{label}.val.run/","DENOLS_URL","https://lsp.val.town","NODE_ENV","production","CLOUDFLARE_TURNSTILE_SITE_KEY","0x4AAAAAACtaIZJH7s9rBMZ6","GOOGLE_CLIENT_ID","875989696899-jtsqkhlbdv5jrlsk8631p7ipjg3c6o62.apps.googleusercontent.com","STRIPE_PUBLISHABLE_KEY","pk_live_51N2zvWEpZTTLipIkVEwXrUG9AT0yeOOEYxszarZT1jpsZV58Ly6wlg3SltZpHN26CUalOBmUSY7adQRQiSkCZz2500qSXNkf8Y","API_URL","https://api.val.town",[],"preferencesKeybindings","default","preferencesTheme","light","preferencesFixers","preferencesCopilot","preferencesSelectionAi","preferencesCloseBrackets","preferencesChatPanelState","preferencesDprint","preferencesDenofmt","notificationCount","flagDenols","isBusinessOrgMember","businessOrgHandle","isBusinessOrgAdmin","isHigherTownieTier","navCollapsed","currentOrgId"]

History · main.ts