Six functions, six vendors, six discovery phases. Each one re-learns the same company from scratch. Year three costs more than year one.
USE CASE TO USE CASE
The second one should be cheaper than the first.
Most enterprise AI spend is repeat spend. New function, new vendor, new discovery, new data modeling, new integration, new security review. Nothing carries. This is the argument for doing it the other way, told through one company over twenty-two months.
THE USUAL PATTERN
WHAT WE ARGUE FOR
One platform, one accumulating model of how your business works. Each use case ships faster than the last because the ground under it is already mapped.
THE HONEST CAVEAT
This only compounds if the use cases are genuinely related. Two unconnected corners of a conglomerate will not help each other, and we will say so.
TWENTY-TWO MONTHS, SIX FUNCTIONS
Seven brands, one team
Seven brands, seventeen sites, three shifts. It began with an overloaded consumer affairs team and ended with marketing reading three years of consumer language. Nothing in between was rebuilt.
MONTH 0 TO 3 · WHERE IT STARTED
One overloaded consumer team, seven brands
This company sells seven household brands. One small consumer affairs team answered every question about all of them, ingredients, allergens, recipes, shelf life, and the one that came up more than any other: where can I actually buy this?
Volume tripled every holiday season and the team could not triple with it. Nights and weekends went unanswered. Nobody was reading what consumers actually said.
The first agent went live across web chat, social messages, email and voice. It answered in brand voice, knew the allergen matrix cold, and turned every “where do I buy this” into a live retailer and delivery option.
WHAT GOT BUILT
- Consumer agent across five channels
- Product, ingredient, allergen and claim model for all seven brands
- Retailer and delivery routing
- Every conversation captured and themed
WHAT IT LEFT BEHIND
- Product master, modeled once, never again
- Brand voice and approved claims
- First connectors: PIM and the DTC store
MONTH 4 TO 7 · SECOND USE CASE
The complaints were already in the building
Quality did not ask for an AI project. They asked a question: the consumer agent is hearing about seal failures, which line and which lot?
Consumer complaints had always been summarized by brand, monthly, in a slide. What quality needed was complaint volume by line, lot, shift and plant, within hours.
Because the product model already existed, this was a matter of adding the manufacturing side of it. Then vision inspection went onto the packaging lines and started feeding the same quality picture from the other direction.
WHAT GOT BUILT
- Complaint-to-lot linking across the plant network
- Print and label inspection on packaging lines
- Defect themes routed to the responsible plant, not the brand team
CARRIED OVER FROM STAGE 1
- The entire product and packaging model
- The consumer conversation history, now a quality signal
- Added: plant, line, shift, lot, defect code
MONTH 8 TO 12 · THIRD USE CASE
Out of stock in exactly the places people were asking
A pattern showed up that nobody had been positioned to see. The consumer agent knew, postcode by postcode, where people were asking to buy a product. Supply knew where it had shipped. Those two things had never been in the same place.
Consumer questions turned out to lead retail out-of-stock by roughly two weeks. Not a forecast, an actual demand signal the company already owned and had been discarding.
The work here was mostly joining: shipments, DC stock, retailer feeds and the existing consumer signal, with agents that flagged the divergence rather than producing another dashboard nobody opened.
WHAT GOT BUILT
- Demand signal from consumer intent joined to shipment and DC data
- Out-of-stock early warning by retailer and region
- Replenishment exceptions surfaced to planners with the evidence attached
CARRIED OVER
- Product, brand and retailer already modeled
- Consumer intent from stage 1
- Added: DC, shipment, retailer feed, forecast horizon
MONTH 12 TO 16 · A DIFFERENT FUNCTION ENTIRELY
The surprise: quality work paid for HR
Different department, different sponsor, no obvious connection. Plants running three shifts, one HR service center, and roughly forty percent of tickets were the same six questions about pay, leave, benefits and shift rules.
The expected answer was a twelve-month HR AI program. It took four months, because the hard part was already done.
Plant, site, shift and crew had been modeled for quality in stage 2. HR needed exactly that structure to answer “what is my accrual as a second-shift line operator at Rochelle” correctly. The quality project had unknowingly paid for the HR one.
WHAT GOT BUILT
- AskHR across policy, benefits, pay and shift rules
- Available on the plant floor, in the languages people actually speak
- Case deflection with clean handover when a human is needed
CARRIED OVER
- Site, plant, shift and crew model, built for quality, reused whole
- Existing identity and permission plumbing
- Added: employee, role, policy, benefit plan, accrual rule
MONTH 16 TO 22 · WHERE GOVERNANCE GOT SERIOUS
I-9, FMLA and the end of the compliance spreadsheet
Answering questions is one thing. Compliance is where an agent can genuinely hurt you, and this is the stage that decides whether an enterprise AI program is real.
I-9 reverification exposure sat across seventeen sites in a spreadsheet one person maintained. FMLA, ADA and STD eligibility maths was done by hand, differently in different regions, with correspondence written from memory.
Nothing here runs autonomously. Every eligibility determination and every piece of correspondence has a named human checkpoint before it becomes a position. What changed is that compliance became a queue with owners, dates and evidence instead of an audit surprise.
WHAT GOT BUILT
- I-9 and work authorization tracking with a reverification calendar
- FMLA, ADA, STD and return-to-work as one governed workflow
- Eligibility maths shown with its working, correspondence drafted not sent
- Full audit trail: who decided, on what evidence, when
CARRIED OVER
- Employee, site, role and shift, unchanged from stage 4
- Added: work authorization document, leave type, statutory clock, jurisdiction rule
- Human-in-the-loop gates and evidence retention now enterprise defaults
MONTH 22 ONWARD · THE DATA THAT WAS ALREADY THERE
Three years of consumer conversation, finally used
By this point the company held something its competitors did not: three years of unprompted, specific consumer language about its categories, tagged by product, region, season and sentiment. The best category research in the business, and nobody was reading it.
Marketing now works from it weekly, what people are confused about, what they compare against, which claims land and which get quietly ignored.
Then a newer question. When a consumer asks an AI assistant which canned tomatoes to buy, what does it say, and on what basis? The same content and claim model that powered the consumer agent turned out to be the lever for being represented accurately there too.
WHAT GOT BUILT
- Weekly category intelligence from real consumer language
- Claim performance and confusion tracking by product line
- Generative-engine visibility: how the brands are represented in AI answers
CARRIED OVER
- Effectively everything, product, brand, claim, consumer voice, retailer
- Twenty-two months of accumulated meaning, at zero marginal modeling cost
A composite drawn from live deployments in food and consumer manufacturing. Sequence, timings and dependencies are real; the company is assembled.
THE PART THAT MATTERS
Stage four is the whole argument
The HR service desk went live in four months instead of twelve. Not because HR is easy, and not because the agent was clever. Because a completely different project, packaging quality, eight months earlier, different sponsor, different budget, had already modeled plant, site, shift and crew properly.
HR needed that structure to answer a shift worker's accrual question correctly. It was already there, already tested, already permissioned.
That is the entire economics of this. Not that the first project is cheap, but that the fourth one is, and the fourth one is where a program has usually earned the confidence and the budget to keep going.
4 moHR service, live, against a twelve-month plan built before the reuse was understood
0Data models rebuilt across six use cases
6Functions on one platform, one security review, one governance model
WHAT TRAVELS AND WHAT DOES NOT
Being precise about reuse
Vendors are vague here on purpose. We would rather be specific, including about the parts that do not carry.
Travels almost completely
The shape of your business. Sites, plants, shifts, products, suppliers, customers, employees, contracts, the identifiers that join them, and the rules about who may see what. Connectors and their permission models. Evaluation harnesses. Governance, approval gates and audit patterns. The security review you already passed.
Does not travel
Domain judgment. A packaging defect taxonomy tells you nothing about FMLA eligibility. Each function needs its own SME time, its own golden set, its own definition of a good answer, and its own failure modes worked out with the people who own the process. Roughly a third of each new use case is genuinely new work.
EXPANSION PATTERNFLEXDAY.AI
- SEQUENCE
- Consumer service → quality → inventory → HR service → HR compliance → marketing
- ELAPSED
- 22 months, six production use cases
- FIRST USE CASE
- Typically 8–14 weeks from workshop to production
- FOURTH USE CASE
- Typically 4–8 weeks, on the same footprint
- REBUILDS
- None. Each stage extends the model rather than replacing it
- SECURITY REVIEWS
- One platform review, then per-use-case data scope only
- REUSE RATE
- Around two thirds of each new use case; the remaining third is genuine domain work with your SMEs
- PREREQUISITE
- The use cases have to share a business reality. Unrelated corners of a holding company will not compound and we will tell you so early
- NOTE
- Ranges reflect delivered engagements across logistics, packaging, CPG and real estate. Your first use case is scoped and estimated before you commit to it.
Where would yours start?
The first use case is usually the one with the clearest volume and a willing owner. Tell us yours.