The situation
A manufacturer of extruded aluminium window and door systems, 1,100 people across three plants, sells through a trade portal to roughly 900 fabricator, glazier and merchant accounts. The board has AUD$400,000 and one seconded manager for AI this year, and wants one initiative funded and the rest sequenced or refused. Four functions have asked.
- Contact centre. Twenty-two agents take about 12,000 calls a month and pick one of fourteen wrap-up codes on each. 38% land on “Other”, so the head of service cannot say why people call. Calls are recorded, none transcribed; agents type free-text notes on roughly half.
- Finance. About 5,800 supplier invoices arrive each month from 1,300 suppliers, almost all PDFs on email in whatever layout the supplier uses. Three accounts-payable staff key them into the ERP at a little over four minutes each, and around 7% go into a query queue.
- Merchandising. The trade portal carries 4,200 SKUs, and the commercial manager wants a “customers who bought this also bought” block on every product page.
- The line. A plant manager wants surface defects caught before a pallet is wrapped: scratches, coating pinholes, colour drift. The three plants wrap about 3,000 pallets a month, and an operator eyeballs them at the wrapping station. Around 1 in 60 of the pallets that ship comes back at roughly AUD$1,800 in freight, rework and credit: 600 returns a year, a bit over AUD$1 million.
All four arrived as “an AI project”, with no capability named and therefore no price.
What actually matters
Four asks cannot be compared until each has been named, because the name carries the price shape. Document extraction bills per page processed. A foundation model bills per token in and out. A custom model trained and hosted for one factory bills per instance-hour whenever the endpoint is up, running line or not. A business intelligence tool bills per seat per month.
Finance has 5,800 documents a month in a mailbox, the portal has six years of order lines against account numbers, and the contact centre has text and audio, though the audio is unindexed recordings rather than transcripts. Computer vision needs labelled images, and nobody has ever photographed a defective profile beside a good one and written down which is which. Three of the four have their input, one does not, and that difference is usually a year.
Bought or built decides who owns accuracy after go-live. A packaged capability arrives with the vendor’s accuracy and improvement schedule, and the business owns the exception process. A built one means somebody here owns a model’s error rate forever, watches it drift as the product mix changes, and retrains it: a standing role, not a project cost, and the line item most often missing from a business case written by a function that has never run a model.
Finance knows minutes per invoice and exceptions per hundred, so a before-and-after number exists already. The contact centre knows call volume but has never recorded why people call in a way anyone trusts. On the line a false alarm costs a few minutes of inspection on a good pallet while a miss ships a defect and costs AUD$1,800, so a single accuracy percentage hides the miss inside the average.
What we’ll filter on
- Would a query or a rule already answer it? If the answer is a definite value that exists in a system, counting it beats predicting it.
- Does the input exist in a form a machine can read? Not “do we have the data” but “is it captured, joined to something, and machine-readable today”.
- Bought or built? A packaged capability behind an API, or a model this business has to train, host and own.
- Is there a before-and-after number somebody already collects? A measure that predates the project, so nobody has to argue about the baseline afterwards.
- Who owns accuracy once it is live? A named function that watches the error rate, works the exceptions, and decides when to retrain.
The landscape
AI, machine learning and generative AI
AI covers systems doing work that would need human judgement. Machine learning is the subset that learns from past examples rather than written rules, so it needs cases with the right answer attached. Generative AI is the machine-learning subset built on foundation models that make text, images or code from a prompt; its training happened elsewhere, so it works the afternoon it is switched on, where a conventional model cannot start until a year of labelled examples exists.
Ranking and recommendation
An ordered list of items for a customer or an item, out of interaction history. Amazon Personalize ships preconfigured ecommerce recommenders, two named “Frequently bought together” and “Customers who viewed X also viewed”. A recommender bills by the hour it is active, at a rate tiered on how many users sit in the dataset, so the meter runs whether the portal is busy or not. AWS will train one on a minimum of 1,000 item interactions from 25 unique users, and recommends at least 50,000 interactions from 1,000 users for recommendations worth showing. This portal has 900 accounts.
Natural language processing over text and speech
Turns language into something countable: topics, sentiment, entities, categories, summaries. Audio is transcribed first, its own step and its own bill. Packaged for the common jobs; a foundation model on Amazon Bedrock covers the rest with no training.
Computer vision
Classifies or locates things in images and video, trained on labelled examples of each defect class shot under the line’s real lighting. Amazon Rekognition Custom Labels asks for at least ten training images and calls a few hundred or fewer typical, shot in a variety of lightings, backgrounds and resolutions; whether that many produces a model good enough to hold a pallet is settled by the first trial and not before it. No general-purpose image service has a label for what this factory calls a pinhole, so inspection means a custom model on this plant’s images, charged for every hour the trained model is available to process them whether the line runs or not. Amazon Lookout for Vision, the packaged answer here, reached end of support on 31 October 2025; a build now runs on Rekognition Custom Labels, Amazon SageMaker AI, or a vision-capable foundation model.
Document extraction
Pulls fields, tables and key-value pairs out of documents whose layout varies. Amazon Textract prices an Analyze Expense capability built for invoices and receipts specifically. Amazon Bedrock Data Automation handles document processing end to end, including classification, extraction, normalisation and validation, with confidence scores on the output. Billed per page. Measured in straight-through processing rate and minutes of keying saved, and owned by finance, whose exception queue it lands in.
AI for customer operations
The contact-centre bundle: call transcription, contact summarisation, reason categorisation and agent assistance mid-call, over recordings, chat transcripts and a knowledge base. Whether it arrives packaged depends on the contact platform, which increasingly sells these features with it. Transcription bills per minute of audio, so the bill tracks how long people talk rather than how many times they call.
Forecasting and classification
A number at a future date, or a label from a fixed set, over a history with past answers recorded; the shape of the answer sorts between the two. Built rather than bought, on Amazon SageMaker AI. Amazon Forecast is closed to new customers, so a business case naming it names something nobody can buy.
Reporting
Not an AI capability, and here because it answers a surprising share of asks: counting, grouping and charting values a system already holds, bought per seat. On AWS that is Amazon Quick Sight, the business intelligence feature of Amazon Quick, formerly Amazon QuickSight. A Reader who only opens dashboards lists at USD$3 per user per month and an Author who builds them at USD$24, and an account carrying any Pro user or natural-language Q&A picks up a USD$250 monthly per-account infrastructure fee on top, so a seat count is not always the whole bill.
Evaluation
Side by side
| Capability | Query or rule would do it | Input exists today | Bought | Baseline number exists | Clear accuracy owner |
|---|---|---|---|---|---|
| Document extraction (invoices) | ✗ | ✓ | ✓ | ✓ | ✓ |
| Ranking and recommendation (portal) | ✗ | ✓ | ✓ | ✗ | ✓ |
| NLP over text and speech (call reasons) | ✗ | ✓ | ✓ | ✗ | ✓ |
| Reporting (call reasons, the coded 62%) | ✓ | ✓ | ✓ | ✓ | ✓ |
| Computer vision (surface defects) | ✗ | ✗ | ✗ | ✓ | ✗ |
Computer vision is the only ask with no machine-readable input, because the images do not exist, and no accuracy owner, because quality has nobody who has run a model. Either one stops the conversation on its own. Reporting’s tick in the first column disqualifies it as an AI initiative and makes it this week’s work for the contact centre. Ranking’s tick in the second column is the one to read closely: the order lines are there, but 900 trade accounts sits under the user count AWS recommends, so the first recommender may be thinner than the commercial manager expects.
“Why are people calling” is two questions, and splitting them makes the ask fundable. The top reasons by volume are already recorded: agents pick a wrap-up code on every call, and counting fourteen codes across 12,000 calls is a query. What the codes cannot say is what sits inside the 38% filed as “Other”, roughly 4,500 calls a month, which needs transcription and language processing over recordings and free-text notes. Run as one project, the reporting ships only when the model does.
Naming the capability from how the ask is worded
The solution
Fund the invoice work first. Document extraction is packaged, so the timeline is an integration timeline rather than a training timeline. At Amazon Textract’s published Analyze Expense price of USD$0.01 a page in Oregon, a single-page run of 5,800 invoices a month comes to about USD$58, or roughly USD$2,100 across three years. Finance already tracks minutes per invoice and the 7% query rate, so nobody renegotiates the baseline at the end. A wrongly extracted field lands in an exception queue three people already staff, and the head of AP owns accuracy the way she already owns that queue. Confidence thresholds are a finance decision, since the setting decides how many invoices post unseen and how many route to a human. Make straight-through processing rate the headline measure rather than accuracy: it converts into hours. The AUD$400,000 pays for the ERP integration, the posting rules and the people who work the exceptions.
Merge the contact-centre and merchandising asks into one stream, sequenced behind it. Two functions asked for two projects over one body of data: the account, the order lines, the contacts attached to those accounts. Split across two teams, it gets assembled twice, joined twice, and disagreed about on the first slide carrying both sets of numbers. The seconded manager owns it: one dataset, three deliverables in sequence. The call-reason dashboard in a fortnight, for a few seats and somebody’s week; then transcription and language processing over the 38% it has sized; then the portal ranking off the same account history. Neither half has a baseline, so the ranking block ships to a slice of product pages and attach rate is read against the pages without it.
Refuse the defect ask and give it a camera. It is the biggest number on the table, a bit over AUD$1 million a year in returns, and still the ask that cannot be funded: there are no labelled images, and no money compresses the time it takes to accumulate examples of a pinhole under real line lighting. Mount a camera at the wrapping station, capture every pallet, have the operator who already inspects them tag what he finds, and re-price in a year. The cost is a camera, a mount and lighting, call it AUD$5,000, plus a change to one person’s routine. A business case that says “a year of images, every pallet the operator pulled tagged by defect type” prices in a way that “we want AI on the line” never will.
Extraction and ranking are both bought, so neither needs a data scientist on staff. The vendor improves the extraction and finance works the exception process it already ran; ranking is judged commercially, so somebody has to read what the portal suggests and say whether it makes sense to a glazier. The defect model, if it is ever built, needs that hire, in its own business case.
What’s worth remembering
- No price without a named capability. Document extraction bills per page, a foundation model per token, a hosted custom model per instance-hour, reporting per seat.
- Generative AI arrives trained. A conventional machine-learning capability learns from this business’s own history first, and that is most of the timeline gap.
- Count before you predict. A definite value that already exists in a system needs a query or a rule, not a model.
- No machine-readable input, no project yet. The ask gets a collection job with an owner and a re-pricing date, not a budget.
- Bought or built decides accuracy. A built capability needs a standing role in its business case, a bought one an exception process.
- Two asks can be one dataset. Differently worded asks over the same data merge into one stream, cheaper than two separate business cases suggest.
Nobody in this scenario opened a console. The board left with one funded initiative, a merged stream sequenced behind it, a refused ask everybody agreed with, and five questions for the next four asks. Which service supplies each capability is the next conversation, and it is a shorter one once the capability has a name.