Choosing an AI Service: The Catalogue Really Only Offers You Two

A shelf with a hole in the middle

Sort the twenty-five artificial-intelligence listings in the VIT MARKET catalogue by how long each one says it takes, and a hole opens up. Seventeen listings quote a starting estimate of three to eight working days. Eight quote ten working days or more. Nothing at all is quoted at nine. Not one row.

That empty space is the only buying decision in the section that matters, because it is not merely a scheduling gap. It is also a price gap and a difficulty gap, and the three line up so exactly that they partition the twenty-five listings into the same two piles. A buyer who already runs a business is not choosing between twenty-five services but a side of one line.

A disclosure before any figure appears. VIT MARKET is VITON13's own services catalogue and VJOURNAL is VITON13's newsroom, which is why nothing below is asserted from memory. Every number here is computed from the catalogue's own priceMinUsd, priceMaxUsd, duration and difficulty fields across all one hundred listings, and a midpoint always means floor plus ceiling divided by two. Where a sentence is our interpretation, it says so.

Twenty-five bands, laid end to end

Start with the shape of the money. Published floors run from $300 to $1,500: first quartile $500, median $700, third quartile $1,000. Ceilings run from $1,200 to $5,000 with a median of $2,000. Midpoints — the figure worth carrying into a budget meeting — begin at $750, sit at a median of $1,350 and stop at $3,250. The mean midpoint of $1,582 sits $232 above the median, the usual signature of a few very large listings dragging an average upward.

For scale, the same arithmetic run across the catalogue's other three sections gives median midpoints of $1,000 for web development, $450 for SEO and traffic, and $340 for automation. At $1,350 the AI section is the most expensive of the four by median, and 3.97 times the automation section.

There is no cheap shelf here, and that is worth stating plainly. The lowest floor anywhere in the section belongs to SEO rewriting of product descriptions at $300 to $1,200. Over in the automation section, six listings have floors at or below $100 and four can be bought outright — ceiling included — for less than $300. In other words, four complete automation services cost less than the cheapest possible entry into the AI section.

One more shape: every band is roughly three times as wide as its own floor. Ceiling divided by floor gives a median of exactly 3.0 across the twenty-five, a minimum of 2.75 and a maximum of 4.0. What moves a project from one end of its band to the other is explained nowhere.

The gap is a staffing line, not a schedule

Now put the duration gap next to the money. Of the seventeen listings that start inside eight working days, the highest floor is $900. Of the eight that start at ten working days or more, the lowest floor is $1,000. The ceilings behave the same way: the short group tops out at $2,500, the long group starts at $3,000. There is no overlap on either axis — not a single dollar of it.

The difficulty labels agree. All eight of the long listings carry the catalogue's hardest label, Advanced. Of the seventeen short ones, fifteen are Experienced, one is Beginner and exactly one is Advanced. The median midpoint of the short group is $1,200; of the long group, $2,300 — a step of 1.92 times.

Three fields that are not derived from one another — a free-text duration string, two integers and a categorical label — agree on the same partition. VJOURNAL's reading, offered as interpretation rather than fact: that is not a pricing quirk but a description of who is expected to do the work. A project that needs a second week is a project that needs a different person.

Below the line: seventeen jobs that finish inside a fortnight

The short group is where a business already trading usually makes its first purchase. It holds the batch and back-office work: AI moderation of comments and user content at $500 to $1,400 over four to eight working days; transcription and summary of calls at $500 to $1,600; an autoresponder for incoming mail at $600 to $1,800; a chatbot on the company knowledge base at $800 to $2,500 over five to ten.

Inside that group sits a cheaper cluster of four, and the catalogue marks it out itself. Four listings carry a duration string that ends the estimate at the machine rather than the work — three to six working days per pipeline, then in batches; three to seven to set up the pipeline, then hours per pack; four to eight for the pipeline, then hours per pack; five to ten for the pipeline. Those four have a median midpoint of $950. The other twenty-one listings in the section sit at $1,650.

Read that discount honestly. They are not cheaper because the work is smaller. They are cheaper because the quoted band stops earlier. What the batches cost once the pipeline exists is quoted nowhere and cannot be computed from any field in the data.

Above the line: eight jobs that name the person they replace

The long group is small enough to list in full. Recognition of invoices and contracts into structured data, $1,000 to $3,500 over ten to eighteen working days. Article production with fact checking, $1,000 to $3,000. Preparing a dataset for further training of a model, $1,000 to $3,000 over twelve to twenty-two. Implementation of AI in an existing CRM, $1,000 to $3,500. AI assistant for legal and accounting documents, $1,200 to $3,500. An AI agent for cold sales and an AI analyst answering database questions, both $1,200 to $3,500 and both fourteen to twenty-five working days — the longest estimates in the catalogue. And the section's most expensive listing, a voice bot for incoming calls, $1,500 to $5,000.

What these eight share is stated in the catalogue's own words, not ours. The invoice listing says it replaces the accountant's manual input. The cold-sales listing says it replaces a junior salesperson at the first-touch stage. The database analyst listing says the analyst ceases to be a bottleneck for routine issues. The CRM listing says salespeople stop spending a third of their day filling out fields. The legal one says a lawyer spends ten minutes instead of an hour on initial analysis.

VJOURNAL's reading: the line between the groups is the line between buying a task and buying a role. That reading has one clear exception, and we would rather name it than smooth it over — preparing a dataset for further training replaces nobody. It produces an input for a model. It sits above the line on price, duration and difficulty, but not on the pattern.

Nothing here is sold by the hour, and this is the only section where that is true

Every one of the twenty-five AI durations is denominated in working days. All twenty-five. No other section is that uniform, and none is that slow to start: all twenty-five web-development listings are quoted in hours, automation mixes nineteen day-quoted listings with six in hours, and SEO runs sixteen in days, seven in hours and two in minutes.

The consequence lands on anyone who wants to test a supplier before testing an idea: there is no small slice of AI work to buy. In automation the shortest listing is two to four hours at $60 to $200. Here the shortest commitment is three to six working days and the cheapest is $300 — the same listing.

It also means the per-hour comparisons that work elsewhere do not work here. VITON13's first-party dataset computes an implied hourly rate for every priced listing with a parseable duration and reports the AI section's median at $19.74, the lowest of the four. That figure rests entirely on a declared conversion of eight hours to a working day, a constant the script chose rather than measured. Restrict the same calculation to listings already denominated in hours and the AI section returns nothing, because it has none. What an hour of AI work costs here, the catalogue does not say.

One listing admits that the bill keeps arriving

One listing in the section is a prompt audit and API bill reduction service: $800 to $2,500, Advanced, five to twelve working days. It describes analysing an AI product already in production and cutting API costs through caching, choosing the right model per task and reworking prompts. It claims a reduction of two to five times. That claim is the listing's, not ours; VJOURNAL has not verified it and it is not derivable from anything in this repository.

Two things follow regardless of whether the multiple holds. The running cost is real — a catalogue does not sell a service to shrink a bill that does not exist. And that bill appears in none of the other twenty-four bands. Every price in this section is a build price; the cost of operating what was built is quoted nowhere in the catalogue, in any section, and no field in the data records it.

For a buyer that is the largest gap between the published number and the eventual one. The $1,350 median midpoint is what it costs to have the thing. It is not what it costs to keep it.

What these bands cannot tell you

A price list is silent about most of what a buyer wants to know. Three of the twenty-five publish a numeric outcome claim: the mail autoresponder says it unloads the mailbox by 50 to 70 per cent, the legal assistant says a lawyer spends ten minutes instead of an hour, the prompt audit claims two to five times. Those are the supplier's claims, none verifiable from this repository, and the remaining twenty-two publish no performance figure at all.

Not one of the twenty-five names a model, a provider or a context window, and sixteen do not carry the letters AI in their title. Nothing records how many have been sold, how many buyers came back, or how often a project landed at the ceiling rather than the floor.

One conversion is ours and should be treated as ours. The section quotes working days; fourteen to twenty-five working days is roughly three to five calendar weeks on a five-day week. That arithmetic is VJOURNAL's. The catalogue never states a working week.

Four questions that pick your side of the line

Does the job have a person attached to it today? Every listing that names a role it displaces sits above the line, and everything above the line starts at $1,000 and ten working days. If a named employee currently does the work, the short group is unlikely to hold your answer.

Is the output consumed in batches or live? The four pipeline listings are the cheapest cluster at a $950 median midpoint, but their estimates stop at the machine. If your output arrives in packs — product cards, banners, translated pages — you are buying a pipeline plus a recurring cost the band does not price.

Can you fund a second bill you cannot yet size? Nothing in the section prices the running cost, and one listing exists specifically to reduce it.

Do you need it inside two weeks? Then seventeen rows are open and eight are not, and the fastest thing on the shelf still takes three to six working days. The useful question was never which of twenty-five AI services to buy. It is whether you are buying a task or a role — the catalogue has already priced those as two products and left a nine-day hole between them.

Decision framework: how to choose an AI service

Choose an AI service by the decision or workflow it must improve, not by the model name. Data access, human review, failure cost, evaluation method and operating owner determine whether a pilot can become a dependable system.

NIST frames AI adoption as continuous risk management across governance, mapping, measurement and management. The service card should therefore expose how quality is tested and who handles exceptions.

Write one measurable use case, assemble representative test cases, define an unacceptable failure, assign a reviewer and compare providers on observed results before committing to integration.