Answer in brief
Every AI purchase looks like buying intelligence and is actually buying throughput on one narrow task. Nine systems, nine different tasks, and one question that separates the ones that pay from the ones that sit unused.
The question that decides every one of them
Before any of the nine systems below is worth its price, one question has to have an answer: what happens today when this task arrives and nobody is available? If the honest answer is that it waits and everything is fine, the automation is buying convenience. If the answer is that it is lost, or done badly, or done by someone whose time costs more than the system, the automation is buying capacity, and capacity is what justifies four-figure prices.
That distinction explains almost every disappointed AI budget. Systems bought for convenience get used for a month and quietly abandoned, because the manual path still exists and is still easier for the person who knows it. Systems bought for capacity get used forever, because the manual path was never actually available.
The VIT MARKET catalogue prices nine of these systems, and reading them in that light rearranges them considerably. What follows is each system, what it genuinely replaces, and the volume at which it starts making sense.
The knowledge-base chatbot: $800 to $2,500
A retrieval-based site chatbot, five to ten working days, marked "needs experience". It replaces the first message of a support conversation — the one where a customer asks something already answered in the documentation and waits four hours to be told so.
What it does not replace is the second message. Retrieval answers questions whose answers exist somewhere in writing. It fails, and should fail visibly, on anything requiring a decision: a refund outside policy, an exception, a promise about a delivery date. A bot configured to attempt those produces confident answers that the company then has to honour or retract, and both are expensive.
The volume threshold is not traffic, it is repetition. A site with modest traffic where eight questions account for most enquiries is a better candidate than a busy site where every question is different. Before buying, count how many distinct questions arrived last month. If the top ten cover more than half, the system will earn its price.
The dependency is the documentation. This system inherits the quality of what it reads, exactly and without mercy. A company whose help pages contradict each other will get a bot that contradicts itself, and no amount of prompt engineering fixes a source that is wrong.
The Telegram assistant: $500 to $1,500
Four to eight working days for a client-facing assistant in Telegram. The interesting thing about this position is that it is cheaper than the site chatbot while doing something similar, and the reason is the channel rather than the intelligence.
A messenger conversation has a thread, a known user, and a natural expectation of asynchrony. Nobody is surprised when a reply takes a minute. A website chat widget carries the opposite expectation — instant, anonymous, and abandoned in seconds if it disappoints. The looser expectation is what makes the messenger version simpler to build and more forgiving in use.
It replaces the repetitive half of an account manager's day: order status, opening hours, where is my invoice, how do I change a booking. The half it does not replace is the relationship, and a business whose Telegram channel is genuinely personal should think carefully before automating the impression of personal attention it has spent years earning.
The email autoresponder: $600 to $1,800
Five to nine working days for inbound mail handling. Email is the channel where automation has the most room and the least tolerance, because an email is a document rather than a chat message: it is forwarded, quoted, printed and occasionally attached to a dispute.
The realistic deliverable is triage rather than replacement. Sorting, routing, drafting a reply for a human to approve, extracting the order number from a wall of text — each of those is a defined win. Sending unreviewed replies to inbound mail is a different risk category, and the price does not change but the exposure does.
The volume that justifies it is roughly where a person is spending more than an hour a day reading mail to decide who should handle it. That hour is pure sorting cost, and it is the part a machine does without getting bored, which matters because bored sorting is where misrouted mail comes from.
Document recognition: $1,000 to $3,500
Ten to eighteen working days, marked "expensive and complex", turning invoices, delivery notes and contracts into structured data. This is the system where the arithmetic is least ambiguous and most often skipped.
Take the real number: documents per month, multiplied by minutes per document, multiplied by the cost of the minute. Below roughly a hundred documents a month, the manual cost rarely clears a four-figure build inside a year. Above a thousand, the manual cost is a job nobody wants to do, which means it is also a job with an error rate that nobody is measuring.
The hidden value is not the time. It is that structured data can be checked. A hundred invoices in a folder cannot be queried for duplicates, for prices that changed without notice, for a supplier who quietly started rounding differently. The same hundred in a table can, and that capability tends to pay for the build faster than the typing it replaced.
The hidden cost is the exceptions. Every document set contains a minority that do not conform, and the correct design routes those to a human rather than guessing. A system with no exception path produces a clean table with wrong numbers in it, which is worse than no table.
CV screening: $600 to $1,800
Five to ten working days, screening applications against a specific vacancy. This is the position that most needs a boundary drawn around it, and the boundary is legal as much as technical.
Used as a sorter, it is straightforwardly useful: does this application meet the stated hard requirements — the licence, the language, the years, the location. Those are facts, they are checkable, and reading two hundred applications to establish them is exactly the work a person does badly by the fiftieth.
Used as a ranker of people, it inherits every bias present in whatever it learned from, and it applies that bias at scale and behind a number that looks objective. The defensible design is to let it exclude on stated criteria, surface the rest to a human, and keep a record of why each exclusion happened. That record is also what makes the decision explainable if anyone ever asks.
The volume that justifies it is a vacancy that attracts more applications than a person can genuinely read. Below that, the system is solving a problem that does not exist, and above it, the alternative is not careful reading but skimming.
The voice bot: $1,500 to $5,000
Twelve to twenty working days, the most expensive system in the group and the one with the clearest justification when it fits. It answers inbound calls.
Nobody should buy this to make an existing receptionist faster. It is bought when calls arrive outside working hours, or in bursts that no realistic headcount absorbs, or in a business where a missed call is a lost customer rather than a callback. Restaurants, clinics, service dispatch, anything with a booking. The return is measured in calls that previously rang out.
The failure mode is specific and worth stating: a voice system that handles eighty per cent of calls competently and the remaining twenty badly can be worse than no system, because the twenty per cent are disproportionately the urgent ones. The design that works routes anything it does not understand to a human quickly and without making the caller repeat themselves. The design that fails keeps trying.
The twelve-to-twenty-day span is honest about that: most of the effort is not the conversation, it is the handover.
The content pipelines: $300 to $3,000, and a per-item price
Three entries share a shape. Batch product-card generation is $400 to $1,500 to set up plus $0.10 to $0.50 per card. SEO rewriting of product descriptions is $300 to $1,200 per project, three to six working days for the pipeline and then batches. An article production pipeline with fact-checking is $1,000 to $3,000, ten to sixteen working days.
Publishing a per-item price alongside the setup cost is the honest part, and a buyer should insist on it for anything that runs per unit. A pipeline is not a purchase, it is a purchase followed by a meter, and a quote that hides the meter is quoting half the cost.
Product cards are the strongest case here because the alternative is genuinely worse. Ten thousand items with no descriptions is a catalogue that cannot be searched or indexed, and hiring that out per card at human rates is arithmetic nobody survives. The fact-checking clause on the article pipeline is the one that matters most: at scale, unverified generated text is the fastest route to publishing something wrong at volume, and the pipeline that costs three times as much is the one with a checking stage in it.
The constraint that applies to all three is the same one search engines enforce: mass-produced pages that differ only by a swapped noun are treated as scaled abuse at the domain level. A generation pipeline is a distribution mechanism for content that would have been worth writing anyway. It is not a substitute for having something to say.
What none of them replace
Every system above narrows to a task with a defined input and a checkable output. None of them replaces a judgement call, and the ones that appear to are the ones that cause trouble, because a confident wrong answer is more expensive than a slow right one.
The practical consequence is that each purchase needs a named owner. Not a supervisor of the machine, but the person who reads what it produced last week and notices when it drifted. Systems without that person degrade invisibly: the knowledge base goes stale, the exception rate climbs, the sorted mail starts landing in the wrong queue, and nobody discovers it until a customer does.
That ownership is also the honest reason to stage these purchases rather than buying three at once. Each system is a small ongoing obligation, and three obligations landing in one quarter on a team that had none is how good tools end up switched off.
An order that works
Start where the input is structured and the output is checkable: document recognition if the volume is there, product cards if the catalogue is large. These have unambiguous success criteria and they teach the organisation to trust machine output by letting it verify machine output.
Then automate the channel with the most repetition — usually the messenger assistant or the email triage, whichever carries more of the same question over and over. Measure the repetition before choosing; intuition reliably picks the noisier channel rather than the more repetitive one.
Leave the voice bot until the case is unarguable, because it is the most visible to customers and the least forgiving of a bad handover. And treat the knowledge-base chatbot as a documentation project with a bot attached, because that is what it is: the bot is three days of the build, and the documentation it reads is the rest of the value.
Questions and answers
Which AI system should a business buy first?
The one whose input is structured and whose output can be checked — document recognition at $1,000 to $3,500 if the volume is above roughly a hundred documents a month, or batch product cards at $400 to $1,500 plus $0.10 to $0.50 per card if the catalogue is large. Both have unambiguous success criteria, which is what builds internal trust for the less checkable systems.
Why does a voice bot cost up to $5,000?
Because most of the twelve to twenty working days goes into the handover rather than the conversation. A system that handles eighty per cent of calls well and twenty per cent badly can be worse than none, since the difficult calls are disproportionately the urgent ones. Buying it is justified when calls arrive outside hours or in bursts, and a missed call means a lost customer.
What decides whether a knowledge-base chatbot is worth $800 to $2,500?
Repetition, not traffic. Count the distinct questions that arrived last month: if the top ten cover more than half of them, the system earns its price. And it inherits the quality of the documentation exactly, so a company whose help pages contradict each other will get a bot that contradicts itself.
Is AI CV screening safe to use?
As an exclusion tool against stated hard requirements — licence, language, years, location — yes, because those are checkable facts. As a ranker of people it applies whatever bias it learned at scale behind a number that looks objective. The defensible design excludes on stated criteria, passes everything else to a human, and keeps a record of why each exclusion happened.