Ofivia blog

What it costs to put AI to work in a company: the real budget

Equipo Ofivia — Mon, 20 Jul 2026 00:00:00 GMT

The question always arrives in the second meeting, after the general manager has already got excited and before finance says anything. And it arrives badly framed: “what does AI cost?” AI does not cost. The software costs, your people’s time costs, model usage costs, and each of those three behaves differently.

This article is the breakdown we would ask a vendor for if we were you. The figures come from real rollouts in Latin American companies of twenty to a hundred and fifty people during 2026, in dollars because that is what the region prices software in.

The four line items, and which one gets underestimated

An honest budget has four lines.

Software licenses. The only one everybody asks about and the smallest of the four in year one.

Model usage. Variable, metered, and the one that scares people because nobody knows how to estimate it. It can be estimated well; how is further down.

Rollout. Connecting, uploading, reviewing permissions, configuring. Usually half of the first-year budget.

Internal time. The line nobody puts in the spreadsheet and the one that actually decides whether the project works. Somebody in-house has to give it hours. If nobody has them, the project does not fail: it simply never happens.

Licenses: between 600 and 3,000 dollars a year

For a forty-person company, the realistic range for a knowledge platform with agents is between eighty and a hundred and fifty dollars a month. That is between 960 and 1,800 dollars a year.

There are two billing models in the market and it is worth understanding the difference, because it changes the invoice six months in.

Per-seat billing charges for every person with access, between ten and thirty dollars per person per month. For forty people that is between 400 and 1,200 dollars a month. It sounds reasonable until you want to give read-only access to the twenty people on the floor, and then each reader costs the same as a heavy user.

Per-company billing with metered usage charges a flat fee plus whatever the AI spends. The flat fee covers the whole team and usage is paid for as it happens. For a company where five people use the agent every day and thirty-five only look documents up, this model costs between a third and a fifth of the other one.

In Ofivia the Basic plan is 80 dollars a month and the Pro plan 150, in both cases for the whole company, with a monthly credit allotment included. We do not charge for readers.

Model usage: how to estimate it without guessing

This is where budgets break, because the vendor says “it depends on usage” and you are left exactly where you started.

It can be estimated from two numbers. The first is what one agent work turn costs. The second is how many turns your team makes in a month.

A turn is a question and its complete answer, with the agent reading the documents it needs in order to answer, in the sense described in AI agents for companies. On the installation we write this from, measured over 119 real work turns, the average came out at 0.19 dollars per turn. That is not a brochure estimate: it is the average of what they cost.

Now the part people calculate badly: how many turns a team makes. Intuition says “forty people times ten questions a day is four hundred a day”. The reality of the rollouts we have seen is very different. In a forty-person company, between four and eight people use the agent daily and make between ten and twenty-five turns each. The rest use it two or three times a week. That gives between 1,500 and 3,500 turns a month.

At 0.19 dollars a turn, between 285 and 665 dollars of monthly usage. Which is considerably more than the license, and that is why two things are worth checking before signing.

Which operations consume nothing. Indexing documents, searching them and walking the link graph should cost nothing, because they do not call a paid model: they run on the platform’s own server. If your vendor charges you to index, you will end up deciding what to upload based on price. In Ofivia indexing, searching, recalling past conversations and people talking to each other consume no credit. Credit is only consumed when the agent works.

How you see the spend. Ask to see the usage screen before signing. It has to answer three questions: how much have I spent this month, which process spent it, and on what day. If the vendor only shows you a monthly total, you have no way of knowing whether the spend came from useful work or from a badly configured automation running every hour.

One warning about spend control, because on this point most of the market over-promises. In Ofivia usage is recorded and shown as it happens, and you see exactly what was spent, by day, by process and by person. What we do not do today is cut off automatically at the ceiling: if a month runs over, you see it in the dashboard and we tell you, and an owner can request more credit from that same screen. We would rather say it that way than sell you a guarantee the system does not yet enforce.

Rollout: between 1,500 and 6,000 dollars, once

This is where most budgets fall short, because rolling out is not installing.

For a forty-person company with around ten thousand accumulated documents, the real work is this.

Permission review: two to five days. Before indexing anything you have to decide who sees what. It sounds administrative and it is the most delicate part of the project. When an importer we worked with indexed its shared folder without reviewing it, a warehouse assistant’s first search returned the suppliers’ purchase price list, margins included. It had been there since 2021 in a subfolder nobody opened. Obscurity was not a permission but for four years it worked like one.

Upload and verification: two to four days. Uploading is quick. Verifying that what went up can be found is what takes time, and it has to be done with real questions.

Department structure: one day. Reproducing the org chart in the system, with its owners and their inherited permissions.

Training: two sessions of two hours. One for everyone, one for the five who will use the agent seriously. The second is the one that changes the results.

At local consulting rates, between twenty-five and sixty dollars an hour, that is between 1,500 and 6,000 dollars. If a vendor tells you the rollout is two hours, either your company is very small or the vendor has not done this before.

The complete first-year budget

For a forty-person company, with five heavy agent users:

Line item	Low	High
Annual license	960	1,800
Model usage (12 months)	3,420	7,980
Rollout (one off)	1,500	6,000
Internal time (80 to 160 hours)	1,200	3,200
First-year total	7,080	18,980

Year two loses the rollout and internal time drops, so it settles between 4,900 and 10,500 dollars.

Put on the scale of a forty-person company: the low end is roughly the total cost of a junior employee for four months. The high end, about eleven months.

When it is not worth it

Three cases where we would tell you to wait.

If you have fewer than fifteen people and all the documentation fits in two heads. The problem does not exist yet. Come back when the third person asks the same question.

If nobody is going to give it the hours. Without an internal owner with time allocated, the system fills with documents and empties of use in two months.

If the material is so out of date that it has to be rewritten. Then the project is not an AI project, it is a documentation project, and it costs what writing costs. AI comes afterwards, and there it genuinely accelerates.

How to ask for a quote you can actually compare

Four questions that separate a serious vendor from one improvising. Ask for them in writing.

Which operations consume and which do not? If indexing consumes, ask for the price per thousand documents.
What is the average measured usage per agent turn on real installations, not estimated?
What exactly happens when the month’s credit runs out? Ask for the literal answer, not the commercial one.
What do I take with me if I leave? The format of the documents and of the database, and in how many days.

If all four answers are clear, the rest of the conversation is about price and you already know how to negotiate that.

The numbers in this article come from the same model we bill on, and it is published with the full breakdown on the pricing page. One credit equals one thousandth of a dollar of real model spend, plans come with their monthly allotment, and the owner can request a mid-month top up. If you would rather see first how that translates into concrete work, the credits explainer has three examples with measured ranges.

RAG explained without the jargon: how AI answers with your company documents

Equipo Ofivia — Tue, 26 May 2026 00:00:00 GMT

RAG stands for retrieval augmented generation, and spelling it out helps nobody. In practice it means something fairly simple: before answering, the system searches your documents and uses what it found to write the answer.

That is all. The hard part is the search, not the artificial intelligence.

It is worth understanding because it decides two very concrete things you care about: what happens when you correct a document, and how safe your material is.

What RAG is not: training a model on your documents

This is the most expensive confusion in the market, because it leads people to buy the wrong thing.

A language model is trained once, on an enormous amount of text, in a process that costs a fortune and takes weeks. What it learns there stays inside the model, blended in, impossible to point at. Training a model on your company’s documents, when it is genuinely offered, means taking an existing model and tuning it with your material. That has three problems for the use you have in mind.

It goes stale the day a document changes. The model learned the old version. For it to learn the new one, it has to be tuned again.

It cannot cite. It knows things, it does not know where from. And an answer with no source is useless inside a company for anything that matters.

It cannot be taken back out. If a document should never have gone in, there is no clean way of removing it.

RAG does the opposite. The model learns nothing about your company. Every time you ask, the system finds the relevant fragments in your documents, hands them to the model along with the question, and asks it to answer using only that.

The three consequences invert. Correct a document and the next answer already uses the correction. It can cite, because it knows which fragments it used. And removing a document from the index removes it from the answers, immediately.

How it works, in four steps

One. The documents get split. A forty-page contract does not fit inside a question, so it is cut into fragments of a few paragraphs. Where you cut matters quite a lot: cutting through the middle of a clause produces two fragments that say nothing.

Two. They get indexed. Each fragment is turned into a representation of its meaning and stored. This is the step that lets you find “commercial premises lease” when you searched for “office rental”.

Three. Retrieval. You ask. The system finds the most relevant fragments. The good ones run two searches at once, one by meaning and one by exact word, and combine both results, because the first is bad with invoice numbers and the second is bad with concepts.

Four. Generation. The model receives the question and the fragments, and writes. With the instruction to answer only from that and to say so when it cannot find the answer.

All the value is in step three. If retrieval brings back the wrong fragments, the best model in the world writes an immaculate, false answer.

Where it fails in practice

Four failures, in order of frequency according to what we have seen in real rollouts.

The right document was not there. The most common failure and the least technical. The answer is written nowhere, or it is sitting in someone’s inbox. The system answers with the closest thing it found, which is the worst thing it can do.

The wrong fragment was retrieved on surface similarity. You ask about the supplier advance payment policy and the system brings back the salary advance policy. They talk about the same thing in almost identical words and they are different things.

The split cut the information in half. The clause said “except as provided in the preceding paragraph” and that paragraph ended up in another fragment. The answer comes out with half the condition.

There were two versions and it brought the old one. This is the quietest and the most dangerous. A law firm we worked with had the same contract in three versions with no mark of which one governed. No retrieval system can know which is the good one if the file does not say.

Three of the four are documentation problems, not technology problems. That is the most useful practical conclusion in the whole article: RAG amplifies the quality of your archive in both directions.

What this means for the security of your documents

The consequence of the model learning nothing is that your documents do not travel anywhere whole: what travels are the fragments needed to answer that particular question, and only for as long as the answer takes. That is a real difference from the alternative of uploading the entire file to a generic tool every time somebody wants it summarized.

What does decide who sees what is the permission layer that sits before the search. If the system retrieves fragments from the whole archive without filtering by who is asking, search turns into an open corridor to documents that person should not open. The right question for a vendor is not whether they encrypt traffic, which everybody does, but whether retrieval respects the asker’s permissions before retrieving rather than after displaying.

What to ask a vendor who offers you RAG

Five questions. The answers separate whoever implemented it from whoever read about it.

Is the search semantic only, or hybrid? If it is semantic only, it will fail on invoice numbers, tax IDs and proper nouns. If they tell you semantic alone is enough, they have not tried it on a real billing archive.

Does it cite the exact fragment or only the document? Citing the file is fine. Citing the section is what makes verifying cost ten seconds instead of ten minutes.

What does it do when it finds nothing? The correct answer is that it says so. If the system always answers something, it is not measuring its own confidence and you will never know when to doubt it.

What does indexing and re-indexing cost? If you are charged per document, you will decide what to upload on price and the archive will end up incomplete, which is failure number one on the list above. In Ofivia indexing, re-indexing and searching run on our own server against a local model and consume no credit. Credit is only consumed when the agent works.

Are the documents used to train anything? The answer has to be no, in writing, in the data processing agreement. It is the question the buying committee asks and the one that ends the conversation fastest when the answer is ambiguous.

What you can do to make it work better

Three things, all of them documentation and none of them technology.

Mark which version is current. One line in the header with the date it takes effect and who answers for the document. It solves the most dangerous of the four failures.

Link the documents to each other. When a procedure links to the policy behind it and that policy to the minutes where it was approved, the system can follow the chain and you can walk it in both directions. In Ofivia those links are written wiki-style inside the text and the graph builds itself, so you can also ask which documents point at this one.

Archive what no longer applies. A procedure for a system the company abandoned in 2023 is not history, it is a trap. Take it out of the index and store it separately.

With those three, the same technology goes from being right half the time to being right nearly always. Without them, no vendor is going to fix it for you.

How the documents are stored, how the index is rebuilt from them, and why the disk outranks the database is explained on the vault page. The detail matters: if the index were deleted it regenerates, and if the folder were deleted it does not, and that asymmetry is what decides who really owns your knowledge.

How to build an internal knowledge base your team will actually use

Equipo Ofivia — Tue, 14 Apr 2026 00:00:00 GMT

Almost every internal knowledge base dies the same way. Somebody sets the tool up with enthusiasm, writes forty articles in three weeks, and two months later the last change is dated forty days ago. Nobody deletes it. It simply stops being opened, and when somebody has a question the answer is once again a direct message to the person who knows.

This article is the method we have seen survive. It is not the most ambitious one. It is the one still alive a year later.

Before you start: do not tidy first

The natural instinct is to clean before uploading. Reorganize the folders, delete duplicates, rename everything to a convention. Nobody finishes that project, because it has no end and because the person doing it does not have that job.

The right order is the reverse: upload what is there, exactly as it is, and tidy afterwards with evidence of what people actually search for. A decent search engine finds things in messy material. What it cannot do is find what is not there.

There is one exception and it matters. Before uploading anything, look at what is inside the folder you are about to upload. At an importer we worked with, the shared administration folder had the purchase margin table by supplier in a 2021 subfolder nobody opened. It had been accessible to half the company for four years and nobody knew, because nobody searched. The new system’s first search surfaced it in third place. Obscurity is not a permission, but while nobody searches it works like one.

Step 1: pick a department, not the company

Starting with the whole company guarantees you will not finish. Pick the area where it hurts most today. Usually operations, sometimes legal, sometimes HR.

Signs that you picked the right department: somebody answers the same question more than once a week, a new hire’s onboarding depends on another person’s patience, and there is at least one process only one person knows how to run.

A clinic we worked with started with nursing rather than administration, because nursing was where the knowledge was most in people’s heads and least written down. It was the right call: within a month they had written down what had gone eleven years unwritten.

Step 2: four document types and nothing else

The temptation is to invent a taxonomy. There is no need. Almost all of a company’s useful knowledge fits into four types, and naming each one makes people aware of what they are writing.

Procedure. How something is done, step by step, so that somebody who has never done it can. “How to onboard a new supplier.”

Policy. What is allowed and what is not, and who decides. “Supplier advance payment policy.”

Reference. Data you look up rather than read through. Price lists, code tables, directories, responsibility matrices.

Decision. What was decided, when, by whom and why. Minutes, committee summaries, notes from a discussion that ended in something.

The fourth type is the one almost nobody documents and the one worth most. Half the arguments that repeat inside a company repeat because the reasoning from the first time was never written down, only the conclusion.

Step 3: a seven-line template

The longer the template, the fewer people write. This is the one that works, and it fits in your head.

Title:          verb + object. "Onboard a new supplier"
Type:           procedure | policy | reference | decision
Owner:          who answers if this is wrong
In force since: date
Applies to:     who uses it
Body:           whatever it needs to be, no format rules
Related to:     links to the two or three neighboring documents

The two lines people skip are the ones holding everything else up.

Owner is what stops the document from dying. A document with no owner ages without anyone noticing. With an owner, there is at least somebody to ask.

Related to is what turns a pile of files into something you can walk through. When the supplier onboarding procedure links to the advance payment policy and that links to the minutes where it was approved, anyone arriving at one of the three finds the other two. In Ofivia those links are written with double brackets, wiki style, and the system builds the graph on its own, so you can also walk it backwards: which documents point at this one.

Step 4: filling it without stopping the operation

Three sources, in this order.

What is already written. Upload it as it is. Old manuals, decks, long emails that were really a procedure. Do not rewrite it yet.

What people ask. For two weeks, every time somebody answers a question over chat or email, have them paste the answer into a new document using the template. Unedited. A raw written answer is infinitely better than a perfect article that does not exist.

What only one person knows. This is where an AI agent earns its place, and it is the use that surprised us most by how well it works. Sit the person who knows down, have them talk for twenty minutes about how they do their job, and have the agent turn that transcript into a draft with the template filled in. The person spends ten minutes correcting what they would never have spent three hours writing. Resistance to documenting is almost never attitude: writing from nothing is expensive and correcting is not.

Step 5: the review that keeps it alive

A knowledge base with no expiration date turns into a cemetery with a search box. The minimum rule we have seen hold is a single one: every document has an owner and an effective date, and once a quarter the owner gets the list of theirs untouched for more than six months. Not to rewrite them. To do one of three things: confirm they still apply, correct them, or archive them.

Archiving is the option nobody uses and the one that cleans most. A procedure for a system the company stopped using in 2023 is not history: it is a trap for whoever joins next.

In companies where this works, the quarterly review takes between twenty and forty minutes per owner. In companies with no owners assigned, the knowledge base averages eighteen months old and nobody trusts it, which amounts to not having one.

How to tell at three months whether it is working

Four signs, in order of importance.

People search before they ask. You notice it in internal chat: fewer messages starting with “quick question”.

New documents are written by more than one person. If the same person writes everything, you built a personal blog.

Somebody corrected another person’s document. That is the sign the material is felt to belong to the company rather than to whoever wrote it.

And the one that measures itself: how long a new person takes to work unsupervised. If it was three weeks and it is now eight days, it has already paid for itself.

All of this needs a place where documents are real files rather than rows in a vendor’s database. In Ofivia’s vault each company gets its own folder with standard Markdown and the originals it uploaded, so the template above is a text file, the links are links, and the day you want to take it all away you take a folder that opens in any editor.

AI agents for companies: what they are, what they do, and what they cost

Equipo Ofivia — Thu, 02 Apr 2026 00:00:00 GMT

At an importer in Guayaquil, the head of purchasing spent Monday mornings with twenty tabs open. Supplier price lists as PDFs, each with its own date, each in its own currency, three of them revised by hand in a later email nobody filed. Her Monday question was always the same: what is the current price for this item from this supplier, and since when.

That is exactly the question an agent answers in about twenty seconds: it reads the same PDFs she was reading and returns the answer with the file name and the page each figure came from. The head of purchasing still checks what comes back. That is part of it working, too.

This article is about what that thing actually is, what it does well, what it does not do, and what it costs.

An agent is not an AI chat

The distinction looks like vocabulary and it is not. It decides what the thing can do for you.

An AI chat takes text and returns text. Everything it knows about your company is what you pasted into the message. If you ask about a supplier contract, you have to paste the contract. It is useful for drafting, summarizing and translating. It knows nothing about your company between one conversation and the next.

An agent takes a task and has tools. It can search your documents, open the ones it needs, read them, write a new file, leave a task noted. It works in several steps: it searches, reads, decides something is missing, searches again, and only then answers. That is why it takes longer than a chat, and why it answers things a chat cannot.

The practical consequence is that an agent is for questions whose answer you do not know the location of. A chat is for working with material you already have in front of you.

The four things an agent does well today

After sitting through several rollouts, these are the four that hold up in daily use. They are not the most impressive in a demo. They are the ones people are still using in month three.

Answering with the source. “What did we agree with this supplier about delivery times?” The agent searches, finds the contract and the later email where it was amended, and answers citing both. The value is not the summary, it is the citation: you can verify it in ten seconds.

Consolidating what is scattered. “List every confidentiality clause we drafted in the last two years and group them by equivalent wording.” Over a hundred and sixty contracts, that is one agent turn and three days of a trainee. It is the kind of work that sits on a firm’s to-do list for eighteen months because nobody has the free week.

Keeping a document alive. A process manual that gets updated when something changes, with the change noted and linked to the minutes where it was decided. This sounds minor and it is what stops documentation from dying, which is the cause of death of ninety percent of knowledge bases.

Preparing the boring draft. The meeting minutes from the notes, the follow-up email, the new supplier record with the details already in the file. Nobody is going to send it unread, but starting from a draft with the right facts saves half the time.

The three things it does not do

This is where expectations are better lowered by you than by reality.

It does not replace judgment. The agent will tell you something correct and something badly inferred from an out-of-date document with exactly the same confidence. Any answer going to a client or a regulator gets reviewed by a person. If your plan depends on nobody reviewing, the plan is something else.

It does not understand what was never written. If the reason that client has a twelve percent discount lives in the commercial manager’s head and in no set of minutes, the agent will not find it. It will find the discount, not the why.

It does not fix your permissions. And here it is worth being very concrete, because this is where people get nervous. When you put an agent over your documentation, information that was hidden in a subfolder becomes searchable. That is what you wanted. It also means the salary table, if it was in there, now surfaces. Before switching anything on, somebody has to review who sees what. It is half a day of work and it avoids a very uncomfortable conversation.

What an agent needs in order to be worth anything

Three things, and only the first is technical.

A place where the documents are. It can be messy. It cannot be spread across four personal mailboxes, because the agent cannot reach those and, more to the point, neither can whoever replaces that person.

A clear boundary of what it can touch. In Ofivia the agent works confined to the file subtree of the department where you opened the conversation. Sibling departments are not mounted into its working environment: the agent knows they exist because we pass it a summary, but it cannot open their files. That is the real boundary, and we say it that way rather than promising a per-file permission we do not enforce on writes today.

Somebody who corrects things. When the agent answers badly because the document was wrong, the work is not tuning the agent: it is fixing the document. The companies where this works are the ones that understood every bad answer is a finding about their documentation.

What they cost

Two costs, and they behave differently.

The software runs between fifty and a hundred and fifty dollars a month for a whole company, when billing is per company. When billing is per seat, between ten and thirty dollars per person per month, which for forty people turns into a considerably larger figure.

The model usage is what you pay for the agent’s real work. Over 119 measured turns on the installation we write from, the average came out at 0.19 dollars per complete turn. A team of forty where five use the agent daily runs between 1,500 and 3,500 turns a month, so between 285 and 665 dollars.

Two questions worth asking any vendor before signing. First: which operations consume nothing. Indexing documents and searching them should cost nothing, because they do not call a paid model. If you are charged to index, you will end up deciding what to upload based on price. Second: what exactly happens when the month’s credit runs out. In Ofivia usage is recorded and shown by day, by process and by person, and you see exactly what was spent. What we do not do today is cut off on our own at the ceiling: you see it in the dashboard, we tell you, and an owner can request more credit from that same screen.

How to evaluate it in two weeks without committing

The mistake is evaluating with demo questions. Do it like this.

Week one. Pick one department and upload what it has, without cleaning it. Write ten real questions somebody had to answer last month, with the correct answer noted separately. Have them asked by someone who was not involved in the upload.

Week two. Compare. Count how many it got right, how many wrong, and how many it said it did not know. The third category is the good sign: an agent that says it cannot find something is more useful than one that invents.

Seven right, two do-not-knows and one wrong is a good rollout. If it comes out four right and six wrong, look at those six source documents before blaming the software. In the evaluations we have sat through, between half and two thirds of the failures came from documentation that said something different from what the company believed it had written down.

That alone justifies the two weeks.

If you want to see how that working boundary is built, which tools the agent has and what gets recorded about everything it runs, it is explained in detail on the agent page.