AI agents
A mid-sized company can run the way the big ones do. In a large company, the daily routine that takes judgement is not sitting on one person's desk: it has been written down, split between what a system can settle and what a person has to, and the people are left with the cases that are genuinely hard. Your company already has the judgement. What it does not have is anywhere to put it.
Because there is a process in your company that repeats every day, that takes judgement, and that was never automated for exactly that reason: check, cross-reference, decide, act. Somebody good does it, and it eats their morning.
This is not an assistant that answers your questions. It is a process worker that runs on its own against your ERP and your documents, acts below the limit you set, stops and asks for approval above it, and leaves every step on record with the reason for it. Nobody is asked to trust a black box.
It starts with five weeks dry: the agent proposes and executes nothing, and at the end you see in how many cases it would have been right. If that number does not suit you, I tell you so in writing and that is where it ends.
USD 2,400 for the five-week dry run. From USD 8,900 for the agent in production, ten to fourteen weeks, one project at a time. USD 590 a month to keep it running.
It is not that you are short of people. It is that the people who know are busy doing what anyone could do if they had everything in front of them at the same time. That is the gap between a mid-sized company and a large one, and it is not a gap in talent.
Go through the receivables customer by customer, match each balance against what that customer's contract says, decide who gets written to today and in what tone, write it, log it, and come back tomorrow. No part of it is difficult. All of it together is half a working day.
The automatic reminders that send everyone the same message ended up switched off, because you do not write to your largest account the way you write to the one that always pays late, and the system could not tell the difference.
It lives in the head of whoever has been doing it for years. That is why it never became a set of rules, and why it falls over the week that person is on holiday.
Every hour that person spends on the routine is an hour not spent on the work that genuinely needs their judgement. That cost appears in no account, which is why it never gets fixed.
Launch offer: the first three to reach step 2 get three months of operation and custody at no cost. It is not a discount: the price does not move, and once the three slots are gone it stays where it is. Do the arithmetic before you talk to me. Count how many times a day that process repeats and how long each one takes. If it does not come to an hour a day, an agent is not worth it for you, and I will say so.
Three different things are sold under the same word, and they are worth separating before paying for any of them.
It runs a script written by hand. It works perfectly for as long as the world behaves the way the script assumed, and it stops the moment anything falls outside. It is rigid by design, and that rigidity is also its guarantee.
It automates the steps
It reads your documents and answers you with the citation of where it took the answer from. It saves you the looking. The decision and the work are still yours. That is what the other page of this site sells.
It automates the looking
It is a language model with tools and with permission to use them. It looks at the state of things, decides the next step, carries it out, observes what happened and corrects. If the expected route fails, it looks for another instead of getting stuck.
It automates the outcome
Ordinary automation automated the steps; an agent automates the outcome. That is why it is worth more, and that is why it has to be given a brake.
Which is why the price
This is the section almost no proposal contains, and it is the one that decides whether you can sleep. It comes before everything else here for a reason: the answer to "how do I know what it will do" has to be a design, not a reassurance.
What it is able to do is settled beforehand, with tools rather than with instructions. Asking the agent to be careful is not enough. If it must not delete anything, it is not handed the ability to delete: it is given read-only access, and then it cannot delete however wrong it gets about everything else. Everything that changes something is handed over separately, listed in writing, and can be revoked in a minute.
Where it stops. You draw the line. Below it, the agent acts and tells you. Above it, it prepares everything, leaves it ready and waits for approval from the person you name. The line can be an amount, a type of customer, a type of document, or all of those at once, and it can be moved whenever you like without touching anything else.
A second agent that reviews it. Before anything outside the routine is carried out, another agent, holding your company's rules, reviews what the first one proposes and can stop it. Having a third party check the work is more reliable than asking the same one to police itself.
And one thing that does not change. The agent has no legal personality and answers for nothing. Responsibility stays with a person, always. That is why everything it does is on record with its reason, and why you draw the line in writing before the agent touches anything.
If your company has no written rule yet about what may be handed to a model and who approves a new use, that conversation comes before this one. It is a product of its own: an AI use policy your people sign.
One question settles this before any other: how much room to decide are you going to leave it? If the answer is none, what you need is not an agent, it is a program, and it costs a great deal less. I will tell you if that is your case.
It repeats many times and each time is slightly different. There is a judgement that a person applies well but never managed to write down as a rule. The result can be verified: you can see whether it came out right or wrong. And the exceptional cases can be left to a person without the rest falling over.
The judgement is subjective and personal, of the "I like it but not like that" kind. Or what decides the matter lives in people's heads and in conversations rather than in the system: negotiating a large contract, granting credit to somebody there is no information about, handling a crisis.
An error that costs more than a correct answer saves. If getting it wrong once means a fine or a legal problem, that process does not get automated: it gets supported, and a person signs the decision.
That last one is not a caution of mine: it is what the engineering literature on agents recommends, and it is the line between a project that goes well and one that goes expensively.
Take a concrete case: a company that imports and distributes, with Odoo running, receivables spread across several hundred customers, and one person who spends the morning chasing them. The company is made up; the shape of the work is not.
Before any technology is touched, somebody sits with the person who does it today and writes down what they do: what they look at first, what makes them change their mind, who never gets written to and why. Out of that come the rules that can actually be written, and what is left over is marked as judgement. That separation is the whole project.
Exactly what it will be able to do is listed: read the overdue receivables from Odoo, read the customer's contract, read the payment history, draft a message, log it on the customer's record, send it. Each one is classified as either reading only or changing something. The ones that change something are few, they are enumerated, and they are reversible.
Up to such an amount, and for customers in good standing, the agent writes and sends. Above that amount, or if the customer is in dispute, or if the contract carries a special clause, it prepares the message and leaves it for approval. That limit is agreed with you and written into the contract.
The agent runs every day against real data and sends nothing. It proposes, and each proposal is compared against what the person in charge would have done. At the end there is a number: in how many cases it agreed, in how many it was wrong, and in how many it asked for help. That number is what the decision to carry on is made with.
It starts with the stretch that is easiest and cheapest to get wrong, and widens when the numbers support it. The whole process is not switched on the first day.
Every run leaves its trace. How many times it called a person, how many times a tool failed on it, how long it took and what it cost are all measured. That dashboard is yours, and the monthly review comes out of it.
All of this runs wherever you say: on your server, in your cloud or in mine. And the interviews are the part that cannot be skipped. If the person who does the process today is not in the room, what gets built is somebody's idea of the process rather than the process.
What changes a person's working day is not that the machine is clever. It is that what it did can be reviewed in thirty seconds and argued with.
What it observed: "Distribuidora del Norte, two overdue invoices for USD 4,180, the oldest of them 38 days old. Framework contract of 14 March 2024, clause seven: sixty-day terms. Last twelve months: pays on average at 52 days, no defaults. No open disputes."
What it decided: "A reminder is what applies here, not a demand. The customer pays late but pays, and the amount is below the approval limit. Cordial tone, reference to clause seven, no mention of interest."
What it did: drafted the email, sent it to the collections contact on record, and logged it on the customer's record in Odoo with the full text.
What it left for a person: "Comercializadora del Sur, USD 11,900, 74 days. Over the approval limit, and the contract has an early-payment discount clause I cannot read confidently in this case. Message prepared, awaiting approval."
Nobody had to open Odoo, or find the contract, or remember who pays how. And what needed judgement reached the right person with everything ready to decide in a minute. That same contract, when what is needed is to ask it something rather than act on it, is answered by the assistant: the two pages are the same case from its two sides.
An agent is only good for what it can reach. A company with no system has an earlier problem: there is nothing to reach in any ordered way, and it would have to be built first.
A company with Odoo, with SAP or with whatever else has already solved that. The agent's tools are the operations your ERP already performs and your people already use: look up a balance, log a note, create a draft, move a status. Nothing has to be invented; the right amount has to be exposed, and no more.
And the permissions are already defined. Your ERP already knows what each profile may see and do, and the agent comes in through that door, with a user of its own, its own permission profile and its own trail in the system's audit log. Everything it did is visible in the same place where you see what anybody in the company did.
There is an important difference with how this is sold elsewhere. Letting the model write free-form queries against your database looks excellent in a demo and cannot be audited: it invents field names, ties itself in knots over the joins, and returns something different each time. Here it is the other way round. A short list of operations already written and reviewed with you, each one with its parameters. The agent picks which to use and with which values; the operation that runs is always the same one, and it is logged.
Five questions separate a serious proposal from a demo.
Ask for the list of operations, split between the ones that only read and the ones that change something. If that list does not exist, nobody knows what the worst possible case is. And if the answer to "what stops it deleting something" is that it has been instructed not to, the correct answer was that it has no way to.
Ask to see the trace of a complete run: what it looked at, what it decided, which tool it called, with what data and how long it took. An agent that only shows the result cannot be audited, and what cannot be audited cannot be corrected.
It is the best indicator there is and almost nobody offers it. If it calls too often, it is not saving you work. If it never calls, it is deciding things it should not be deciding. Ask for the number from the trial onwards, and ask to see it month by month.
An agent can act with complete confidence on a wrong figure that a system handed it. Ask what it verifies before acting and in which cases it stops because it is unsure. If the answer is that it trusts whatever arrives, the first bad figure becomes a bad action.
A person, always. Ask who that is, what they can undo, how quickly, and how they find out. If this is not written down before anything is switched on, it is undefined.
You do not buy the automated process up front. You buy a trial that tells you with numbers whether the agent would have been right, without its having touched anything.
Includes: the process written down step by step with your people, the list of operations split between reading and acting, the approval limit agreed, the agent running daily against real data without executing anything, and a report on in how many cases it agreed with your people's judgement, in how many it was wrong, in how many it asked for help and what the rest would cost. It covers one process and up to six read-only operations; beyond that is quoted. If you decide to go on, it is credited against step 2.
One project at a time: if the slot is taken, I tell you the date of the next one before you invest any time in this. Includes: the agreed operations connected to your ERP with a user and permissions of its own, the brake by amount and by exception, the reviewing agent, the full trace of every run, the dashboard of indicators, the staged go-live and training for whoever is going to supervise it. The price depends on how many operations there are and on whether it runs on your premises or in the cloud.
Includes: watching that it runs, how long it takes and what it costs, a monthly spending cap with a warning on reaching eighty per cent of it, the monthly review of the cases where it asked for help and the ones it resolved badly, the adjustment of the approval limit and the addition of new rules. No minimum term: cancel whenever you like. A year paid in advance is USD 5,900, ten months instead of twelve.
Prices in United States dollars; under my Ecuadorian tax regime, RIMPE Negocio Popular, my services are invoiced with VAT at 0 %, so the figure is the final one. The cost of the model and of the infrastructure is invoiced separately, at cost and with the provider's own detail, because it depends on how much it is used. An agent costs more than an assistant because it acts: what you pay extra for is the brake, the record and the review, not the model. Guarantee on step 1: before we start, we agree in writing what proportion of cases the agent has to match your people's judgement in for carrying on to be worth it. That proportion is measured over every case in the five weeks, not over a sample picked afterwards, and the report publishes it case by case. If it is not reached, there is no step 2 and half of step 1 is refunded; the other half pays for the work of finding out, which is an answer too. For it to apply, two things are needed on your side: the interviews are held with whoever runs the process today, not with whoever supervises it, and the same person judges the cases across the five weeks, as they happen rather than all at the end. Full conditions in the proposal and in the refund policy.
No, and anyone promising you that has never put one into operation. What goes is the routine: looking, cross-referencing, drafting and logging. What stays is deciding the cases that are not obvious, which are the ones that genuinely need that person. In practice, the indicator that the agent is working is how many times a month it calls somebody, not that it stops calling.
It will do something wrong, the way a person does. The difference is that the worst case is defined beforehand: the agent can only perform the operations it was handed, and every one of those that changes something is reversible and on record. An agent that was never given a way to delete does not delete, even when it is wrong.
Yes, in a minute and without touching code. The user is revoked in the ERP and the agent is left with no hands. The process goes back to being done the way it was, because nothing that was there was taken away.
Almost certainly not, and it is the first thing you will be sold. Several agents cost more, take longer and, when something goes wrong, make working out which one failed much harder. They are justified in three cases: when there is data that has to stay walled off, when there are so many operations that the agent gets confused choosing, and when different teams maintain different parts. Outside those, one agent, done properly.
Longer than an assistant, and it does not matter. An agent goes round the loop of looking, deciding and acting several times, and each round adds up. Since it runs on its own and nobody is sitting in front of a screen waiting, what gets watched is different: that it finishes, that it does not go round in circles, and that the cost per run does not run away. That is measured and reported.
More use than rigid automation, because the judgement is explained to it in ordinary language and there is no script to reprogram. But whatever changes has to be told to it, and that is the work of step 3. An agent nobody tells that the policy changed will go on applying the old one with complete confidence.
No. What is needed is that the process lives in some system that can be entered with permissions. With Odoo I also know the ground and the operations already exist. With another ERP it connects through its programming interface, with a user of its own and permissions of its own.
The assistant answers questions about your documents and your ERP, citing where it took the answer from. The agent does the work. Plenty of companies start with the assistant, see which questions keep repeating, and discover from that which process is worth automating. It is a good order, and there is no need to buy both.
Yes. The work is remote, and what the agent runs against is your system, wherever it is hosted. Quito is on UTC-5 all year round, so the working day overlaps with North America almost entirely and a European afternoon is a Quito morning. The interviews in step 1 are done by video call with whoever runs the process, and the reports and the dashboard are in English.
11 /
Tell me which process repeats every day in your company and who does it today. I reply with the proposal for the dry run within one business day.
contacto@andresbrinez.comSix years building systems for companies in three countries: Colombia, Ecuador and Australia · Member of the Odoo Community Association