Skip to content
Gathena
da

What an AI agent needs to do real work

What does an AI agent need to do real work in a company? A model on its own can reason, but it knows nothing about your company. Here are the seven things that turn it into a colleague, and how your people stay in charge of it.

Why the chat tool didn't do it

Assistive AI answers. Agentic AI works toward a goal.

You asked a chat tool to close the month. It explained beautifully how to do it. Nothing was booked.

A chat tool answers the question it is asked, and waits for the next one. A knowledge worker, our name for an AI agent that does real work in your systems, gets a goal instead. It breaks the goal into steps, uses tools, checks its own work, and comes back when the job is done, or with the one question it could not answer alone. A good answer is not a finished job.

Chatanswers, and stops Accountingnothing was bookedKnowledge workerworks toward the goal, uses tools, checks its own work Accountingbooked, checked, one question back

The mental map

A brilliant brain with nothing connected.

Picture a brain with Einstein's reasoning and no connection to anything.

It has never seen your records. It cannot open your accounting system. It does not know your bookkeeping rules, your standard terms, or how you speak to partners. It does not remember what was decided yesterday. It can reason. It cannot know.

It can reason.It cannot know.

The model is not the system

A model alone cannot know your company. Seven things make it a colleague.

Give the same brain what a colleague has on day one, and the task gets done.

New colleagues arrive with reasoning of their own. On day one they get a login, the right systems, the rulebook, an introduction to how things are named here, and a desk with their own files. From then on, they write down what they did.

A knowledge worker needs the same: your language, its own records, tools, access, rules, limits, and a logbook, with an operator who signs off. The parts work only together, and we build all of them.

The operator signs offa human, always outside the computerInside the computer the same brain Dictionary Own recordsTools Access Rules Limits Logbook
The operator signs offa human, always outside the computerInside the computer the same brain Dictionary Own recordsTools Access Rules Limits Logbook
The operator signs offa human, always outside the computerInside the computer the same brain Dictionary Own recordsTools Access Rules Limits Logbook
The operator signs offa human, always outside the computerInside the computer the same brain Dictionary Own recordsTools Access Rules Limits Logbook
The operator signs offa human, always outside the computerInside the computer the same brain Dictionary Own recordsTools Access Rules Limits Logbook
The operator signs offa human, always outside the computerInside the computer the same brain Dictionary Own recordsTools Access Rules Limits Logbook
The operator signs offa human, always outside the computerInside the computer the same brain Dictionary Own recordsTools Access Rules Limits Logbook

Your language

First, it has to speak your company's language.

Your people carry a shared understanding in their heads. Nobody wrote it down, because everyone knows.

A "customer" is someone with a signed contract. An "open invoice" is one unpaid past its due date. A "project" has one owner and one budget line. Ask three colleagues and they agree, because they learned it by working here. A knowledge worker has never worked here.

So we write it down once: what you call things, what each thing means in your company, and how the things connect. An invoice belongs to a supplier, a supplier belongs to a project, a project belongs to a budget line. The technical word for this is an ontology, which is the company's dictionary and map in one. Every knowledge worker you add uses the same one, so they never argue about what a word means. When the company changes, the dictionary changes with it, in one place.

belongs tobelongs tobelongs tobelongs tobelongs toBudget lineProjectone owner, one budget lineSupplierCustomersomeone with a signed contractOpen invoiceunpaid, past its due dateContract One of the seven parts · DictionaryWritten once.Your own words, never technical ones.Shared by every knowledge worker.So they never argue about what a word means.Changes in one place.When the company changes, the dictionary changes with it.

The template under your dictionary

Your dictionary starts from a template we built in our own finance work.

Your dictionary does not start from a blank page. Under it sits our template: a general dictionary and map of how knowledge workers do the work.

The template defines what a job is: what comes in, which method applies, which records it keeps, which checks it runs, when the operator signs off, and how a lesson becomes part of the method. That template is an ontology too. Your company's words fill in the blanks; the structure, the checks, and the way of working are already there. We built the template in our own finance work. Because every job starts from it, each job reuses what the last one wrote down.

Our templatehow knowledge workers do the work, in any company · built in our own finance workpart ofpart ofpart ofpart ofpart ofpart ofA jobone objective, one ownerWhat comes inThe methodthe steps, reusableRecordsChecksSign-offwhen the operator signs offA lessonreviewed, then part of the method carriesYour dictionarywhat things mean here · your words, written oncebelongs tobelongs tobelongs tobelongs tobelongs toBudget lineProjectSupplierCustomerOpen invoiceContract

Access and guardrails

It gets access to what its work needs, and no more.

You control access to your systems, through your IT, in-house or your IT partner, and you own the cloud project where the knowledge worker's tools and data live.

Access
Each knowledge worker gets access only to what its job needs. The finance knowledge worker can read the accounting system and propose entries. It cannot send email. The correspondence knowledge worker can draft email. It cannot touch the ledger. Access can be taken back at any time. Every use of it is written to the logbook.
Rules and guardrails
Rules tell it what to do: apply your bookkeeping rules, use your standard terms, write in your tone. Guardrails block what it must not do: no payments, no deleting records, nothing sent outside without sign-off. A rule is something the model follows, and it can be argued with. A guardrail is enforced outside the model: without access there is no payment, and the pipelines are read-only in code.
You control access Access for one job. Reversible. Rules: what to dorules, terms, tone · can be argued with Guardrails: where it must not gono payments, no deleting records, nothing outside without sign-off · cannot be passed and at the end, your sign-off

Where you sign off

Three zones decide when the operator steps in.

Where it acts alone. Where it asks first. Where the operator signs off.

The operator is always in the loop when money leaves the company, when something goes to an outside party, when a commitment is made, or when a rule is unclear. The operator decides, and nothing goes ahead until the operator has decided. Every decision is written to the logbook.

Every answer shows its source: where it found it, file and page or table and row. When there is no source, it says so instead of guessing.

Acts alone Asks first The operator signs offThe operator is always in the loop whenmoney leaves the companysomething goes to an outside partya commitment is madea rule is unclear Everything is written to the logbook.What access was used, what was proposed, who decided, and on which sources.

You stay accountable for what your knowledge workers do, and we help your people build the AI literacy to carry that responsibility with confidence.

Several knowledge workers

Several knowledge workers, no crossed wires.

Finance, legal, and correspondence can all deal with the same partner in the same week.

Three things keep them from confusing each other. They share one dictionary, so a word means the same thing to all of them. Each has its own records, so one cannot overwrite another's work, and one client's records never mix with another's. Every handover is written in the logbook, so a draft stays a draft until the right person has signed it off, and the next knowledge worker knows what was decided and what is still open.

They share methods: a better way to check an invoice is available to all of them the next day. They do not share records; each keeps its own.

One dictionarya word means the same to all of them Financeits own records Legalits own records Correspondenceits own records a draft, written in the logbook, not yet signed off One logbook everyone writes into. One library of methods everyone shares. Methods are shared; records stay where they belong.