Insights — Artificial intelligence
A document assistant does not cost what its demonstration costs.
A proof of concept takes three days and costs about as much as an internship. The tool your teams open every morning costs something else: an index that ages, rights that change, answers nobody measures. Here are the seven cost lines missing from every quote — and what they weigh over three years.
01 — The observation
The budget of a RAG is an operating cost disguised as an IT project.
A document assistant quote looks like a software quote: a fixed implementation fee, an interface, an acceptance test. That is a misreading. A RAG is not software you install, it is a chain you maintain: documents change, models evolve, user rights move, and answer accuracy degrades without warning.
As a result, year one is written in the quote and the next two are not. That is exactly where the difference lies between an assistant used every day and a demonstration abandoned after six months — and that is also where overruns hide.
01The seven cost lines absent from every initial quote
None of these lines is exotic: they exist in every document assistant project, including the ones we take over after an abandonment. They are simply missing from quotes, because they are paid later — in person-days, monitoring and rebuild — rather than in licence fees.
- Continuous re-indexing — your procedures, price lists and contracts change every month. A frozen index answers confidently from outdated documents: the worst of both worlds. Plan for monthly incremental re-indexing, never a full rebuild.
- Source document quality — scans, spreadsheets, composite PDFs, procedures written by five different people. This is the most thankless and most decisive line: no model compensates for a badly built index.
- Access governance — rights must apply to retrieval itself, not just to the home screen. An assistant answering a salesperson with a board-level document is not a theoretical leak, it is an incident.
- Continuous evaluation — a set of real questions, the expected answer beside each one, the share of correct answers, the share of inventions, and correct refusals when the information simply does not exist. Without that measurement, nobody knows whether the tool is improving or decaying.
- Operations and on-call — chain monitoring, quotas, logging, supplier updates, incidents. It is little work per month, never zero.
- Supplier drift — retired models, changed pricing, deprecated APIs. This line costs nothing as long as a multi-model gateway lets you switch engines without touching the application.
- Reversibility — who holds the index, the test questions and the logs? If the answer is 'the supplier', changing supplier means rebuilding everything. It is the highest hidden cost of all, and it surfaces the day you need it most.
02Total cost, year by year
Over three years, the full cost of an internal document assistant typically lands between one and a half and four times the amount of the initial quote. The gap has nothing to do with the number of users: it comes from how often your documents change and how strict your access requirements are.
Year one is dominated by implementation — document preparation, development, acceptance, initial evaluation. Years two and three are dominated by operations, on a decreasing rhythm: the first months of an assistant are always the most expensive to run, because that is when the quality of the original documents shows itself.
- Year 1 — document preparation, go-live, initial evaluation set, team onboarding: the heaviest part of the project
- Year 2 — incremental re-indexing, retrieval tuning, scope growth: an operating budget in the hundreds or low thousands per month depending on volume
- Year 3 — model or supplier change, rights review, extension to a second business line: the year when the gateway and proper ownership of the index make the difference
03Three decisions that lower the total cost
They are taken on day one, not on the first overrun.
- Reuse an existing PostgreSQL database with the pgvector extension rather than adding a dedicated vector store: one less service to back up, secure and monitor.
- Place a multi-model gateway in front of the application: an intermediate model for everyday questions, a more expensive one reserved for genuinely hard cases, and the ability to switch engines without rewriting the tool.
- Treat the index, the test questions and the logs as assets that belong to you, exportable at any time. That is what turns leaving a supplier into a migration instead of a rebuild.
03 — Quoted / Real
What the quote says, what is actually paid
| Before | With Zorglon |
|---|---|
| A fixed implementation fee, then nothing | A monthly operating line: index, rights, evaluation, updates |
| 'We will index once and for all' | Incremental re-indexing every month, one to three person-days |
| The most powerful model, chosen for three years | An intermediate model by default, replaceable in a day |
| Index and logs kept at the supplier | Index, test questions and logs on your side, exportable |
| Accuracy observed in daily use | A reference question set, replayed at every change |
04 — Common questions
What we are asked about the full cost
What annual budget should be planned after go-live?
On a scope of a few hundred to a few thousand documents, operations count in person-days per year, plus a monthly hosting cost in the hundreds when the infrastructure is shared. The factor that moves this amount threefold is not the model: it is how often your documents change.
Can continuous re-indexing be avoided?
It can be made incremental, automatic and dated: only modified documents are reprocessed, and the previous version stops being served. What cannot be removed is the link between index freshness and answer accuracy — a six-month-old index is a misunderstanding generator.
Why budget for reversibility?
Because leaving costs too: exporting the index, transferring the test questions, taking back the logs, reinstalling elsewhere. Prepared, it is a migration of a few days. Unprepared, it is a new project — and you no longer choose when.
Does a cheaper model degrade quality?
Rarely, when retrieval is good. In most cases we see, wrong answers do not come from the model but from the wrong passage being retrieved. Investing in index quality pays back more than paying for the most powerful model.
Do you want the total cost, not the price of a demonstration?
RAG scoping: we start from your documents, how often they change and your access constraints, then hand back a three-year budget — implementation, operations and reversibility included.