Short answer: an assistant over Business Central starts from $3,000, an agent from $6,000, a solution around a specific business process from $10,000, after a discovery from $750. But the main variable is not those figures — it is the state of your setup and whether your calculations are defined in writing. The same scenario can differ by a factor of two between two companies for that reason alone.
Ranges and what they mean
| What | From | Timeline | Includes |
|---|---|---|---|
| AI Discovery | $750 | 1–2 weeks | Use cases, data, permissions, architecture, estimate |
| AI assistant | $3,000 | 2–4 weeks | Read-only over a defined data set |
| AI agent | $6,000 | 4–8 weeks | Approved actions, permissions, approvals, logging |
| Process-specific solution | $10,000 | Varies | Multiple sources, integrations beyond BC |
"From" is not a marketing hedge here. An assistant covering three question types and one covering twenty are different pieces of work under the same name. That is why any serious estimate comes after discovery rather than before it.
Why an agent costs twice an assistant. Not because the code is much bigger. Because of testing. Verifying that a system answers correctly means comparing output against figures computed by hand. Verifying that it refuses what it should refuse means inventing the scenarios where it might fail and walking through each. The second takes longer and decides whether this is production-ready.
What the cost is made of
Effort distribution on a typical first project. Useful because it shows where the money actually goes.
| Stage | Share | What it really is |
|---|---|---|
| Scenario description and definitions | ~25% | Conversations with your people, not code |
| Designing operations | ~20% | Translating questions into the BC data model |
| Development | ~25% | Integration, orchestrator, logging |
| Testing | ~20% | Mostly verifying refusals |
| Deployment and documentation | ~10% | Environments, permissions, written scope |
Half the project is the first and fourth rows, and neither is programming. This is why "we'll do it in a week" estimates do not hold: a week is roughly the third row.
For budgeting purposes: the cost depends less on how much code is needed and more on how clearly you can describe what you want.
Running costs
The part most often missing from a budget. One-off build cost is not total cost of ownership.
Model usage. You pay the provider for the volume of text processed. It depends on the number of requests, the volume of data in each, and how economically the operations were designed — an operation returning the three rows you need is cheaper than one returning a table for the model to filter. Design affects the invoice as well as security. Work it out from the provider's current pricing multiplied by expected volume; those prices change, so use current ones.
Hosting. The integration layer has to run somewhere. For most scenarios a small figure, noticeable only against very small projects.
Support. A check after each major Business Central release — twice a year, a few hours if the solution is built on the standard API. Plus periodic review of the permitted operation list.
Business Central licensing. The integration needs an identity like any user. The exact configuration depends on the authentication model chosen during architecture.
What makes a project more expensive
In descending order of impact.
- Undefined terms. If "overdue", "active customer" or "margin" are calculated differently in different departments, the project waits until the business decides. Not technical work, and not something a vendor can resolve — but the clock runs.
- Inconsistent dimensions. Analysis over data where the fill-in rules changed mid-period either fails or is wrong. Sometimes the setup has to be fixed first — see dimensions in Business Central.
- Data outside Business Central. Every additional system is a separate integration with its own authentication and error handling.
- Role-based data separation. Delegated access is materially harder than a service account. Decide early: switching later is a rebuild.
- Write access. Every data-changing operation brings retry handling, partial-failure handling and separate testing.
- A broad first scope. Trying to cover three scenarios at once extends testing by more than three times, because they start interacting.
What makes it cheaper
- A narrow first scenario. One process, running weekly, with a visible result.
- Read-only. An assistant delivers most of the value at half the cost of an agent.
- Definitions ready. If margin and overdue are already written down, you save a quarter of the project.
- A tidy setup. Payment terms filled in, payments applied to documents, item cards current.
- The second and later scenarios. Materially cheaper than the first: the integration layer, authentication and logging already exist.
That last point matters for annual budgeting. The first agent is an investment in infrastructure; the second and third reuse it.
Costs nobody warns you about
Your people's time. The process owner is needed for several sessions, not one hour. Whoever agrees the calculation definitions likewise. Real cost, just not on the vendor's invoice.
Cleaning up data. If discovery shows dimensions are inconsistent or payments are not applied, that gets fixed first. Separate work with a separate cost, and an honest vendor mentions it before the build rather than after.
A second round after launch. A month in, it becomes clear that some scenarios were unnecessary and two unplanned ones are needed. Normal — budget it as extension rather than treating it as an estimating error.
Publishing data through AL. If the fields you need live in your extensions and are not exposed through the standard API, API pages need publishing. Small work, but it should surface during discovery.
How to compare quotes
Comparing headline prices is pointless: the same figure can cover very different scope. Nine questions that make quotes comparable:
- What is the full list of operations available to the system? Show it.
- Does the integration run as a service account or with delegated access?
- What permission set, and why that one?
- What happens if it is asked to do something outside the list?
- What does the system do when there is no data — answer, or refuse?
- How are retries handled when the connection drops?
- Which model provider, deployed where, under what contract?
- Who pays for model usage and hosting, and roughly how much?
- What does support include, and what happens at a Business Central release?
Questions 4, 5 and 6 are the most informative. Vague answers there usually explain a price difference: that work is not in the quote.
How to reduce the number
Four actions that genuinely lower cost, all of them yours.
Fix the definitions before talking to anyone. What "overdue" means, from which date, which amounts are excluded; which of the three possible margin definitions is yours. That is a quarter of the project, and only your team can do it.
Write down ten real questions. Not hypothetical ones — the ones that currently arrive in somebody's inbox as a request for a report. Under each, note what data the answer is made of.
Switch on what is already licensed. A month of using the built-in capabilities produces a factual list of what they do not cover — a specification written for free, from real cases.
Start with read-only. An assistant costs half an agent and delivers most of the value: the time goes into gathering data, not into the final click.
A company arriving with a list of questions and written definitions gets both a lower estimate and a shorter project. Discovery from $750 closes exactly that uncertainty, and the document is yours whether or not you continue.
Frequently asked questions
Why can't you quote a firm price up front?
Because identically named solutions cover different scope, and because the main variable is the state of your setup, which cannot be assessed without access to the system. An assistant over three question types and one over twenty are different pieces of work.
How much does model usage cost per month?
It depends on the number of requests and the data volume in each. Work it out from the provider's current pricing multiplied by expected volume. Operation design affects this materially: returning the rows you need is cheaper than returning a table.
Can we start small and expand?
Yes, and it is the cheapest route. The first narrow scenario builds the infrastructure — integration layer, authentication, logging — and the second and third reuse it at significantly lower cost.
Which is cheaper, an assistant or an agent?
An assistant is roughly half the cost and delivers most of the value, because the time saving is in gathering data rather than in the final action. An agent is justified where the prepared work also has to be executed.
What if our data is in poor shape?
Then it gets tidied first, as separate work with a separate cost. We say so at discovery rather than after the build starts — analysis on top of inconsistent data costs more and produces worse results.
Is support included?
Build and support are quoted separately. Support typically covers checking after major Business Central releases twice a year, error review, and periodic review of the permitted operation list.