Start with the budget band
An AI automation quote usually doubles because the workflow was described as a task, while the real system contains permissions, exceptions and several business tools. The model rarely caused the jump.
Our published project intake uses four practical bands: under €5k, €5k-€15k, €15k-€40k, and €40k+. For automation work, we read them like this:
- Under €5k: discovery, a workflow map or a constrained feasibility test. It should not be sold as a production system.
- €5k-€15k: a narrow pilot with one clear input, one useful output and limited access to business data.
- €15k-€40k: a production workflow with authentication, integrations, evaluations, monitoring and a real interface for operators.
- €40k+: several connected workflows, sensitive actions, difficult legacy systems or a product that must serve multiple teams and markets.
The uncomfortable answer is that a dependable automation is often a small software product. Budgeting for a clever prompt produces a clever demo.
Discovery is paid risk removal
Discovery identifies the event that starts the workflow, the systems it reads, the actions it may take and the person responsible when it stops. We also need representative examples, failure cases and a baseline for the manual process.
This work looks expensive when the brief says only automate support or qualify leads. It becomes cheap when it reveals that the CRM cannot expose the required field, that customer consent is missing or that two departments use different definitions of a qualified lead.
A serious discovery phase ends with a scoped workflow, an access plan, an evaluation set and a decision about what remains human. Without those, the implementation quote contains contingency for unanswered questions.
Integration costs more than intelligence
Connecting a model to a clean API is straightforward. Connecting it safely to a CRM, inbox, document store and internal database is not. Each integration needs authentication, field mapping, retries, duplicate protection, logging and a recovery path.
The quote rises when data is trapped in PDFs, exports or undocumented legacy software. It rises again when the automation can send messages, alter records, issue refunds or expose personal data. Those actions need narrower permissions, approval states and an audit trail.
The model bill may still be modest. OpenAI currently lists GPT-5.6 Luna at $0.20 per million input tokens and $1.20 per million output tokens. That does not make the whole system cheap. Search, storage, external APIs, observability and human review sit outside that token price.
What makes a quote double
A quote can change sharply when a read-only assistant becomes an agent that writes to production systems. The same happens when a single-language workflow becomes multilingual, a clean knowledge base becomes a document cleanup project, or normal office traffic becomes a customer-facing service with unpredictable peaks.
Exceptions are the other multiplier. The happy path may take an afternoon. Handling missing records, conflicting data, expired credentials, provider outages and partial writes is the work that makes the automation trustworthy.
Then comes the running cost nobody quotes: reviewing failures, updating prompts and evaluations, responding to API changes, reconciling incorrect actions and deciding who owns the system after launch. A system without an operator slowly becomes an unmonitored liability.
What to price this week
Write down the trigger, required inputs, permitted actions, approval points and failure owner. Collect real examples, including the awkward ones. List every system involved and confirm that usable API access exists.
Then request separate prices for discovery, implementation, third-party usage and ongoing operation. Ask what assumption would change the quote most. If the answer is vague, the scope is still vague. If the running cost is absent, it has not disappeared. It has merely been left for the buyer to discover later.