Applied AI in Europe: buy a pilot, not a promise

What is confirmed on 5 October

The European Commission published the ten Apply AI Startup Award finalists on 2 October; the news page shows an update on 5 October. The European AI Innovation Month page, updated on 14 September, announces events from 14 October to 17 November 2026. These are announcement and scheduling dates, not outcomes already observed.

For a Portuguese SME, this calendar can organise research and supplier conversations. Our interpretation: a competition helps discover candidates but does not replace a procurement assessment. Finalist status alone does not validate a product for your documents, users or decisions.

Write a task whose outcome can be accepted

Velqory recommendation: start with a bounded operation, such as drafting a classification of incoming requests. Describe the input, output and person who decides whether the work is correct. Exclude payment decisions, hiring decisions and final customer replies from the first pilot.

Do not ask only for a demonstration using documents selected by the seller. Prepare a reserved set representative of the languages and formats your team uses, including incomplete examples and cases that should be refused. Keep expected answers separate. An impressive presentation does not answer the question of how the system handles the exception that arrives the following Tuesday.

Acceptance comes before the outcome

On the page reviewed, NIST describes the AI RMF as a voluntary framework for incorporating trustworthiness into AI design, development, use and evaluation. It is neither certification nor a Portuguese obligation. Velqory proposes translating that framing into a small acceptance agreement signed before work starts.

Define mandatory fields, verifiable references and behaviour when information is insufficient. Identify errors that block approval altogether instead of diluting them in an average. Also measure review time and human corrections. If an output becomes useful only after it has been rewritten, do not count the initial draft as completed work.

Set boundaries for data and authority

Ask the supplier for written answers on retention, training use, access, subprocessors, deletion and export. Until those answers have been assessed, use synthetic or appropriately anonymised data. A pilot can have low volume and still expose important information; size does not define authorisation.

OWASP distinguishes direct instruction injection from indirect injection through external content such as files. It recommends least privilege and human approval for high-risk operations. For this pilot, Velqory's proposal is specific: documents are material for analysis, not authorisation to send messages, open new access or change records.

The matrix to include with a request for proposals

This matrix is a Velqory editorial proposal, not a competition requirement. Fill in names and concrete criteria before comparing offers; a missing answer is an unresolved question, not implicit confirmation.

Ask for evidence that can challenge the sale

Reserve part of the set for evaluation without supplier assistance. Repeat relevant cases and record version, configuration, failures and interventions. Do not change criteria after seeing the answers to make the outcome favourable. If the task changes, open a new assessment and preserve the previous one.

Include a controlled check of trust boundaries without real data or production connections. The conclusion should state which behaviours were observed and which remain untested. Even a satisfactory exercise cannot establish immunity to attacks or general security; it only supports discussion of evidence limited to the scope examined.

The useful decision for October

Choose between continuing within the same scope, correcting and repeating, or closing the pilot. Record the decision-maker, reservations and next review. The European calendar should not become an artificial procurement deadline: a conversation at an event can start an investigation without creating a commitment.

Velqory recommends first buying a verifiable answer to an operational question. Discuss scale afterwards. A pilot's value lies in revealing conditions for use and reasons to say no, not in an obligation to produce a success story.