Guide · Consultant, startup or AI agency

How to design and deliver AI systems for clients

A deliverable AI project connects the client’s problem to an observable result, accountable roles and the conditions under which the system will keep working after the demo.

Consulting, product and integration follow different business models, but share one risk: promising a capability without making data, dependencies, evidence, operating cost and delivery boundaries legible.

Editorial noteDraft generated by the system on 14 August 2026. Not yet reviewed.

From context to usable evidenceFive relations to preserve
  1. ProblemDescribe the work or decision to improve, not the technology to sell.
  2. EvidenceDefine data, baseline, failure cases and acceptance criteria.
  3. DependenciesExpose models, APIs, people, permissions, costs and suppliers.
Open the mapPrepare a test

Starting question

A demo asks whether it can work; delivery asks who will keep it working.

A proof may rely on curated data, temporary access and hidden manual intervention. Before turning it into an offer, separate what has been demonstrated, what depends on the supplier team and what requires the client’s systems, people or authorisation.

A commercial proposition becomes more credible when it states the result, unit of work, service level, shared responsibilities and exit conditions.

Operational relations

Five steps that must remain connected.

01

Problem

Describe the work or decision to improve, not the technology to sell.

02

Evidence

Define data, baseline, failure cases and acceptance criteria.

03

Dependencies

Expose models, APIs, people, permissions, costs and suppliers.

04

Delivery

Assign ownership of configuration, data, documentation and components.

05

Continuity

Assign monitoring, correction, updates, incidents and exit.

Boundaries and responsibility

Promised autonomy must match demonstrated autonomy.

A workflow that needs frequent manual repair is not an autonomous service. A prototype using personal accounts is not yet an integration. An initial price does not describe model, support, observability and maintenance costs.

Delivery should distinguish generated, installed, connected and active. These states change responsibility, security and the value of the offer.

Practical object

AI Delivery and Integration Brief

Complete the AI Delivery and Integration Brief. Each field exposes a relationship to verify before extending the system.

01

Client and work

Who uses the result, in which activity?

02

Outcome

What observable change is being purchased?

03

Evidence

Which data and criteria accept or reject it?

04

Dependencies

What remains outside the solution?

05

Responsibility

What belongs to supplier, client and third parties?

06

Ownership

Who owns data, configuration and artefacts?

07

Operation

Who monitors cost, quality, incidents and versions?

08

Exit

How are data exported and the service stopped?

First test

A short test should produce knowledge, not merely an output.

  1. 01
    Interview the work

    Observe tasks, exceptions and decisions before proposing a solution.

  2. 02
    Write the baseline

    Describe current quality, time, cost and problems without inventing figures.

  3. 03
    Build the proof

    Use representative cases, including errors and stop conditions.

  4. 04
    Simulate operation

    Test access, logs, correction, cost and responsibility beyond the demo.

  5. 05
    Deliver the boundary

    Document inclusions, exclusions, dependencies, ownership and handover.

Public sources

References for verification and further work.

These sources support initial design. Legal, professional, ethical and organisational requirements depend on the case and the responsible functions.

Frequently asked questions

Consultant, startup or AI agency: questions to clarify before the project.

How should an AI project for a client be scoped?

Connect one activity to an outcome, sources, acceptance criteria and responsibility. Unproven integrations and future functions remain stated possibilities, not part of delivery.

What should a proof of concept demonstrate?

It should show behaviour on representative cases, including failures, while exposing manual intervention, dependencies and excluded costs.

Who owns prompts and configurations?

That depends on the agreement and objects involved. Client data, project-specific configuration, pre-existing components and third-party services should be distinguished.

How can vendor lock-in be reduced?

Document formats, dependencies, export procedures, realistic alternatives and exit conditions before the system becomes critical.

When is a managed phase needed?

When quality, incidents, updates or integrations still require an operational owner. Describe the service level without calling it autonomy.