AI consulting should leave your team with clear decisions, tested software and the information needed to run it. The deliverables depend on what you commission: an assessment may end with a prioritised plan, while an implementation should include integration, evaluation results and an operational handover.
Consider a support team that copies customer emails into an AI tool, then pastes the draft answers back into its service portal. An integrated system could retrieve the relevant account records, prepare a reply inside the portal and let the agent review it there. Getting that workflow to work reliably involves data access, software development and operating procedures as well as the model.
Discovery: identify the process and measure the problem
The first deliverable is a description of the work to improve. Record who does it, how often it happens, how long it takes and what happens when it goes wrong. Check whether changing the process or connecting two systems would solve the problem before choosing AI.
Recurring emails, document searches, support triage and weekly reporting are useful places to look. Staff can usually show you where they repeat work, retype information or wait for an answer. Include those workarounds in the process map.
The AXA/Sotomo 2025 survey of 300 Swiss SMEs found growing use of AI for correspondence, translation, work-step optimisation and data analysis. Among respondents already using AI, 57% reported time savings. That supports investigating these tasks, but your project still needs its own baseline.
Measure recurring work over a full week
A ten-minute task repeated thirty times takes five hours. Use the actual frequency and handling time from your team, then account for the review and exception handling that would remain after automation. This gives you a more useful estimate than counting the time spent on one demonstration.
Assess use cases before choosing technology
An assessment should explain why a use case is worth testing and what could prevent it from working. These four questions give you an initial comparison. Error consequences and data access also need explicit review before a candidate is approved.
| Question | Evidence to collect | How it informs the decision |
|---|---|---|
| How often does it occur? | Cases per day or week | Shows whether a small improvement could matter at scale |
| How much manual effort does it take? | Handling time, review and rework | Helps estimate the effort that could be recovered |
| Can the system access suitable data? | Source quality, freshness and permissions | Exposes data preparation and access work |
| What business result should improve? | Cost, speed, error rate or service quality | Provides a measure for evaluating the pilot |
Frequency alone does not make a task suitable for AI. A frequent, rules-based transfer may only need an API connection; a less frequent task may still be valuable if it prevents expensive errors. The assessment should document the alternatives and recommend which uncertainty to test first. The AI use case guide covers prioritisation in more detail.
Implementation: connect AI to company systems
An assistant answering account questions needs the right customer records. A document search needs current files and their access permissions. A workflow that updates the CRM needs validation, an audit trail and a way to recover from a failed write. These are concrete integration requirements to include in the implementation scope.
The Model Context Protocol (MCP) provides a standard way to connect AI applications to external data and tools. It can simplify connector development, but it does not decide who should be allowed to read a record or approve a change. The application and connected services still need to enforce those controls.
Specify how these components work together, how they are tested and who maintains them.
These dependencies also appear in EY's 2026 survey of 604 respondents in Switzerland. Data quality and silos were the most commonly reported implementation obstacle, at 20%, followed by security and data protection at 19%. Ask your consultant to identify the specific source systems and access constraints behind the proposed design.
Architecture: choose cloud or private AI
An architecture recommendation should compare the options that fit the task. These might include a managed AI service, a retrieval system using approved company documents, a private cloud deployment or a self-hosted model. Retrieval-augmented generation (RAG) supplies relevant source material to a model; it can be used with several hosting options.
Compare output quality, data handling, latency, operating cost, integration effort and vendor dependence. Document where prompts, retrieved records, responses and logs will be processed and stored. For sensitive information, the decision also needs to cover access controls, retention and the provider's contractual terms. Hosting location alone does not settle those questions.
Handover: agree the deliverables your team will receive
For an implementation engagement, put the expected handover in the scope before development starts. Depending on the project, it should include:
- The process map, baseline measures, acceptance criteria and decisions made during discovery
- The agreed source code, configuration, integration documentation and deployment instructions
- Representative evaluation cases, results and known limitations
- Monitoring for service health, output quality, usage and cost
- Operating instructions, incident contacts, recovery steps and a named system owner
- User guidance on permitted data, required checks and reporting a problem
Staff need to practise the workflow with realistic examples, including incorrect answers and unavailable source systems. They should know when to review an output, how to correct it and where to get help. The receiving team also needs access to the accounts and documentation required to operate the system.
Training and data handling also have legal implications. The European Commission states that AI literacy obligations have applied since February 2025. In Switzerland, the FDPIC confirms that the Data Protection Act applies to AI-supported processing of personal data. Review the obligations that apply to the actual deployment as part of discovery.
After launch: measure results and maintain the system
Compare the deployed workflow with its baseline. Track completion time, review effort, error rates and cost per successful task. Check whether people use the system and whether saved capacity improves service or supports additional work. These measures help decide whether to expand the deployment, change it or stop it.
Ongoing operations should include incident handling and evaluation after changes to models, prompts, retrieval sources or connected systems. Agree who reviews failures, who can release a change and how the previous version can be restored. A support arrangement needs those responsibilities as well as a response time.
At handover, the team should know how to use the system, check its results and recover when it fails.
Begin with one workflow whose result you can measure. Use the assessment to establish what it needs, then agree the implementation and handover that will let your team keep using it. For questions about selecting a supplier, see the AI consulting partner checklist.