AI consulting for small business is worth paying for when it ends in three things: a working prototype on your own documents, a written data-protection answer, and a build or no-build decision. A proposal that ends in a slide deck instead should stay unsigned.
The failure pattern is familiar: the workshop felt productive, the deck looks sharp, and three months later every tender, invoice, and support ticket is still handled by hand.
Caution is the right instinct, because a small AI budget only survives one bad engagement. This article lists what a paid assessment must deliver, how honest pricing works, and which work a serious consultant refuses. The receipts come from webvise's own offer documents and from a construction client that went from discovery call to a production AI portal.
- Demand six deliverables in writing: workflow selection and mapping, prototype scope, data and privacy requirements, prompt and tool-flow design, a working prototype, and a production build plan.
- Honest pricing is estimated after discovery. A fixed quote before anyone has seen your workflow is a guess.
- The DSGVO answer is a deliverable: an Art. 28 processing agreement with no-training guarantees, local models, or a hybrid where a local model redacts personal data before cloud processing.
- Refusals are a quality mark: generic AI strategy without a named workflow and autonomous legal, tax, clinical, or credit decisions both get a no at webvise.
If one workflow already hurts every week, webvise's AI consulting service runs the assessment in 2 to 4 weeks and ends with a prototype your team can judge. If no use case has surfaced yet, read where to start with AI when there is no use case before paying anyone.
The deliverable list to demand before signing
webvise publishes the assessment scope on the service page instead of negotiating it per deal, and the default first engagement pairs the assessment with one focused build. The list below is the published offer copy. Measure any proposal, from any consultant, against it.
- Workflow selection and process mapping, including inputs, decisions, exceptions, and review points
- Prototype scope for one high-friction task
- Data, privacy, and human-review requirements, written down before any build
- Prompt, retrieval, and tool-flow design for that specific workflow
- A working prototype, tested on realistic examples from your team
- A production build plan with architecture, cost drivers, risks, and next steps
Pricing follows the same rule. The one-pager webvise sends to prospects states it plainly: effort is estimated after discovery. A consultant who quotes a fixed number before mapping your process is pricing a template, and a template was never your problem.
| Stage | Duration | You leave with |
|---|---|---|
| Discovery call | 30 minutes, free | Fit check and workflow candidates |
| Workflow assessment | 2 to 4 weeks | Process map, working prototype, DSGVO setup, build plan |
| First focused build | Estimated after discovery | One production workflow with review gates |
How the arc ran at MP Bau
MP Bau, a German general contractor, is the public reference for this arc. The engagement moved through four steps: a discovery call, a ranked list of use cases, an on-site workshop with the people who own the workflows, and a build decision.
Two workflows survived that workshop as validated candidates. The engagement ended in a production AI portal the team works with today, and no strategy deck existed at any point in between. Every stage produced something the client could inspect: a ranked list, a validated workflow, a running system.
Whether your business is ready for that arc is a different question, and the AI readiness assessment answers it with a checklist. This article assumes you pass and are deciding what to buy. The same sequence, assessment first and one focused build second, is the default shape of webvise's AI consulting engagement.
The data-protection answer belongs in the deliverables
German small businesses ask the same question in nearly every discovery call: where does the data go? A real assessment answers it in writing, with one of three setups.
- Cloud with contracts: an Art. 28 DSGVO processing agreement with no-training guarantees from the model provider.
- Local models: inference runs on your own hardware, and no document leaves the building.
- Hybrid: a local model redacts personal data first, then the cleaned text goes to a cloud model.
One client boundary shows how concrete the written answer gets: external documents such as incoming tenders may go to a cloud model, internal data never leaves the network. That single sentence is a deliverable. A consultant who cannot produce its equivalent for your business has not finished the assessment.
Compliance is designed in, and irreversible steps stay human
From August 2, 2026, Art. 50(1) of the EU AI Act requires that people are told when they interact with an AI system. webvise maps that obligation into every build as disclosure by design: the chatbot identifies itself as AI, and generated content is labeled where the law requires it. The details for small-business chatbots are in the AI Act transparency guide.
The same design pass fixes the review gates. The operating rule across webvise builds: humans approve anything irreversible. A drafted reply can go out after review, while a quote, a contract, or a payment never leaves without one.
What gets refused, and why that protects you
Two categories get a no at webvise regardless of budget. Generic AI strategy without a named workflow is refused because it produces exactly the deck this article warns about. Autonomous legal, tax, clinical, or credit decisions are refused because no review gate makes them safe enough for a small business to carry the liability.
The boundary holds in real scoping work. For a tax firm's client-facing chat, the design limit was hard from the start: no tax advice in the chat, ever. The assistant may sort documents and draft internal summaries, and the advice itself stays with the professional who signs it.
Ask any consultant for their refusal list before you sign. A list with real entries tells you more than the reference list.
Three questions before you sign anything
First, ask for the deliverable list in writing and compare it against the six items above; a working prototype must be on it. Second, ask how the price was set: estimated after discovery is the honest answer. Third, ask what gets refused, and walk away from a yes to everything.
webvise has delivered 25+ projects on these rules and runs the AI assessment for small businesses that can name one workflow worth fixing. If you want the workflow, the privacy setup, and the build decision settled in one engagement, describe your workflow and the discovery call covers the rest.
Development practices are aligned with ISO 27001 and ISO 42001 standards.