Most SMEs do not need a data warehouse, a data lake or a data team before they use AI. They need five far more ordinary foundations, covering records, identifiers, documents, ownership and a basic cloud setup, and most of them make the company easier to run whether or not the AI project goes ahead.
The question "is our data ready for AI?" is usually asked the wrong way round. Data is ready for a specific use or it is not, and an assistant that answers staff questions from internal documents sets a far lower bar than a model that sets prices. A 40-person wholesaler can be ready for the first and years away from the second.
What data foundations does an SME need before AI?
Five, and none of them require a data team. One system of record for each core entity such as customer, product and order. Identifiers that stay the same in every system. Documents stored in one searchable place, with access rights that match who should actually see them. A named person who owns each dataset. And a basic cloud setup with tested backups and logging switched on.
One system of record per core entity. If customer addresses live in the CRM, the accounting package and the webshop, and all three can be edited, there is no customer record. There are three opinions, and any AI built on top inherits the argument. Decide which system wins for each entity and make the others read from it.
Consistent identifiers. The article number in the webshop should be the article number in the ERP and on the supplier invoice. In the companies we assess, mismatched codes are the usual reason a forecasting project stalls early, and the fix is clerical rather than technical. It is the same cleanup that dominates ERP data migration.
Documents in one place, with access that reflects reality. A retrieval-based assistant answers from whatever the user can read. If the HR folder is open to everyone, the assistant will quote salaries to anyone who asks. Oversharing that was invisible becomes visible the day the assistant goes live.
A known owner of each dataset. Not a governance board. One person whose name sits next to the product master, who decides what a valid record is and fixes it when it is wrong.
A basic cloud setup. Multi-factor authentication, backups that someone has actually restored, and logs that show who accessed what. Logs matter more once an AI tool reads across systems, because someone will ask what it touched.
Do we need a data warehouse or a data team first?
Usually not. A company of 10 to 200 people with clean operational systems can run a document assistant, email triage and a first forecast directly from those systems or from a scheduled export. A warehouse earns its cost when several reports or models need the same combined data every day and the exports keep breaking. Built on mismatched identifiers, a warehouse only centralises the mismatch.
The same applies to people. In a 60-person company the work belongs to someone who knows the operation, often the finance or operations lead, with a few protected hours a week and the authority to decide which record wins.
A two-week readiness check with a spreadsheet
One sheet, one row per dataset, and these columns: dataset, system, owner, identifier used, editable or read-only, sample match result, access reviewed, backup restored, status.
Week one is inventory. List every place customer, product, order and document data lives. Fill in the owner column honestly; a blank cell is a finding. Mark each system where the same entity can be edited, because every extra editable copy is a future conflict.
Week two is testing. Take fifty customers from the CRM and look each one up in accounting by identifier. Do the same for products between the webshop and the ERP. Open the five most-used document folders and check who can read them. Ask whoever runs IT to restore one file from backup while you watch.
Then score each row red, amber or green against the AI use you chose. The output is not a report. It is a ranked list of three fixes, each tied to that use.
What good enough looks like, by AI use
The bar rises sharply as the AI moves from reading to deciding.
| AI ambition | Minimum data needed | Typical gap in SMEs | First fix |
|---|---|---|---|
| Document assistant | Current documents in one store, permissions that match roles | Old versions everywhere, folders open to all staff | Archive superseded files, review access on the busiest folders |
| Email triage | Agreed categories, labelled past emails, customers findable by email address | Categories exist only in people's heads | Agree the categories and label a recent month of email |
| Demand forecasting | Two or more seasons of order history on stable codes, stockouts and promotions recorded | Codes changed in a migration, stockouts invisible in sales data | Map old codes to new, start logging stockouts now |
| Pricing optimisation | Net transaction prices including discounts, current unit costs, a market reference | Discounts agreed off-system, costs updated once a year | Record the actual net price on every order |
A document assistant tolerates mess, provided the mess is permissioned correctly. It retrieves and quotes, and a person reads the answer. Forecasting and pricing are unforgiving because the model learns from history, so a recording error becomes learned behaviour. The specific traps in sales history are set out in demand forecasting without a data science team.
Does our data have to stay in Switzerland?
No. The revised Federal Act on Data Protection does not require personal data to be stored in Switzerland. It regulates disclosure abroad instead: data may go to countries the Federal Council lists as providing adequate protection, and to other countries only with appropriate safeguards such as standard contractual clauses. Swiss hosting is a legitimate choice for a client contract or internal policy, but it is a business decision, not a legal requirement.
Three practical points when choosing a cloud region.
Storage and processing can sit in different places. The large cloud providers run Swiss regions, so residency is available if a contract demands it. An AI service may still process requests elsewhere while your files sit in Zurich. Ask where inference runs.
Support access counts. Remote access to your data from another country can amount to disclosure abroad, so a provider's support model matters as much as its region.
Your privacy notice has to keep up. Under the FADP, people must be told which countries their data is disclosed to and on what safeguards. Adding an AI tool that processes abroad can change that notice. The Federal Data Protection and Information Commissioner publishes guidance on cross-border disclosure. The same argument in a regulated setting is in cloud migration in Swiss banking.
Where to start
Pick the AI use first, run the check against it, and fix the red row that blocks it. A document assistant is often a sensible first project because its bar is the lowest. Candidate uses are compared in generative AI use cases for small businesses.
If the check passes and the project still stalls, the problem has moved from data to ownership, which is the pattern in getting AI past the pilot stage. Our cloud computing work usually starts with exactly this spreadsheet. The model choice comes afterwards, in AI and machine learning.
Related reading
- Getting AI past the pilot stage in Swiss companies
- ERP data migration: the part that overruns, and why
- Demand forecasting without a data science team
- Cloud computing consulting
Working on this right now?
Tell us where you are in two questions. A consulting partner reviews every enquiry within 2 business days.