The real problem isn't in the tool
Most German companies now use AI. According to a Bitkom survey of 603 companies with 20 or more employees (Bitkom is the German digital industry association; in German), the figure is 57%, up from 36% the year before. In the same survey, 59% of AI users say that, by their own assessment, they are not making use of its potential at all.
In daily work it looks like this. For a follow-up quote, the AI assistant suggests a contact person who left the customer months ago. The automated email flow writes to the same customer twice, because the customer is in the CRM under two spellings. The revenue overview for the annual review shows figures that accounting corrected long ago.
None of these errors originates in the tool. The tool works with what it finds, only faster and without asking.
The gap between what AI promises and what it delivers is rarely technical. It lies in trust in your own customer data. The 90-minute data check shows whether this data can carry the planned use case before you put money into the first tool.
What the trust-data gap is and why it becomes visible in 2026
The data exists, in the CRM, in accounting, in the email tool. Still, nobody would base an important decision on it alone. Before sending a quote, sales would rather call the customer once more; before revenue planning, someone checks the figures against their own list. There is enough data. What is missing is trust in it. We call this the trust-data gap.
As long as people work with the data, experience covers the gap. The experienced employee knows the customer is in the system twice and corrects it in her head. An AI assistant doesn't know that. A person makes a given error once a week; the assistant repeats it a hundred times a day. The email to the wrong contact sounds just as fluent as the one to the right contact, and it often goes out before anyone checks it.
What the studies show
The figures come mostly from international surveys. They show a direction. They are no benchmark for your company.
In the State of Data and Analytics by Salesforce (November 2025, more than 7,600 respondents in 18 countries), data and analytics leaders estimate that around 26% of their organisation's data is not trustworthy. 84% say their data strategy needs a fundamental overhaul before their AI plans can succeed. Of those already using AI in production, 89% have experienced incorrect or misleading AI results.
The report State of CRM Data Management in 2026 by Validity (August 2026, 500 marketing leaders in the US, the UK, Brazil, Australia and New Zealand) goes a step further:
- Only 21% consider their CRM data very well prepared to support AI.
- Almost 78% of respondents from top management have acted on an AI recommendation that they later believed to be wrong because of bad data. Among employees without a management role, the figure is 41%.
- Just under 60% of top management feel under pressure to introduce AI now, even though they know their data isn't ready.
- Two thirds of organisations have handed more marketing decisions to autonomous AI agents over the past year.
So top management acts on a flawed basis much more often than staff at the working level, while more decisions move to systems that don't ask questions.
In the Bitkom survey, half of German companies name data preparation as a major cost factor in using AI, almost level with infrastructure (51%) and far ahead of software licences (21%). Among companies without AI, 40% name missing data as an obstacle. In many places, the decision to use AI has been made. What remains open is whether you would pass on the results unchecked.
How to tell in your own company that your data won't carry the load
Bad customer data rarely shows itself all at once. Usually it hides in small workarounds that everyone has got used to. We see four patterns in mid-sized companies again and again. If two of them sound familiar, no AI should access this data unchecked.
Sales keeps its own list
The truth sits in an Excel file on the head of sales' laptop, and the CRM gets updated when management needs a report. This isn't a discipline problem: the official system is too cumbersome for daily work or has the wrong fields. An AI then learns from the system nobody uses.
The same customer has three different statuses
In the CRM a customer is active, in accounting there is an open payment reminder, and in support a complaint has been running for weeks. Each system is right on its own terms, and none of them knows the whole picture. Anyone who wants to know how things stand with the customer calls three colleagues.
This slows AI down too. In the sales study State of Sales 2026 by the same publisher (4,050 sales professionals in 22 countries), 51% of sales leaders already using AI say that disconnected systems are slowing down their AI initiatives. An assistant that sees only one of the three sources still formulates a recommendation, with the same confidence as if it had read all three.
Nobody knows who owns a field
Ask who is responsible for the industry or customer class field. Usually there is silence. The consequence: the same metric produces different values depending on the department, and people argue about the number instead of the decision.
Cancellations only show up once they take effect
The cancellation sits in the account manager's inbox, and nothing changes in the CRM until the payment stops arriving. Until then, the customer appears in every revenue plan. An AI meant to detect churn risk rates them as stable.
A quick self-test
- Besides the CRM, are there lists that individual teams rely on more than on the system?
- Can you say for any customer, in under five minutes, whether they pay, whether they are satisfied and who their contact person is?
- Does every important customer field have a named owner?
- Do you learn about a cancellation on the day it arrives, or only when the payment fails to come in?
Every answer where you hesitate marks a place where your data can't carry the load today.
The 90-minute data check: structure and preparation
The data check is a working session rather than a project: three blocks of 30 minutes each. At the end you have an assessment you can base a decision on.
Who sits at the table
Three roles are enough: management with the planned AI use case, one person from sales or service who works with the customer data every day, and whoever looks after the CRM. That is often a task alongside day-to-day business, and that is enough. You only need to ask IT for the export in advance.
What you prepare
- An export of the customer master data as a table (companies, contacts, status, revenue fields, date of last change), limited to the subset your use case would use.
- A list of the tools holding customer data, including the teams' Excel lists. One row per tool: which customer data is stored here, and where does it come from?
You don't need anything more than a spreadsheet and a sheet of paper.
What the check aims to achieve
The result is a traffic light for the specific use case. An audit asks whether all the data is in order. The check asks a narrower question: does the data this one tool needs carry a decision?
One rule for the session: count instead of estimating.
Block 1 (30 minutes): Completeness and duplicates
Do the fields your tool needs contain usable values? And does every customer exist only once?
Step 1: Define mandatory fields (about 5 minutes)
Check only what the use case needs. Typical candidates: company name, contact person with job title, email address, customer status, industry, responsible account manager, latest revenue. Stick to a few fields.
Step 2: Draw a sample (about 5 minutes)
Take roughly every twentieth entry or a selection using a random function. Don't choose by hand: otherwise you reach for the well-known customers, and they are usually better maintained than the rest.
Step 3: Count empty and incorrect entries (about 10 minutes)
Assign each entry to one of three categories: usable, empty or obviously wrong. Obviously wrong means placeholders such as xxx or unknown, a general info address instead of a person, a contact who has left the company, or an active customer who cancelled long ago. No formula catches that last one; only someone at the table who knows the customer does.
At the end, you have the share of usable entries for each mandatory field. There is no universally valid threshold. Decide at the table what share is enough for this use case. A field the AI reads for every answer tolerates fewer gaps than one that only helps occasionally.
Step 4: Look for duplicate companies and contacts (about 10 minutes)
To an AI, two entries for the same company are two customers. Then the cancellation sits with one, and the renewal offer goes to the other. Use the entire export here:
- Sort by company name and look for spelling variants: with and without the legal form, abbreviated, typos, old and new name.
- Sort by email domain. The same domain with different company names is a strong indication of a duplicate.
- Check contacts that have been created more than once, for example with an old and a new address.
Note how many companies and contacts appear more than once.
Look out for patterns too. If a field is empty mainly for customers from the web form or an old import, the cause lies in a handover between systems rather than in maintenance. That is where block 2 comes in.
Block 2 (30 minutes): Freshness and data flows
A record can be complete and still wrong. Block 2 checks whether the content is still correct and where it comes from.
Part A (10 minutes): When was the record last touched?
Decide when a record counts as outdated for your use case. For quotes, the limit may differ from the one for service requests. Sort the sample from block 1 by date of last change and count how many records fall beyond the limit.
Do a cross-check: for recently changed records, it was often just an automatic import, while the contact person has been the same for years. Note the share of outdated records.
Part B (20 minutes): Where does the flow of information break?
The email tool knows about unsubscribes and undeliverable addresses, accounting knows actual revenue, and support knows about complaints and intentions to cancel. Does this information arrive where the AI will later read?
Sketch the customer journey, for example:
Enquiry → Quote → Order → Invoice → Account management and support → Renewal or cancellation
Under each step, enter the leading tool and who enters data there. Then check each handover with three questions:
- Is the information transferred automatically or by hand? Every manual transfer is a place where information gets lost or arrives too late.
- Which system is right in case of conflict? If nobody in the company can say whether the revenue in the CRM or in accounting applies, the AI won't know either.
- How long does the handover take? A cancellation that only appears in the CRM weeks later is invisible to any tool giving recommendations in the meantime.
Mark every handover red that runs manually, has no clear precedence or is noticeably delayed. Only the handovers your use case needs count.
Block 3 (30 minutes): Ownership and approval
This is the most important part. A field that nobody owns looks the same three months after the clean-up as it did before.
Step 1 (10 minutes): Name field owners
Write a name after every mandatory field from block 1. One person rather than a department. Customer status should be owned by whoever actually processes cancellations. For industry or segment, it is often nobody, and that is precisely the finding. If the column stays empty, record it that way.
Step 2 (10 minutes): Check editing rights
Who may change the value today, and who actually does? Often everyone may change everything. Then sales overwrites what service entered, and an interface to accounting overwrites both overnight. Record which system has the last word for which field.
Step 3 (10 minutes): Decide on the AI's access
Who decides whether an AI tool may access this data and what it may do with it? This is where governance begins in a practical sense. Set an approval level for the use case:
- Read: the AI analyses and changes nothing.
- Suggest: the AI makes suggestions, and a person checks and adopts them.
- Execute with approval: the AI only acts after confirmation by the owner.
- Execute independently: people only check samples.
Rule of thumb: the less you trust a field, the lower the level. Name one person who knows the business process, owns the level and, when in doubt, can say: this is where we stop.
When an AI accesses customer data, data protection questions arise too: which data, processed where, on what basis. The check does not replace a legal review. Involve your data protection officer before the pilot starts.
The results sheet
At the end you have a table with four columns: field, owner, system with the last word, permitted approval level. It is the basis for the traffic light and for the week after.
Reading the traffic light: start, clean up or switch use case
The traffic light tells you whether this data is solid enough for this purpose. You set the limits yourself, guided by one question: what does it cost if the tool gets it wrong based on this data? A flawed internal summary is annoying. A quote sent automatically to the wrong contact costs you the customer's trust.
Lay the counted values side by side and read them like this:
| Colour | What the figures show | What you do |
|---|---|---|
| Green | All mandatory fields are above the limit you set at the table. No red handover without a solution. Every row in the results sheet has a name. | Start a pilot with a clearly limited subset of data. |
| Amber | Individual fields are below the limit, or there are duplicates or red handovers. Ownership, however, is settled. | Clean up only the fields the use case needs. |
| Red | Most mandatory fields are below the limit, or central fields have no owner. | Start with a use case that needs less customer data. |
Green: pilot with a limited subset of data
Start with one region, one customer group or one product segment rather than the whole customer base. The pilot is time-limited, has a fixed goal and one person who regularly checks the results against reality. If the tool doesn't help, the errors stay small. After the pilot comes the handover checklist from pilot to routine operation.
Amber: clean up selectively
We advise against a large clean-up project. If your use case needs company name, contact person and contract status, put exactly these three fields in order and decide who will maintain them from now on. Everything else waits. With costs that, according to Bitkom, are major for half of all companies, this limitation pays off.
Red: switch use case
Red doesn't mean AI has to wait. Choose a use case without customer master data: drafts from your product documentation, search in internal manuals, summaries of meetings. There you can settle approvals and responsibilities while putting your customer data in order in parallel. That way you respond to the pressure described in the studies above without bad data doing damage.
When the blocks show different colours
Then a simple rule applies: the overall result is never better than the colour from block 3. If completeness is amber and ownership is red, the light is red. Gaps in fields can be closed. Without clear ownership, they open up again.
What you should not do
After the check, the pressure is often greater than before. Three reflexes then cost a lot of money and rarely make the data more trustworthy.
A big data clean-up without a use case
Cleaning up everything first sounds responsible, and in practice means months of work on fields no tool will ever read. Without a use case, there is no yardstick for when a record is good enough. The team loses patience before any benefit becomes visible.
A new CRM as a fresh start
A new CRM takes over the old data, with its duplicates and the open question of who maintains what. The Excel lists in sales stay. A switch can be the right move, for example if the system can't be connected to your other tools. But decide only once ownership and data flows are settled. Then you know what a new system has to do.
AI for data maintenance before responsibilities are settled
AI can help find duplicates. But who decides which of two records is correct? Without that answer, the AI writes assumptions into your system, and nobody notices. The order is fixed: first the field owners and approval rules from block 3, then tools that make suggestions, and last of all automation that changes data itself.
What changes on the Monday after
Three steps for the week after, none of which needs budget or a new tool.
1. Make the results sheet binding. Fill in the empty rows from block 3. Where the group found no name, management decides. Have a short conversation with each named person about what they own from now on.
2. Repair one handover. Take exactly one red spot from the sketch in block 2. A typical one: sales closes a deal, service hears about it by email and creates the customer again by hand. Clarify who supplies the information, where it ends up and how you can tell it has arrived.
3. Define the pilot use case. Write it down in two sentences, in line with the traffic light: what should the AI do, and which fields does it need to read for that? Whether your sources are fit for that is checked by the checklist of eight data readiness questions.
After that you don't yet have an AI project, but you have a foundation on which you won't be automating errors.
If you don't want to run the check alone, we can run it together with your team. If the real breaks lie between your systems, we take a closer look in a journey analysis. More on this under AI Operations Diagnostic.
What decision would you base on your CRM data today without checking back first?
Frequently asked questions
Do I need an IT department or special software? No. An export, a spreadsheet and three people at the table are enough.
How clean does our customer data have to be? It has to fit the use case; it doesn't have to be perfect. An assistant that follows up on open quotes needs correct contacts and up-to-date quote versions. It doesn't need the complete history of all contacts.
How often should we repeat the check? Before every new use case, because different fields then become important, and also at intervals that suit your business. If you win a lot of new customers or often get new contacts, check more often.
This article is not legal advice.
Sources
- Bitkom: Erstmals nutzt Mehrheit der Unternehmen KI (for the first time, a majority of companies use AI; in German), press release of 14 September 2026
- Validity: State of CRM Data Management in 2026, press release of 25 August 2026
- Salesforce: State of Data and Analytics, 4 November 2025
- Salesforce: State of Sales 2026, 3 February 2026