A company does not need to start with a tool. It needs to identify a commercial process that already has a recurring question, a recognizable source of information, and a person who can review the result.
The conversation about artificial intelligence often starts too late: with a demo, a license, or a list of automations. The problem is not that those tools are useless. The problem is that they arrive before anyone has decided which part of the commercial process should improve and how the company will know whether it has actually improved.
In a small or midsize company, commercial friction rarely lives in a single system. It appears in the meeting whose information never reaches the CRM; the proposal that is rebuilt because nobody can find the approved version; the opportunity with no clear next action; or the distributor asking a question that someone else answered months ago. AI can help at those moments. But it does not replace the decision to put the process in order.
The starting point is not AI. It is a specific commercial friction.
An initial application deserves attention when the work repeats, affects a commercial decision, and leaves a trace the team can review. The process does not need to be perfect. It does need to be identifiable.
A simple way to recognize it is to describe the situation in one sentence: “Whenever this happens, someone has to find this information, decide this action, and record it in this place.” If the sentence cannot be completed, there is not yet a use case. There is an intuition.
Four entry points that are often reasonable
- Proposals and quotes. When the team searches for references, requirements, or previous answers before building a first draft.
- Commercial knowledge. When answers about products, terms, cases, or customers are scattered across files, emails, and people.
- CRM and follow-up. When a conversation is not consistently translated into a next action, an owner, and a date.
- Meeting preparation. When preparing for a visit requires reconstructing the context of an account from several known sources.
The common denominator is not “using AI.” It is reducing a search, an information handoff, or a delay that currently prevents the team from moving forward with sound judgment.
What information a company needs before automating
AI does not automatically turn scattered information into reliable information. Before asking a system to summarize, classify, or draft, it is worth answering five operational questions.
- What is the current process? Describe the starting point, the steps, the expected result, and the most common exceptions.
- Which sources may be used? Distinguish approved documents, working materials, historical information, and content that should no longer be reused.
- Who is accountable for the result? There must be a person who can validate a proposal, correct a CRM update, or resolve an exception.
- What may the system not decide? Price, margin, technical scope, contract terms, customer commitments, and final versions must have explicit boundaries.
- Which examples will be used for testing? A pilot needs real but controlled cases to assess usefulness, errors, and review time.
This order has a practical foundation. The NIST AI Risk Management Framework organizes the work around governing, mapping, measuring, and managing. It is not a commercial recipe or a certification of results. It does reinforce one essential idea: roles, context, evaluation, and oversight are not details to add after buying a tool.
Proposals, knowledge, CRM, or follow-up? How to choose the first workflow
Not every problem should be addressed at once. The first choice should follow the friction that is most visible and most reviewable today.
| If this is happening today… | The first focus could be… | The question the company must be able to answer |
|---|---|---|
| Answers must be found in too many places. | Commercial knowledge. | Which sources are approved, and who maintains them? |
| Proposals start from scratch even when sections repeat. | Proposals and quotes. | Which content can be reused without inventing scope or terms? |
| Opportunities lose continuity after calls or meetings. | CRM and follow-up. | What is the next action, who owns it, and when is it due? |
| A meeting requires manually reconstructing the account history. | Meeting preparation. | What information must reach the owner before speaking with the customer? |
The table does not prescribe a tool. It helps separate an information problem from a process problem. In the first case, the company may need to establish sources and access. In the second, it may need to define a stage, an owner, or a follow-up rule before automating anything.
Common mistakes when buying tools
1. Starting with the largest integration
Connecting every system may sound ambitious, but it also multiplies dependencies, permissions, exceptions, and the potential for error. A controlled initial pilot often teaches more than an extensive integration without a clearly defined use case.
2. Confusing a chatbot with a commercial process
A standalone assistant can be useful for exploration. It is not, by itself, a commercial system. To deliver sustained value, it must know which sources it may use, what it may not claim, when it must request review, and where the result is recorded.
3. Measuring activity instead of usefulness
The number of messages generated, documents summarized, or automations activated does not prove improvement. It is more useful to track a measure that belongs to the workflow: search time, the percentage of opportunities with a next action, time to first draft, material corrections, or questions that remain unanswered.
4. Delegating decisions that remain human
AI can propose. It should not assume, without control, a price, a specification, the closing of an opportunity, or a commitment to a customer. In commercial processes, speed without clear accountability can be costly.
5. Treating every existing document as valid
A file being available does not mean it is current, approved, or appropriate for a new proposal. A useful system needs to distinguish origin, validity, and ownership.
Useful context for Spanish SMEs
Adoption is not just about choosing software. In its 2025 analysis of Spain, the OECD reported that approximately 10.4 percent of Spanish SMEs say they use AI applications. The same source links the adoption gap to limited awareness, skills, regulatory concerns, and integration complexity. This is national context, not an E2M survey or a prediction of results for any particular company.
That is precisely why a disciplined first decision is worth more than a spectacular demo. The goal is not to “catch up” with AI. It is to choose a commercial process in which the company can learn without compromising information quality or customer relationships.
When a Commercial AI Diagnostic makes sense
A Diagnostic makes sense when a recognizable friction already exists, but it is not clear which workflow should come first, which information may be used, or what scope a responsible test should have.
E2M’s Commercial AI Diagnostic reviews one priority process, its sources, tools, people, and dependencies. It delivers prioritized use cases, a recommendation for an initial pilot, and explicit boundaries. It is not unlimited implementation, a full CRM migration, or a promise of guaranteed results. It is a way to make a better decision before expanding the scope.
Which commercial process should be reviewed first?
If there is a specific friction and an internal owner, E2M can assess whether an initial conversation is a good fit.
Sources and limits of this guide
- NIST AI Risk Management Framework, Core. A voluntary risk management framework; it does not certify commercial results or replace legal advice.
- OECD Economic Surveys: Spain 2025. The 10.4 percent figure refers to Spanish SMEs that reported using AI applications in the context described by the OECD.