The question we get almost every week is not whether AI is useful, but what exactly it is useful for in a company of twenty, fifty or two hundred people. This article answers with a catalog: 40 concrete use cases by department, with what each one solves, how to ask for it, what data you need, how much time it can save as an estimate, and what to watch. At the end, the table of where to start based on the size of your company and the list of what you should never delegate.
Short answer: An SMB starts with one department, three use cases and one metric measured before kickoff. The cases that work best are the ones that happen every week and whose output can be verified in five minutes: qualifying inquiries, quoting, logging calls in the CRM, answering frequently asked questions and reconciling reports. If you handle customer or employee data, a commercial plan is the sensible choice, since by default it does not train on your content.
- Where does an SMB that wants to use Claude seriously start?
- How do you read each use case fact sheet?
- Sales: seven use cases
- Customer service: six use cases
- Marketing and content: five use cases
- Administration and finance: six use cases
- Purchasing and logistics: five use cases
- Human resources: six use cases
- Leadership and management: five use cases
- Where should you start based on the size of your company?
- What data do you need before you start?
- What should you NOT delegate to AI?
- How do you know whether this is working?
- Tricks that are not in the manual
- What to do this week
Where does an SMB that wants to use Claude seriously start?
With a single department, three concrete use cases and one person accountable, not with a general transformation plan. The SMBs that get results pick tasks they already do every week and whose output can be verified in five minutes. The rest comes later, once the team trusts the tool and knows how to correct it.
This article is a catalog: 40 use cases spread across seven departments of a typical SMB in Mexico or Paraguay, all with the same fact sheet. You do not have to read it end to end: find your department, pick two or three cases you recognize as a real pain, and start there.
One clarification that saves disappointment: the difference between a consumer plan and a commercial one is not only about features, it is about default data policy. Anthropic documents that in its commercial products (Team, Enterprise and API) your inputs and outputs are not used to train models by default, whereas on the Free, Pro and Max plans that depends on a setting the user turns on. You can see it on the official pages about training in commercial products and in consumer plans (both accessed on September 19, 2026). If you are going to work with customer or employee data, that is the first decision, before choosing use cases.
On costs: as of September 2026, Anthropic's pricing page lists Pro at 17 dollars per month billed annually or 20 per month, and the standard Team seat at 20 or 25 per month, with a two-seat minimum. All prices are in dollars, before tax, and vary by region: confirm on their site before you budget.
How do you read each use case fact sheet?
Every sheet has the same five fields, so you can compare them and decide quickly.
- What it solves. The concrete problem, not the technical capability.
- How to ask for it. The shape of the request, in short. The full structure of a good brief is in the guide to prompts and tricks.
- Data you need. What you need at hand before you start. If you do not have it, the case does not apply yet.
- Estimated time saved. These are estimates to calibrate expectations, not measurements. They depend on the team, the quality of the data and the practice. Measure your own baseline before taking them as given.
- Risk to watch. What can go wrong and always has to be reviewed.
The rule that applies to all 40 sheets: in every one of them, a person reviews the output before it leaves the company. AI drafts, organizes and compares; the responsibility still belongs to whoever signs it.
Sales: seven use cases
This is where the return shows up first, because almost all of a rep's administrative work is text: notes, emails, quotes and reports.
1. Qualifying the leads that come in through the website and WhatsApp
- What it solves: The team treats someone asking for a price out of curiosity the same as someone with an approved budget.
- How to ask for it: Ask it to sort every inquiry into A, B or C using criteria you define, with the reason for the classification and a suggested opening line for the reply.
- Data you need: Your qualification criteria in writing and 20 examples of real inquiries you have already classified yourself.
- Estimated time saved: 3 to 5 hours a week on a team of three reps.
- Risk to watch: Classifying a large customer who wrote a terse message as a C. Review the Cs once a week.
2. Preparing visits and calls
- What it solves: People show up to the meeting without remembering what was promised last time.
- How to ask for it: Paste in the customer's history and ask for the three facts you cannot forget, what was promised and not delivered, and the open question.
- Data you need: Emails, notes and tickets for that account.
- Estimated time saved: 15 to 25 minutes per prepared meeting.
- Risk to watch: Filling in with generic industry context. Require that everything come from the history.
3. Building quotes from the price list
- What it solves: Every rep builds the quote their own way, with calculation or validity errors.
- How to ask for it: Give it your price list and your template, and ask for the quote for a specific case, respecting terms, lead times and validity.
- Data you need: Current price list, commercial terms and an approved template.
- Estimated time saved: 20 to 40 minutes per complex quote.
- Risk to watch: Out-of-date prices. Set a rule: if the list is more than 30 days old, it does not get used.
4. Turning the call into a CRM record
- What it solves: The CRM gets filled in late, badly or never.
- How to ask for it: Paste your notes as they are and ask for the exact fields in your CRM, using only the allowed values you give it.
- Data you need: The list of fields and allowed values in your CRM.
- Estimated time saved: 10 to 15 minutes per opportunity recorded.
- Risk to watch: Filling in fields that were never discussed. Require that it leave blank whatever does not appear.
5. Staged follow-up without losing contacts
- What it solves: Follow-up depends on the rep's memory.
- How to ask for it: Ask for a 90-day plan with dates, channel and a useful reason for each touch, plus the text of each message.
- Data you need: The typical length of your sales cycle and the useful material you can share.
- Estimated time saved: 2 to 4 hours a week per rep.
- Risk to watch: Every message sounding the same. Ban opening with the question of whether they have decided yet.
6. Post-mortem on lost deals
- What it solves: The same deal gets lost for the same reason three times a year.
- How to ask for it: Paste the full thread and ask for the timeline, the moment it went cold, the signal that was ignored, and a single actionable recommendation.
- Data you need: Emails and notes from 10 to 20 deals closed as lost.
- Estimated time saved: 2 to 3 hours per quarter, with a lesson you were not getting before.
- Risk to watch: Pleasant but false conclusions. Ask it to cite the specific message it is relying on.
7. Weekly pipeline report
- What it solves: Management gets a spreadsheet that nobody reads the same way.
- How to ask for it: Paste the CRM export and ask for half a page: what moved, what stalled, what came in, what was lost, and three decisions to make.
- Data you need: The CRM export and your definition of a stalled deal.
- Estimated time saved: 1 to 2 hours a week of the sales manager's time.
- Risk to watch: Softening what is going badly. Ask it explicitly to flag stalled deals even if the rep says otherwise.
Want a hand applying this?
Free 30-minute working session. Uniamos is a remote AI automation agency (Austin and Mexico City) working over video calls and WhatsApp.
Customer service: six use cases
The value is not in answering faster, it is in answering equally well every time and spotting what repeats so you can fix the cause.
8. A bank of answers to frequently asked questions
- What it solves: Every person on the team answers the same thing differently.
- How to ask for it: Ask it to group six months of inquiries into 30 frequently asked questions and write the approved answer for each one in two versions: email and WhatsApp.
- Data you need: Your inquiry history and the answers you currently consider correct.
- Estimated time saved: 4 to 8 hours up front, which then save several hours a week.
- Risk to watch: Answers that promise more than the company delivers. Review them one by one before publishing.
9. Draft replies to complaints
- What it solves: Replies to complaints go out late or with a defensive tone.
- How to ask for it: Paste the complaint and ask for three drafts: one that acknowledges and resolves, one that asks for information before committing, and another for when the customer is not right.
- Data you need: Your returns policy, warranties and what you can actually offer.
- Estimated time saved: 10 to 20 minutes per complaint.
- Risk to watch: Commitments you cannot keep. Write into the instructions that it must never offer unapproved compensation.
10. Classifying and routing tickets
- What it solves: Urgent cases get mixed in with routine ones.
- How to ask for it: Ask it to classify each ticket by type, urgency and owning department, with one line of justification.
- Data you need: Your case taxonomy and who handles each type.
- Estimated time saved: 30 to 60 minutes a day on teams with medium volume.
- Risk to watch: Misclassified urgencies. Define hard rules: certain words always go to top priority.
11. Summarizing long WhatsApp conversations
- What it solves: A 200-message thread that nobody wants to read before replying.
- How to ask for it: Paste the thread and ask: what the customer asked for, what was promised, what is still open, and by what date.
- Data you need: The thread export, with any personal data you do not need already removed.
- Estimated time saved: 10 to 20 minutes per escalated case.
- Risk to watch: Mixing what the customer said with what the company said. Ask it to mark who said each thing.
12. Translating the technical manual into plain language
- What it solves: The customer does not understand the product documentation and calls.
- How to ask for it: Ask it to rewrite each section for someone with no technical background, with numbered steps and one idea per step.
- Data you need: The current manual and two examples of how the best technician on the team explains it.
- Estimated time saved: It reduces repeat inquiries; the saving depends heavily on the product.
- Risk to watch: Simplifications that change the meaning on safety topics. Those sections get reviewed by the technical team.
13. Spotting the topics that keep repeating
- What it solves: The same inquiry gets answered a thousand times without fixing the cause.
- How to ask for it: Ask it to group a quarter's inquiries by root cause and sort them by volume and by cost to serve.
- Data you need: A ticket or conversation export for the period.
- Estimated time saved: 3 to 4 hours per quarter, with an impact on future volume.
- Risk to watch: Arbitrary groupings. Ask it to show real examples inside each group.
Marketing and content: five use cases
In an SMB, marketing is usually one and a half people. The goal is not to produce more, it is to sustain a rhythm that does not depend on one Friday's inspiration.
14. Quarterly content calendar
- What it solves: You publish in bursts and then abandon it.
- How to ask for it: Ask for a 12-week calendar with topic, format, channel and objective per piece, built from real customer questions.
- Data you need: Your list of frequently asked questions, priority products, and the channels where you already have an audience.
- Estimated time saved: 4 to 6 hours per quarter.
- Risk to watch: Generic industry topics. Require that every piece answer a real customer question.
15. Product sheets and catalog copy
- What it solves: Hundreds of products with no description, or with descriptions copied from the supplier.
- How to ask for it: Give it the spec sheet and ask for a description with the fixed structure you define, at the exact length your store or catalog needs.
- Data you need: Spec sheets and three examples of descriptions you like.
- Estimated time saved: 5 to 10 minutes per sheet, versus 20 or 30 by hand.
- Risk to watch: Invented attributes. Ban any detail that is not in the spec sheet.
16. One piece of content, several formats
- What it solves: An article gets written and that is where it dies.
- How to ask for it: Paste the content and ask for the spin-offs: a social post, a customer email, a short video script, and a summary for sales.
- Data you need: The original content and the rules for each channel.
- Estimated time saved: 1 to 2 hours per repurposed piece.
- Risk to watch: Repeating the same idea in every format. Ask for a different angle per channel.
17. Customer newsletter from internal material
- What it solves: The newsletter does not go out because nobody has time to write it.
- How to ask for it: Paste the month's news in raw form and ask for the newsletter with three blocks and a single call to action.
- Data you need: Real news from the month and your mailing list with consent.
- Estimated time saved: 2 to 3 hours per send.
- Risk to watch: Sending to a list without consent. That is a legal problem, not a marketing one.
18. Reading the competition's public messaging
- What it solves: You compete blind, against what you assume the other side is doing.
- How to ask for it: Ask for a comparison of public messaging: what each competitor promises, who they are talking to, and what gap is still open.
- Data you need: Websites and public materials from three to five competitors.
- Estimated time saved: 3 to 5 hours per analysis.
- Risk to watch: Conclusions with no source. Require that it cite where each claim comes from.
Administration and finance: six use cases
This is where it pays most to be cautious, and where, properly scoped, you recover the most time. The rule here is absolute: AI prepares and reviews, it does not decide and it does not file anything with a tax authority.
19. Reconciling reports and spotting inconsistencies
- What it solves: Two reports that should match and do not.
- How to ask for it: Upload both files and ask only for the differences, with row, amount and a hypothesis about the cause.
- Data you need: Both files, with the same key columns.
- Estimated time saved: 1 to 3 hours per monthly close.
- Risk to watch: Assuming equivalences between columns. Ask it to state what matching it did before comparing.
20. Checking invoices against the purchase order
- What it solves: You pay price or quantity differences that nobody caught.
- How to ask for it: Give it the order and the invoice and ask only for discrepancies in unit price, quantity, lead time and terms.
- Data you need: Legible purchase orders and invoices, in text or PDF.
- Estimated time saved: 5 to 10 minutes per invoice reviewed.
- Risk to watch: Misreading a poor-quality scanned PDF. Every difference found gets verified by hand.
21. Monthly report in plain language
- What it solves: The income statement arrives and only the accountant understands it.
- How to ask for it: Ask for half a page with what changed versus last month, the three figures that explain it, and the two questions management should be asking.
- Data you need: The income statement and the previous month's.
- Estimated time saved: 2 to 3 hours a month of the manager's time.
- Risk to watch: Interpretations with no basis. Ban any statement that is not supported by a figure in the file.
22. Cash flow scenarios
- What it solves: An investment gets decided without seeing what happens if a large customer pays late.
- How to ask for it: Give it your projection and ask for three scenarios with the assumptions stated and the critical month in each one.
- Data you need: Your current projection, real collection terms, and payment commitments.
- Estimated time saved: 2 to 4 hours per exercise.
- Risk to watch: Assumptions hidden inside the calculation. Require that it list them before the numbers.
23. First read of contracts
- What it solves: Contracts get signed without reading the clauses that matter.
- How to ask for it: Ask first for the literal quotes about terms, penalties, auto-renewal, exclusivity and termination, and then for a summary anchored in those quotes.
- Data you need: The full contract in text.
- Estimated time saved: 1 to 2 hours per contract, before it goes to the lawyer.
- Risk to watch: This is not legal advice and does not replace a professional: it helps you arrive prepared to the consultation, not skip it.
24. Preparing the month-end documentation
- What it solves: The accountant asks for the same things every month and gets them incomplete.
- How to ask for it: Ask for a checklist of what is missing, comparing what has been delivered against the list they asked for.
- Data you need: The list of requirements and an inventory of what you already have.
- Estimated time saved: 1 to 2 hours a month.
- Risk to watch: Treating something as delivered when it is not. Ask it to mark as pending anything it cannot confirm.
Purchasing and logistics: five use cases
Buying well means comparing well, and comparing quotes in different formats is exactly the kind of tedious task worth delegating.
25. Comparing supplier quotes
- What it solves: Five quotes with different formats, currencies and terms.
- How to ask for it: Paste them all in and ask for a table with the same columns: comparable unit price, lead time, payment terms, warranty and inclusions.
- Data you need: The complete quotes and the exchange rate you are going to use.
- Estimated time saved: 1 to 3 hours per significant purchase.
- Risk to watch: Comparisons that ignore what is not included. Ask for an explicit exclusions column.
26. Standardized requests for quotation
- What it solves: Everyone requests quotes their own way and the replies cannot be compared.
- How to ask for it: Ask for a request template with specifications, quantities, lead times, delivery, and the exact format for the supplier's reply.
- Data you need: Your technical specifications and usual terms.
- Estimated time saved: It pays for itself the first time you use it.
- Risk to watch: Ambiguous specifications. Check that every requirement is verifiable.
27. Import document control
- What it solves: A document is missing and the goods get held up.
- How to ask for it: Give it the list of required documents and the ones you have, and ask for a list of what is missing and of inconsistencies between documents.
- Data you need: Your customs broker's checklist and the shipment documents.
- Estimated time saved: 1 to 2 hours per shipment.
- Risk to watch: Requirements that change. The reference list is confirmed by your broker, not by the AI.
28. Planning replenishment
- What it solves: You run out of your fastest-moving stock and sit on what does not sell.
- How to ask for it: Give it the sales and purchase history and ask for a replenishment proposal per product, with the lead time stated as an assumption.
- Data you need: 12 months of history, current stock, and each supplier's real lead times.
- Estimated time saved: 3 to 5 hours a month.
- Risk to watch: Projections on histories whose seasonality is misread. Ask it to flag products with an irregular pattern.
29. Evaluating suppliers with real criteria
- What it solves: The decision gets made out of habit or personal relationship.
- How to ask for it: Ask for a sheet per supplier with on-time delivery, incidents, relative price and dependency, built from your own records.
- Data you need: Purchase history, incidents and complaints.
- Estimated time saved: 2 to 4 hours per semiannual review.
- Risk to watch: Conclusions drawn from too little data. Ask it to state how many orders support each sheet.
Human resources: six use cases
This is the department with the most benefit per hour invested and the most delicate one, because you are working with people's data. Nothing that follows replaces a human decision.
30. Job descriptions and job postings
- What it solves: A posting copied from another one gets published and the wrong candidates apply.
- How to ask for it: Ask for the description based on what the person actually does, separating responsibilities, must-have requirements and nice-to-haves.
- Data you need: A conversation with whoever holds the role today or with their direct manager.
- Estimated time saved: 1 to 2 hours per opening.
- Risk to watch: Disqualifying requirements that are not needed and shrink the candidate pool for no reason.
31. First read of applications
- What it solves: Two hundred resumes and nobody with time to read them.
- How to ask for it: Ask for a five-line summary of each resume against the requirements, with no scoring and no rejecting, flagging which requirement cannot be verified.
- Data you need: The resumes and the list of objective requirements.
- Estimated time saved: 4 to 8 hours per hiring process.
- Risk to watch: Bias. Do not delegate rejection: a person makes the selection, and it is worth stripping from the files any data that is not relevant to the role.
32. Structured interview script
- What it solves: Every interviewer asks different things and candidates cannot be compared.
- How to ask for it: Ask for eight questions tied to the real responsibilities, indicating which answer reveals experience and which one only theory.
- Data you need: The job description, already defined.
- Estimated time saved: 1 to 2 hours per process, plus real comparability between candidates.
- Risk to watch: Questions about personal matters that have no place in a hiring process.
33. A 30-day onboarding plan
- What it solves: The new employee spends two weeks not knowing what to do.
- How to ask for it: Ask for the day 1, week 1 and day 30 plan, with who supports each milestone, what they should be able to do by the end, and how that gets verified.
- Data you need: The org chart, the tools they will use, and the tasks of the role.
- Estimated time saved: 2 to 3 hours per hire, with less early turnover.
- Risk to watch: Unrealistic plans that nobody executes. Ask that every milestone have a named owner.
34. Internal policies people can actually read
- What it solves: The handbook exists but nobody has read it.
- How to ask for it: Ask for each policy rewritten in plain language, with real examples and an FAQ section.
- Data you need: The current policies and the cases that generate the most questions.
- Estimated time saved: 3 to 5 hours per policy.
- Risk to watch: The simplified version contradicting the formal text. The legal version always governs and gets reviewed by a professional.
35. Reading an employee survey
- What it solves: A hundred open-ended answers that nobody processes.
- How to ask for it: Ask for the comments grouped by topic, sorted by frequency and with anonymous examples, without interpreting intent.
- Data you need: The answers, already anonymized.
- Estimated time saved: 3 to 4 hours per survey.
- Risk to watch: Re-identifying people from very specific comments. Remove anything that would let someone be identified.
Leadership and management: five use cases
The value for management is not in generating text, it is in reaching decisions with the information organized and the questions already framed.
36. Weekly dashboard with commentary
- What it solves: The indicators get looked at, but nobody writes down what they mean.
- How to ask for it: Paste the week's numbers and ask for three paragraphs: what changed, what explains it, and what decision it enables.
- Data you need: Your five or six main indicators, with the figures from previous weeks.
- Estimated time saved: 1 to 2 hours a week.
- Risk to watch: Invented causality. Require that it distinguish between what the data shows and what is a hypothesis.
37. Preparing committee meetings
- What it solves: Long meetings that end without decisions.
- How to ask for it: Ask for an agenda with a maximum of three items, the pre-read for each one, and the decision that has to come out of each item.
- Data you need: The open topics and who is attending.
- Estimated time saved: 30 to 60 minutes per meeting, and shorter meetings.
- Risk to watch: Agendas that avoid the uncomfortable topic. Ask it to flag which issue is being postponed.
38. Analyzing a decision with scenarios
- What it solves: Decisions get made with the information at hand rather than the information needed.
- How to ask for it: Describe the decision and ask for three scenarios with stated assumptions and what piece of information would change the decision.
- Data you need: The business numbers that bear on that decision.
- Estimated time saved: 2 to 4 hours per significant decision.
- Risk to watch: False precision. Ask for ranges and assumptions, not exact figures.
39. Internal announcements
- What it solves: Changes get communicated badly and create more noise than necessary.
- How to ask for it: Ask for 200 words covering what changes, from when, who it affects and where to ask questions, plus the five questions the team will ask.
- Data you need: The decision already made and its real consequences.
- Estimated time saved: 30 to 60 minutes per announcement.
- Risk to watch: Softening it until nobody understands what changed.
40. Minutes and tracking of agreements
- What it solves: Something gets agreed and two weeks later nobody remembers who was going to do it.
- How to ask for it: Paste the notes and ask only for the agreements with an owner and a date, and separately the topics left undecided.
- Data you need: The meeting notes, even if they are messy.
- Estimated time saved: 20 to 30 minutes per meeting.
- Risk to watch: Interpreted agreements. If it was not clear, it has to show up as an undecided topic.
Where should you start based on the size of your company?
The order depends less on your industry than on your headcount, because that determines who has time to sustain the habit.
| Size | Start with | Who owns it | What not to do yet |
|---|---|---|---|
| 1 to 9 people | Cases 1, 3, 4 and 8: qualifying inquiries, quoting, recording and answering frequently asked questions | Whoever answers the messages today; an individual paid plan is enough to start | Integrations and connectors. First make sure the tool gets used every day |
| 10 to 49 people | Cases 1, 5, 8, 19 and 30: sales, service, report reconciliation and job openings | One person per department, with a shared project per department | Complex permission policies. A Team plan and basic roles are enough |
| 50 to 250 people | The same ones, plus 36 (dashboard with commentary) and 13 (root cause of inquiries) | A cross-functional owner with support from IT | Rolling out across all seven departments at once. Two departments per quarter is a sustainable pace |
In all three cases the sequence is the same: one department, three cases, four weeks, one metric. If the metric did not move, change the use case before you change the tool. The phase-by-phase implementation guide works through that calendar.
Want a hand applying this?
Free 30-minute working session. Uniamos is a remote AI automation agency (Austin and Mexico City) working over video calls and WhatsApp.
What data do you need before you start?
Less than you think, but organized. Almost every failed attempt comes down to the base material not existing, not to the tool.
- A current price list and commercial terms, in a single dated file.
- Between 10 and 15 real examples of what you consider a good result in each department: a good email, a good quote, a good summary.
- Your business taxonomy: CRM stages, ticket types, product categories, with their allowed values written down.
- An exportable history: sales, tickets or purchases from the past 12 months, even if it is just a spreadsheet.
- The rules that are not negotiable: what we promise, what we do not, what deadlines we can commit to.
All of that lives better in a shared space than in one person's inbox. Projects have their own knowledge base and instructions, and on paid plans they use retrieval-augmented generation to expand capacity by up to ten times. To package the way your company works, skills are folders of instructions in Markdown available on every plan. We go into it in the Projects guide.
Base instructions for a department project (copy and adapt):
You work for [company], which [what it does] in [country]. This department handles [function].
Always: answer in the format I ask for, with sentences under 20 words and numbers before adjectives.
Never: invent data, prices or deadlines. If a piece of information is missing, write [MISSING: what data] and move on.
Before delivering: tell me what information would have helped you do it better.
To connect with what you already use there are native connectors to Google Drive, Gmail, Calendar, GitHub, Microsoft 365, Slack and Salesforce, and MCP connectors for everything else. On Team and Enterprise the administrator can authorize them for the whole organization.
What should you NOT delegate to AI?
There are tasks where the time saved does not justify the risk, and others where it simply is not appropriate. This table is worth pasting into your internal policy.
| Do not delegate | Why | What you can do instead |
|---|---|---|
| The final decision to hire, fire or discipline | It affects a person's life and requires human judgment and accountability | Organize information, prepare questions and structure the process |
| Filing tax obligations | It is done on the official platforms and the responsibility belongs to the company and its accountant | Review, reconcile and prepare documentation before filing |
| Legal advice or opinions | It requires a licensed professional and knowledge of the specific case | Read a contract, extract quoted clauses and arrive prepared to the consultation |
| Any figure that leaves the company unverified | A misquoted number in a proposal or a report costs credibility or money | Draft the document flagging the gaps for a person to fill in |
| Health, safety or technical compliance assessments | There is regulation and professional liability involved | Summarize the applicable rule and build the checklist the specialist will review |
| Sensitive communications without human review | Bad news written badly does more damage than the news itself | Prepare drafts and anticipate the questions the recipient will ask |
| Personal data you do not need for the task | Less data exposed means less risk, and data protection regulation requires it | Work with the minimum information, already anonymized |
A short and necessary disclaimer: this is not legal or tax advice. In Mexico and in Paraguay, the handling of personal data and tax obligations are governed by their own regulations and are worth checking with your advisor. On electronic invoicing we have material for Mexico and Paraguay, and on customer data, Law 7593.
How do you know whether this is working?
With one metric per department, measured before you start. Without a baseline, any improvement is just an impression.
| Department | Simple metric | How to measure it |
|---|---|---|
| Sales | Minutes between the end of a call and the follow-up going out | Timestamp in the CRM or in the sent email |
| Customer service | First response time and number of repeat inquiries | A report from the ticketing system, or a manual count one week a month |
| Marketing | Pieces published per month against the calendar | The calendar itself, reviewed on the last Friday of the month |
| Administration | Days the monthly close takes | Start date and delivery date to the accountant |
| Purchasing | Hours per quote comparison | A simple log kept by the purchasing lead |
| Human resources | Days from posting the opening to the final shortlist | Dates from the process itself |
| Leadership | Agreements with an owner and a date per meeting | The minutes, counting how many agreements ended up with no owner |
Four weeks is a reasonable window to see movement in an operational metric. If it does not move, review the use case before the tool: almost always, the task chosen is one that is not done every week. The adoption and return guide goes into detail.
Tricks that are not in the manual
The shortcuts that separate a team that tries AI for two weeks from one that actually adopts it.
- Pick weekly tasks, not annual ones. A case that happens once a quarter never builds a habit. Start with what repeats every week, even if it looks minor.
- Upload the file, do not describe it. Anthropic documents file uploads and the creation and editing of spreadsheets, documents and presentations. Working on the real file saves half the work.
- One project per department, not one per person. Knowledge that lives in a personal chat disappears when that person goes on vacation.
- Always ask for the list of what is missing. Before the output, ask it to list the data it would need and does not have. That prevents 80% of invented results.
- Ban filling gaps, in writing. The instruction to write a marker such as [MISSING: what data] instead of filling it in belongs in the project instructions, not in every message.
- Start with the task the team hates. Adoption is not won with the most profitable task, it is won with the most annoying one: that is where people pitch in on their own.
- Save the brief, not the output. A shared document with the ten briefs that work is worth more than a hundred scattered conversations.
- Strip out personal data you do not need. Names, ID numbers, phone numbers and addresses that are not useful for the task get deleted before pasting. It removes the problem at the root.
- Turn off chat rating if you handle sensitive data. On Team and Enterprise the owner can disable feedback submission in the organization settings: that is the route by which a conversation could be retained for up to five years, according to the privacy documentation.
- Use the model that fits the task. The official models page, accessed on September 19, 2026, recommends starting with Claude Opus 5; Sonnet 5 balances speed and intelligence, and Haiku 4.5 is the fastest, which is usually enough for short, repetitive tasks.
- Take advantage of the free training. Claude Academy publishes courses and tutorials at no cost. They do not grant a credential and there is no official Anthropic certification: they are common ground for the team, not a qualification.
- Review once a quarter, not every week. If the instructions change all the time, the team stops trusting them.
What to do this week
Five days that leave one thing working, rather than seven half-done.
- Monday. Pick one department and three cases from this list. Write down that department's metric and its value today. Without that number, do not start.
- Tuesday. Gather the base material: price list or taxonomy, plus 10 to 15 examples of what you consider a good result. Anonymize whatever needs it.
- Wednesday. Create a project for the department, upload the material and write the instructions: how you write, what you never promise, and what to do when information is missing.
- Thursday. Run the three cases on real work from last week, whose outcome you already know, and compare. Correct the instructions, not the outputs.
- Friday. Half an hour with the department team: what gets done, what never gets delegated, and who reviews before anything goes out. Schedule the measurement four weeks out.
Checklist before scaling to another department
- Did the department's metric move within four weeks?
- Are the three cases being used by more than one person?
- Is it written down which data never gets pasted into a conversation?
- Is there a named owner for the department's project?
- Do the instructions include the rule about flagging gaps instead of inventing?
- Does someone review everything that goes out to a customer or an employee?
- Has the plan you pay for been reviewed against the data you are handling?
- Is the quarterly review of the base material scheduled?
When you want to move from scattered cases to a program with a calendar, the next step is the eight-week training program, and the ground rules are worth setting with an internal AI use policy. If you would rather do it with support, at Uniamos we work through these cases with each department's real data.
Frequently asked questions
How many use cases should you start with at once?
Three, in a single department. More than three scatters attention and none of them turns into a habit. When those three work without anyone having to be reminded, you add the next block. Two departments per quarter is a sustainable pace even in companies of 50 to 250 people.
Are the time savings in this article measured data?
No. They are estimates to calibrate expectations, based on our training experience, and they are declared as such in each sheet. They depend on team size, data quality and practice. The only reliable figure is the one you measure yourself: note the baseline before you start and measure again at four weeks.
What plan do I need if I am going to handle customer data?
A commercial one. Anthropic documents that in Team, Enterprise and API your inputs and outputs are not used to train models by default, whereas in Free, Pro and Max that depends on a user setting. On top of that, in Team and Enterprise the owner can disable feedback submission in the organization settings. As of September 2026 the standard Team seat costs 20 dollars per month billed annually or 25 per month, with a two-seat minimum; confirm on their site.
Can I use AI to file taxes or electronic invoices?
Not to file them. Filing happens on each country's official platforms and the responsibility belongs to the company and its accountant. Where it does add value is in the preparation: reconciling reports, checking invoices against purchase orders, and building the list of what is missing before the close. This is not tax advice.
Is it a good idea to use it to screen resumes?
For summarizing, yes; for rejecting, no. The sensible approach is to ask for a summary of each application against objective requirements and to have it flag what it cannot verify, leaving the selection to a person. It is also worth stripping from the files any data that is not relevant to the role, because that reduces the risk of bias and of unnecessary exposure of personal data.
What about personal data in Mexico and Paraguay?
It is governed by each country's data protection regulations, and this is not legal advice. The practical rule you can apply today is minimization: work only with the information needed for the task and anonymize the rest. For the Paraguayan case we have a dedicated article on Law 7593.
Do I need to integrate it with the ERP or the CRM from day one?
No. Most of the 42 cases work by copying and pasting or uploading a file. Integrations make sense once the use case is established and the bottleneck becomes moving data around. There are native connectors to the usual tools and custom MCP connectors for the rest, which on Team and Enterprise the administrator authorizes for the whole organization.
What do I do if the team does not use it?
Almost always it is because the case chosen was not a real pain or does not happen every week. Change the case before you change the tool, and start with the task the team hates most: that is where adoption happens on its own. The second most common cause is that nobody owns keeping the base material up to date.
What is the most expensive mistake people make at the start?
Letting a figure that nobody verified leave the company. You avoid it with two written rules: that the instruction requires flagging gaps instead of filling them, and that a person reviews everything addressed to a customer, an employee or an authority. Neither one costs time; both prevent expensive trouble.
Related articles
- How to implement Claude in your company, phase by phase
- Claude prompts and tricks for sales teams
- Projects: your company's shared brain
- An internal AI use and governance policy
- Measuring adoption and return on AI training
- An eight-week AI training program for teams
- Process automation with AI in Paraguay
- More from the blog