Buying AI licenses is easy, cheap and fast. Having your company work differently six months later is another matter. Between those two sentences there's a project, and that project is the one almost nobody does. This guide is the full plan: diagnostic, plan selection, governance, phases from day 0 to day 90, integration with what you already use, security and personal data in Mexico and Paraguay, training, measurement and mistakes.
It's written for a company of 15 to 300 people, with an ERP nobody wants to touch, a half-filled CRM, WhatsApp as the main sales channel and people already using AI on their own. The prices and limits cited come from official Anthropic pages consulted in September 2026, with a link to the source.
Short answer: Implementing Claude at a company in Mexico or Paraguay isn't handing out accounts: it's choosing a plan based on the data regime you need, naming someone accountable, picking two or three measurable use cases, training a pilot group, connecting the tool to your email, CRM and WhatsApp, setting a usage policy and measuring real adoption. With a phased 90-day plan, the project holds up; without one, it fizzles out in six weeks.
- Why do so many AI implementations fail at companies in Mexico and Paraguay?
- How do you know if your company is ready? The pre-diagnostic in one afternoon
- Which Claude plan do you need: Free, Pro, Max, Team or Enterprise?
- Roles and governance: who decides what in an AI rollout?
- Phase 0 and days 1 to 30: prepare the ground and run the pilot
- Days 31 to 60: extend to the people already asking to get in
- Days 61 to 90: consolidate, automate and decide the next investment
- Which pilot use cases actually work, and which are better left for later?
- How does Claude integrate with email, the CRM, the ERP and WhatsApp?
- Security and personal data: what Mexico requires and what Paraguay requires
- Team training: how you go from «trying it» to «knowing how to use it»
- How do you measure whether the implementation is working?
- The classic mistakes you'll make if nobody warns you
- Tricks they don't put in the manual
- What to do this week: seven concrete steps
Why do so many AI implementations fail at companies in Mexico and Paraguay?
They fail because a tool gets bought before a problem is defined, access is handed out without training and nobody stays on the hook for the result. Technology is almost never the bottleneck: the bottleneck is treating AI as a software purchase instead of a change in how people work.
If you've watched a pilot die, it probably died like this: someone in management tried the tool over a weekend, bought twenty licenses on Monday and sent an email with the subject line "we now have AI." Three months later it was canceled "because we didn't see the return." The problem wasn't the tool: nobody defined what working meant, nobody taught people how to ask for things properly and nobody measured anything except the invoice.
| Symptom you see | Real cause | Antidote |
|---|---|---|
| "We tried it and it doesn't work for what we do" | It was tested on the hardest case, with no context or instructions | Start with repetitive, verifiable text tasks |
| Heavy use in week one, flat by week four | Novelty without habit: it wasn't anchored to any weekly process | Tie each use case to a fixed slot on the calendar |
| "It makes up data" | People ask it for information it doesn't have instead of giving it the document | Your own knowledge base and a requirement to cite the source |
| Legal halts the project halfway through | It started without deciding the data regime or the plan | Decide plan and policy before the pilot |
| Nobody knows whether it helped | There's no baseline: nobody measured how long the task used to take | Time three tasks before handing out the first license |
There's one local factor that gets underestimated: most commercial work in Mexico and Paraguay happens on WhatsApp, not in the CRM. A project that only touches the official CRM is optimizing a fraction of the real operation, as the guide to AI agents on WhatsApp Business explains. The last reason is the most human one: the team is afraid the tool will replace them, so they use it just enough and don't share what they discover. Without an honest conversation about that up front, adoption stalls at the usual 20%.
How do you know if your company is ready? The pre-diagnostic in one afternoon
You're ready if you can name three tasks that eat hours every week, you know who does them and how long they take, you have someone willing to own the project and you know what information can and cannot leave the company. If any of those four things is missing, fix it before you buy anything.
It doesn't take six weeks of consulting: one afternoon, one room, three or four people from different areas, and these questions answered in writing.
Diagnostic checklist
- Which three repetitive text tasks eat the most hours per month?
- How long does each one take today, measured rather than estimated, and who does them?
- Can someone verify the result in under five minutes?
- Where does the necessary information live: email, Drive, SharePoint, the ERP, somebody's head?
- What percentage of your sales conversations happen on WhatsApp?
- Is your CRM clean enough to be worth connecting?
- Does your ERP have a documented API, or do you get at it through manual exports?
- Do those tasks involve personal data from customers or employees? Is any of it sensitive?
- Who signs off that a given piece of information can be uploaded to an external tool?
- Do you have an identity provider (Google Workspace, Microsoft Entra ID, Okta)?
- How many people already use AI with personal accounts?
- Who can commit four to six hours a week for three months?
- What would have to happen for you to say "this stays" in 90 days?
The question about personal accounts is the one that surprises management most: at almost any company with more than twenty people, AI is already being used off the books. That's an argument for rolling out something official, not against it: that information is already leaving the building, just without policy, without logs and without a contract.
| Maturity level | How to recognize it | The right first move |
|---|---|---|
| 1. Curiosity | Nobody uses AI systematically; there are unresolved legal questions | One or two licenses, one use case and a one-page policy |
| 2. Scattered use | Several people with their own accounts and uneven results | Team plan, shared knowledge bases and adoption metrics |
| 3. Established use | Use cases in production and demand for automation | Formal governance, integrations and cost control per use case |
Copyable template — diagnostic record (one page).
Company and date:
Three candidate tasks (description, area, current hours/month, who validates):
Information needed and where it lives:
Personal data involved (yes/no, whose, category):
Project owner (role) and weekly hours committed:
Sponsor who signs off on decisions (role):
Success criterion at 90 days, in one measurable sentence:
Pilot review date:
If you'd rather have someone from outside run the diagnostic and the initial training, Uniamos does that work; you can also do it with this article and one afternoon blocked off. What doesn't work is skipping it.
Which Claude plan do you need: Free, Pro, Max, Team or Enterprise?
To try it out, Pro. To work as a team with customer data, Team or Enterprise: those are the commercial plans, and on them Anthropic states that by default it doesn't use your inputs or outputs to train models. The difference between a consumer plan and a commercial one isn't only features, it's the default data policy.
This is the table published at claude.com/pricing. Amounts are in US dollars, before tax, and the page itself warns that they vary by region. As of September 2026; confirm on their site before deciding.
| Plan | Annual price | Monthly price | Min./max. seats | Who it's for |
|---|---|---|---|---|
| Free | $0 | $0 | 1 | Poking around. Maximum 5 projects |
| Pro | $17/month billed annually | $20/month | 1 | Whoever is going to lead the pilot |
| Max 5x | — | $100/month | 1 | Heavy individual use |
| Max 20x | — | $200/month | 1 | Very heavy individual use |
| Team Standard | $20/seat/month | $25/seat/month | min. 2, max. 150 | The natural plan for a small or mid-sized business |
| Team Premium | $100/seat/month | $125/seat/month | min. 2, max. 150 | High-usage profiles |
| Enterprise self-serve | $20/seat/month + usage at API rates | — | min. 20, annual | When you need audit logs and retention |
| Enterprise via sales | Custom pricing | — | min. 50, annual | Large or regulated organizations |
Sources by plan: Pro, Team and Enterprise. Max prices are for web subscriptions; they may differ in mobile app stores.
What actually changes between plans
| Capability | Free | Pro | Max | Team | Enterprise |
|---|---|---|---|---|---|
| Projects | Max. 5 | Yes | Yes | Yes | Yes |
| Enterprise search, SSO, central billing, admin controls | — | — | — | Yes | Yes |
| No training on your content by default | — | — | — | Yes | Yes |
| Custom roles, SCIM, audit logs, compliance API, configurable retention | — | — | — | — | Yes |
| Customer-managed encryption keys | — | — | — | — | Yes |
Usage limits, no nonsense
Anthropic doesn't publish an absolute number of messages: it publishes multiples and windows. Pro has a session that resets every five hours plus a weekly limit; Max 5x and Max 20x are five and twenty times that allowance; a Team Standard seat is 1.25 times Pro's and Premium is 6.25 times, and those limits are per member, not pooled across the team. Usage-based Enterprise has no plan or seat limits. If someone promises you "so many messages a day," they're making it up: the reference is Settings > Usage.
The decision rule in three lines
If you only want to know whether this is worth it, one Pro license for a month: it costs less than the meeting you'd hold to debate it. If you're going to work with customer, pricing or HR information, Team at minimum, because the default data regime is different. If you need audit logs, configurable retention, SCIM or custom roles, Enterprise: those controls don't exist in Team no matter how much you pay.
The cost almost nobody calculates. Licenses are the floor, not the ceiling. If you automate through the API in phase 3, a second cost shows up for token usage, with its own rates and cache discounts. Budget the two line items separately from the start and review usage every month.
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.
Roles and governance: who decides what in an AI rollout?
You need four figures: a sponsor who signs off on decisions and budget, an operational owner with four to six hours a week, two or three champions per area who answer questions in the hallway, and a contact in IT or legal for access and data. At companies under fifty people the sponsor and the owner can be the same person; the champions, never.
There are two layers you shouldn't mix: platform roles, defined by Anthropic, and your company's roles, defined by you. The first are covered in roles and permissions.
| Platform role | What it can do | Who should have it |
|---|---|---|
| User | Their own chats and projects | Everyone by default |
| Admin | Invite and remove members; analytics only on Enterprise | One or two people from IT or operations |
| Owner | Billing, roles, integrations, security and audit logs on Enterprise | The project owner and a backup |
| Primary Owner | Everything, including provisioning seats, requesting exports and transferring ownership. Only one per organization | A person identified by their role, with a written stand-in |
| Custom | No permissions by default; everything is defined by group. Enterprise only | Organizations with separation of duties required by audit |
Three day-one decisions. First: centralized identity, with SAML single sign-on against Okta, Microsoft Entra ID, Google, OneLogin, JumpCloud or Duo; it requires domain verification via a TXT record and lets you enforce its use. Second: automatic onboarding and offboarding with JIT or SCIM provisioning, so whoever leaves loses access without anyone having to remember. Third: who the Primary Owner is; there's only one, and if it's the partner who travels three weeks a month you've got a single point of failure on billing and data exports.
| Internal figure | Time commitment | Sign it isn't working |
|---|---|---|
| Sponsor: signs off budget and unblocks areas | 2 h/month | Never shows up at a single demo |
| Adoption owner: plan, training, metrics, knowledge bases | 4-6 h/week, 90 days | Only spends time installing licenses |
| Champions by area: translate use cases and spot resistance | 1-2 h/week | They're all from the same department |
| IT: SSO, access, connectors, integrations | By milestone | Finds out once there are already twenty licenses |
| Legal: usage policy, personal data, contracts | By milestone | Says no to everything because nobody explained the controls |
Documented governance fits in three documents: a one- or two-page usage policy, a use case register with owner and status, and an inventory of what information may go in. The full templates are in the guide to internal AI usage policy and governance.
Copyable template — governance record in eight lines.
1. Sponsor (role) · 2. Adoption owner and hours committed · 3. Champions by area · 4. Primary Owner and backup · 5. Plans and number of seats · 6. Information that may NOT be uploaded, as a closed list · 7. Authorized connectors and who approves them · 8. Review cadence and the metrics that go to the committee.
Phase 0 and days 1 to 30: prepare the ground and run the pilot
In phase 0 you don't use the tool: you pick two or three use cases, time how long they take today, write the usage policy, decide the plan and select the pilot group. In the first 30 days, eight to twelve people should be handling those tasks better and faster, and it should be documented.
Phase 0: the two weeks before you hand out the first license
- Pick two or three use cases, not ten. A good pilot case happens every week, is done by more than one person, produces text or analysis, can be verified in five minutes and doesn't require sensitive data in its first version. If it fails two of those five criteria, save it for phase 2.
- Time the baseline: minutes and repetitions logged for a week by whoever does the task today.
- Decide the data regime before the plan. If the pilot touches personal data, you're going to Team or Enterprise from the start. What you can't do is start with personal accounts "just to try it" and discover in month two that half the team uploaded the customer database.
- Write the one-page usage policy, with three lists: what's allowed, what's never allowed and what to ask about first. It gets published the same day as the first license, and that same day you schedule the pilot sessions and the day-30 review.
- Choose the pilot group with your head: eight to twelve people, from at least three areas, mixing enthusiasts with useful skeptics, who turn out to be the best trainers in phase 2. Don't build it only from the youngest team: it'll end up seen as "a thing for the kids."
- Prepare the first knowledge base: five to ten documents that explain how the work is done well — catalog with current prices, proposal template, discount policy, real frequently asked questions. No whole folders. How to structure them is in the guide to Projects and knowledge bases.
Week by week, from day 1 to day 30
- Week 1. Two-hour kickoff session: what it's for, what it isn't for, what never gets uploaded. Hand out access and verify that everyone can get in; there are always two who can't and end up left out forever. First shared project and a single channel for questions.
- Week 2. Each person repeats their use case three times without hand-holding. The three best prompts from each area get collected in a shared document: that document is the most valuable asset of the month. One hour to review what went wrong, which is almost always missing context.
- Week 3. The criterion for "this is ready to send to the customer" gets defined in writing and verification is introduced: nothing goes out without a person reading it and taking responsibility for it. The missing documents get added to the knowledge base.
- Week 4. The numbers get closed out, two before-and-after examples get documented, and the decision gets made with the sponsor: extend, adjust or stop.
| Day-30 indicator | Good sign | Warning sign |
|---|---|---|
| Weekly use by the pilot group | 7 out of 10 use it three days a week or more | Fewer than half open it during the week |
| Repetitions per use case | Eight or more | It was tried once and dropped |
| Time per task | Measurable, sustained reduction, even if it's 20% | No data, because nobody measured |
| Prompts saved | Fifteen or more, in a shared document | It all lives in two people's heads |
A method for week 1: the official documentation recommends treating the model like "a brilliant but new employee who doesn't know your norms or your workflows", with one golden rule: show your instruction to a colleague with little context; if they get confused, the model will too. It's in Anthropic's prompting best practices.
Days 31 to 60: extend to the people already asking to get in
The second month extends to the areas that are already asking for it, not to the whole company at once. The criterion is spontaneous demand: if an area didn't ask a single question during the pilot, don't bring it in yet. Training shifts from individual hand-holding to per-area sessions with that area's own use cases.
This is the month most projects fall apart: the tap gets opened, victory gets declared, and then comes the week-six collapse. What does work:
- A second wave of 20 to 40 people, grouped by area rather than by hierarchy.
- One 90-minute session per area, with that area's use cases and real examples from the pilot. Generic training doesn't land; "your own work" training does.
- One shared project per area, with its own knowledge base and instructions, and a prompt library with someone responsible for maintaining it.
- First connectors: email, calendar and document storage, with permissions reviewed.
- A review of the usage policy using the real cases from month 1, and a 30-minute biweekly champions meeting.
The scheduling mistake that ruins month 2
Scheduling the second wave on top of the accounting close, high season or a big trade show. If your company exhibits in October or November, those weeks don't exist for internal training. What you can do is turn the trade show into the use case: preparing materials, qualifying contacts and following up afterward. The method is in the guide to trade shows and lead capture and in the one on post-show follow-up in 72 hours.
How to avoid the week-six plateau
- A fixed ritual and visibility: a weekly slot set aside for using it — Monday sales prep, Friday recap — and a biweekly message with two examples of something that went well, citing the role rather than the person.
- Remove friction: if someone takes longer with the tool than without it, they're missing saved context; you fix that with a project, not a telling-off.
- Use cases by area, not generic ones: the full catalog is in Claude use cases by department.
Copyable template — per-area session script (90 minutes).
0-10: what changed for this area, with two real examples.
10-25: the three use cases we'll work on today and why those.
25-55: guided exercise, each person with this week's actual work on screen.
55-70: what went wrong and live correction of instructions.
70-80: which documents are missing from the area's knowledge base; note them down and assign an owner.
80-90: individual written commitment: "this week I'm going to use it for ___, ___ times."
Days 61 to 90: consolidate, automate and decide the next investment
The third month turns individual use into processes: the three or four tasks that already work get standardized, documented as procedures, and whatever has enough volume gets automated; then you decide with numbers whether to expand the investment, hold it steady or cut it.
Standardize what survived. Every use case that lasted two months deserves a written procedure: who does it, with which project and instructions, who validates it and in how long. Without that, the knowledge walks out when the person does. Then separate what's done by hand from what gets automated: the criterion isn't ambition, it's volume.
| Situation | Right solution | Why |
|---|---|---|
| Fewer than 20 repetitions a month, shifting criteria | A person with a project and instructions | Automating costs more than it saves |
| 20 to 200 a month, stable criteria | Shared project, templates and connectors | The volume justifies standardizing, not programming |
| More than 200 a month, structured input | API integration or an agent | Here the development really does pay for itself |
| Inbound customer conversations at all hours | An agent in the channel they're already in | The cost of not replying exceeds the cost of the agent |
For that leap, the reference is the process automation page and, if you operate in Paraguay, the guide to AI process automation in Paraguay.
Review the real cost. At day 90 you have three figures you didn't have on day 1: licenses, API usage and measured hours freed up. If the project doesn't pay for itself in the first quarter, it's almost always because licenses got handed out to people with no assigned use case; the fix isn't canceling, it's reassigning seats. And decide the next stretch among four options: broaden coverage, go deeper into more complex use cases, automate the high-volume ones, or consolidate without growing — which is legitimate if month 2 left a mess behind.
Copyable template — day-90 report for the committee (one page).
1. What we did: use cases, people, training sessions.
2. Numbers: hours/month freed up per use case, with baseline and measurement method.
3. Cost for the quarter: licenses + usage + internal hours.
4. What worked and what didn't, with one example of each.
5. Open risks (data, dependency, quality) and how they're controlled.
6. Proposal for next quarter, with three options and their cost.
7. The decision being asked of the committee, in one sentence.
Which pilot use cases actually work, and which are better left for later?
What works: cases with text in, text out, high frequency and easy verification — sales proposals, meeting summaries, bid responses, lead qualification and meeting prep. What works badly, at first: anything that requires live ERP data, decisions with legal impact, or information the company never wrote down.
| Use case | Area | Why it works | What you need |
|---|---|---|---|
| Sales proposal from visit notes | Sales | High frequency, repeatable format, immediate validation | Approved template, catalog and discount policy |
| Meeting summaries and commitments | All | Visible time savings from day one | Notes or a transcript and a criterion for what counts as an agreement |
| Lead qualification and prioritization | Sales and marketing | Stable criteria and high volume after campaigns or trade shows | A written definition of a good lead: see B2B lead generation with AI |
| Reconciliation and explaining variances | Finance | Analytical work on already-exported data | Clean exports: see finance |
| Support replies and FAQs | Customer service | Extremely high repetition and verifiable criteria | Ticket history and a catalog of responses |
The use cases worth postponing
- Live queries against the ERP. Without a documented API, the pilot turns into a systems project and your quarter is gone: first read how to choose and integrate an ERP with a solid API.
- Decisions with legal or employment consequences and final tax calculations. AI prepares the material, explains it and spots inconsistencies; the decision and the signature are human. In Mexico, the context is in tax automation and SAT platforms.
- Anything the company never wrote down. If the criterion for a "good proposal" lives only in the sales director's head, write it down first and automate second.
- Unsupervised customer service. Start with drafts that a person approves: see chatbots and AI agents.
A selection trick. Sort your candidates into two columns — weekly frequency and ease of verification — and start with the high-high quadrant, even if it's the least glamorous. The flashy use case that happens once a quarter doesn't build a habit, and without habit there's no adoption.
How does Claude integrate with email, the CRM, the ERP and WhatsApp?
There are three levels: native connectors for email, calendar and storage; custom connectors via MCP for your own systems; and API integration when volume justifies automating outright. The rule is to start with connectors, measure, and only build code for what repeats hundreds of times a month.
Level 1, native connectors. Connectors let you work with Google Drive, Gmail, Calendar, GitHub, Microsoft 365, Slack and Salesforce with no development. Two warnings: a connector inherits the permissions of the account that authorizes it, so an account with access to everything turns the tool into access to everything; and on Team or Enterprise it's best to authorize them centrally.
Level 2, MCP. For your own systems there's the open Model Context Protocol standard: custom connectors using remote MCP. It lets you query your catalog, inventory or ticketing system without copy-pasting: it requires development, but the scope is bounded and reusable.
Level 3, API and agents. When the task happens hundreds or thousands of times a month and the input is structured, it's time to integrate. That's where a little-known cost factor shows up: prompt caching. Writing to a cache with a five-minute lifetime costs 1.25 times the base input price and reading from cache costs 0.1 times, so it pays for itself after a single read. If your integration sends the same catalog on every call, that changes the bill: prompt caching documentation and the integrations and APIs page.
The special case of WhatsApp
In Mexico and Paraguay, WhatsApp isn't just another channel: for many companies it's the channel. The integration happens through the WhatsApp Business API, not through the sales rep's phone, and that distinction determines who owns the contact database and what traceability you have. The full walkthrough is in what an AI agent on WhatsApp is, what it costs to set one up and WhatsApp Business with AI.
| System | Prerequisite | When to connect it |
|---|---|---|
| Email and calendar | Corporate accounts and reviewed permissions | Month 1-2 |
| Document storage | Tidy folders | Month 1-2 |
| CRM | Clean data; if the CRM is empty, so is the result | Month 2-3 |
| ERP | Documented API and read-only to start | Month 3 onward |
| WhatsApp Business | Official API, approved templates, data policy | When volume justifies it |
The golden rule of integration. Connect in read-only first. Let the tool read the CRM for a month before you let it write to it: read errors fix themselves, write errors stay in your database and get discovered by a customer.
Security and personal data: what Mexico requires and what Paraguay requires
In Mexico, the governing law is the Ley Federal de Protección de Datos Personales en Posesión de los Particulares (Federal Law on the Protection of Personal Data Held by Private Parties), published on March 20, 2025 and in force since March 21, 2025, with authority transferred from the now-dissolved INAI to the Secretaría Anticorrupción y Buen Gobierno (Ministry of Anti-Corruption and Good Governance). In Paraguay, the governing law is Ley 7593/2025 (Personal Data Protection Law), enacted on November 27, 2025, with a 24-month compliance window. Using AI doesn't create new obligations: it means applying the ones you already had.
This section is for general guidance and doesn't constitute legal advice. Consult your lawyer or compliance officer before deciding.
Mexico
The law keeps the familiar architecture — privacy notice, consent, and rights of access, rectification, cancellation and objection — with two changes that matter: the change of authority, now the Secretaría Anticorrupción y Buen Gobierno, and tighter rules on using data for purposes other than those disclosed, so changing the use of data you've already collected requires fresh consent. The text in force is on the Chamber of Deputies portal. Translated to your project: if your privacy notice says customer data is used to manage the commercial relationship, processing it with AI to prepare a proposal fits within that purpose; handing it to a third party for a different purpose does not.
Paraguay
Ley 7593/2025 sets legal bases for processing; prior, freely given, informed, unambiguous and revocable consent; data subject rights; written contracts with processors; incident notification; and impact assessments for high-risk processing, and it creates a national agency as the enforcement authority. Its 24-month compliance window makes 2026 and 2027 the period to get your house in order. Practical guide: using AI with customer data in Paraguay and the e-Kuatia rollout schedule by group.
What the platform gives you and what you have to bring
- Training. On commercial products — Team, Enterprise and the API — the published policy says that by default your inputs and outputs aren't used to train models; the exception is explicit consent, for example when you rate a conversation: data usage policy.
- Chat ratings. When you rate a conversation, the related conversation is retained for up to five years, disassociated from identifiers. Team and Enterprise owners can turn it off entirely in the organization settings: it's the first control to switch on.
- Retention. Only on Enterprise can you set your own retention periods for chats and projects, with a 30-day minimum; by default data is kept indefinitely: custom retention.
- Audit logs. Audit logs are Enterprise-only, cover the last 180 days on a rolling window and don't include the title or content of chats or projects. If you need history, export them.
- Certifications. SOC 2 Type I and II, ISO/IEC 27001:2022 and ISO/IEC 42001:2023, plus HIPAA-ready configuration for those who qualify; the stated scope is the commercial products: certifications and trust portal.
- Data processing agreement. The addendum is accepted automatically when you accept the commercial terms, with no separate signature: DPA text.
- Residency. If a customer requires a specific region: on the first-party API, today the workspace supports only the United States and inference supports "global" or "United States."
What no platform does for you: deciding what information goes in, training people to respect that, and reviewing the authorized connectors.
Copyable template — three-level information classification.
Green (goes in without asking): public material, published catalogs, internal templates with no customer data.
Amber (goes in with judgment and anonymization): emails with no sensitive data, proposals with prices, process documentation, aggregated analytics.
Red (doesn't go in without written approval): sensitive personal data, health data, credentials, contracts under confidentiality, complete customer databases.
Person responsible for answering questions (role): ___ · Response time: 24 h.
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.
Team training: how you go from «trying it» to «knowing how to use it»
Effective training isn't a two-hour course: it's short sessions with each person's real work on screen, repeated over six or eight weeks, with a shared prompt library and someone who answers questions the same day. The official courses are free and make a good foundation, but they don't replace practice with your own documents.
Worth clarifying before someone tries to sell it to you: Anthropic doesn't offer any professional certification program. The Claude Academy courses are free and don't grant a credential. If someone offers to "certify your team in Claude," they're using a word that doesn't exist. What does exist is public, high-quality material: a roughly four-hour fundamentals course with the four D's framework — delegation, description, discernment and diligence — another on capabilities and limitations, and short tutorials, including an admin one for IT leads.
| Moment | Format | Duration | Content |
|---|---|---|---|
| Kickoff | Joint session | 2 h | What it is and isn't, usage policy, first three use cases |
| Weeks 1-4 | Per-area session | 90 min | Real work on screen and live correction |
| Weekly and biweekly | Open office hours and champions circle | 30 min | Loose questions; findings and additions to the library |
| Monthly | Open demo | 45 min | Two real cases, presented by the people who use them |
Three principles: never train with generic examples — everyone arrives with a real task from that week; teach people to give context before writing pretty prompts, because that's where most bad results come from; and turn every finding into an asset, uploading it to the shared library the same day.
The full calendar, session by session, is in the eight-week AI training program; the material specific to the sales team is in Claude prompts for sales teams.
How do you measure whether the implementation is working?
You measure with three families: adoption (how many people use it and how often), productivity (time per task against the baseline) and business (response time, proposals sent, cycle time). The metric that doesn't help is the number of licenses purchased, and it's exactly the one almost everyone reports.
| Family | Indicator | How you get it | 90-day target |
|---|---|---|---|
| Adoption | Weekly active users over assigned licenses | Organization analytics or a survey | 60-70% |
| Adoption | Days of use per week; areas with a use case in production | Same source and the use case register | 3 and 3 |
| Productivity | Minutes per task, before and after | Manual measurement on a sample | −25 to −40% on the pilot use cases |
| Quality | Outputs that pass review on the first try | The validator's mark | 80% or more |
| Business | Response time to a quote request | CRM or email | Measurable reduction |
| Business | Leads contacted within 24 h after an event | CRM | 90% or more |
| Cost | Total cost per use case per month | Licenses + usage + internal hours | Known and stable |
The most expensive measurement mistake is reporting "hours saved" times the hourly cost and presenting it as accounting savings. If nobody stopped hiring, that saving doesn't show up in any financial statement, and the first CFO who notices will kill the project's credibility. The honest framing is capacity freed up: this many hours a month now going to something else — and you name that something else.
Cadence: weekly and internal for usage and blockers; monthly to the sponsor, one page; quarterly to the committee, with numbers, cost and the decision being requested. If you build a dashboard, three well-measured indicators are worth more than fifteen estimated ones: see dashboards and analytics and how to measure adoption and return.
Copyable template — monthly report in six lines.
Month: ___ · Active licenses: ___ · Weekly active users: ___ (___%)
Use cases in production: ___ · New use cases this month: ___
Time per task on the main use case: before ___ min → now ___ min
Quality: ___% of outputs approved on the first try
Cost for the month: licenses $___ + usage $___ = $___
Main blocker and what I need to resolve it:
The classic mistakes you'll make if nobody warns you
The mistakes that kill projects most often: buying before diagnosing, handing out licenses with no assigned use cases, training only once, not measuring the baseline, starting with the hardest case, ignoring WhatsApp, leaving legal until the end, not naming an owner, confusing a pilot with production, and not revoking access.
- Buying twenty licenses on day one. A license with no assigned use case is pure expense and it kills the perception of value.
- Handing out access without assigning a use case to each person. "Use it for whatever you want" reliably produces it being used for nothing.
- Training only once. A two-hour session generates enthusiasm for four days; habit needs six or eight weeks.
- Not measuring beforehand and starting with the most ambitious use case. Without a baseline you'll be arguing about feelings on day 90, and the use case that wins management over is the one that sinks the pilot.
- Ignoring the channel where the customer is. If 70% of your sales conversations happen on WhatsApp and your project only touches the CRM, you're optimizing 30%.
- Leaving legal until the end. Two meetings at the start avoid a hard stop in month two.
- Not naming an owner with protected hours. "The committee is handling it" means nobody is handling it.
- Confusing a pilot with production. A process has a written procedure, an owner, a quality standard and periodic review; a successful pilot does not.
- Not revoking access. Without automatic provisioning, the former employee keeps access to the knowledge base: any auditor's favorite finding.
- Banning instead of channeling and inflating results: the first makes usage invisible, the second gets paid for with the project's cancellation three months later.
Tricks they don't put in the manual
The shortcuts you pick up while accompanying implementations, which almost never show up in the documentation. Actionable today.
- Ask for the plan before the result. "Before you write anything, tell me how you'll approach this and what information you're missing." You correct the approach in thirty seconds instead of throwing out three pages. It's the trick that saves the most time and the one fewest people use.
- Long documents at the top, question at the end. The official documentation says this for long contexts: putting queries at the end can improve quality by up to 30% in their tests. If you paste a 40-page tender document, the instruction goes after the document.
- Ask for quotes before conclusions. "First pull the verbatim sentences from the document that support the answer; then answer." It cuts down on invented facts and gives you material to verify in ten seconds.
- Give the reason behind every rule. Instead of "don't use jargon," write "this will be read by a buyer at a family-owned company with no technical department, so avoid jargon." The official documentation uses that pattern.
- Three to five examples, never one. The official guide recommends three to five examples, varied and covering edge cases. A single example produces literal imitation; five produce judgment.
- Separate the parts with tags. Wrapping context, instructions and data in tags like <context> or <data> stops it from confusing your example with your order. It's the fastest improvement for someone who already uses it daily.
- Save the prompt that worked, not the result. The result expires in a week; the prompt gets reused a hundred times. A sheet with three columns — task, prompt, who uses it — is worth more than any course.
- Turn the meeting into input, not a summary. Instead of "summarize this meeting," ask for "list the commitments with owner and date, and the three questions left unanswered." Nobody reads the summary; the commitments list gets read. And ban a bare "make it better": teach people to say what's wrong.
- Review your draft against explicit criteria. "Tell me only what fails against these five criteria and why; don't rewrite it." You learn, and the text stays yours.
- Turn off chat ratings on day one. On Team and Enterprise, owners can switch it off in the organization settings: it's one checkbox and it eliminates at the root a path to retaining conversations for up to five years.
- Export the audit logs with a calendar reminder. They cover 180 days on a rolling window; if your policy requires a year, you won't have it without exporting.
- Reassign seats every month. Check who hasn't logged in for four weeks and move that license to whoever asked for one. Nobody takes offense and cost per active user corrects itself.
Copyable template — a prompt structure that works for almost anything.
<context> Who I am, what company, who this is for, what happened before. </context>
<material> Documents, emails or data pasted in full. </material>
<task> Exactly what I want, in one sentence. </task>
<criteria> Length, tone, what must be included, what must never be included. </criteria>
<format> How I want the output: list, table, email ready to send. </format>
Before answering, tell me what information you're missing.
What to do this week: seven concrete steps
This week, don't buy anything for the whole company. Block two hours for the diagnostic, pick three candidate tasks, measure them, decide the data regime, write the one-page usage policy, name an owner and schedule the pilot kickoff for two weeks out.
- Monday. Two hours with three people from different areas to answer the diagnostic checklist, in writing.
- Tuesday. Pick three candidate tasks and ask people to log their times for five days.
- Wednesday. Decide whether personal data will be involved in the pilot: that answer determines whether you go with a consumer plan or a commercial one. Write it down and share it with your legal lead.
- Thursday. Write the one-page usage policy with the three lists: green, amber and red.
- Friday. Name the adoption owner in writing, with committed hours, and the sponsor. Over the weekend, optional: the free Claude Academy fundamentals course is four hours and you'll speak with authority on Monday.
- The following week. Gather the five to ten documents for the first knowledge base and schedule the kickoff session. That's where day 1 begins.
The rest of the series develops each piece at this same level of detail: the knowledge base with Projects, the prompts for the sales team, the training program, measuring adoption and return, the lead generation system, trade shows and their 72-hour follow-up, the internal usage policy and the use cases by department. If you'd rather first see what can be automated at a Mexican small or mid-sized business, start with automation for SMBs in Mexico.
Frequently asked questions
How much does it cost to implement Claude at a small company in Mexico or Paraguay?
There are two parts. Licenses start at 20 dollars per seat per month on the Team plan billed annually, or 25 with monthly billing, according to claude.com/pricing as of September 2026; confirm on their site, because prices are in dollars, before tax, and vary by region. The second part is internal time: four to six hours a week from the project owner during the first quarter. Almost nobody budgets that line item, and it's the one that weighs most.
Can I start with the free plan?
To poke around, yes; to implement, no. Free caps you at five projects and belongs to the consumer plans, whose default data regime differs from the commercial ones. If you're going to work with customer information, start on Team. If you only want one person to evaluate the tool, one Pro license for a month is the cheapest investment in the whole project.
Does Anthropic train its models on my company's information?
On commercial products — Team, Enterprise and the API — the published policy says inputs and outputs aren't used to train models by default. The exception is explicit consent, for example when rating a conversation; in that case the related conversation is retained for up to five years, disassociated from identifiers. Team and Enterprise owners can disable that feature entirely, and it's worth doing on day one.
Which plan do I need if I'm asked for audit logs?
Enterprise. Audit logs don't exist in Team no matter how much you pay: they cover the last 180 days on a rolling window and don't include the title or content of chats or projects. If your policy requires more than six months of history, you'll have to export them. Configurable retention, with a 30-day minimum, is also Enterprise-only.
How long does it take to see a return?
On repetitive text tasks, the time improvement shows up within the first two or three weeks of sustained use. What takes longer is consolidation: getting the task to stop depending on one enthusiastic person and become a company procedure. That leap happens between day 60 and day 90 if there's training and an owner.
Does my ERP need an API before I can start?
Not to start. The first use cases work on text, emails and documents, not on live ERP data. Integration is a phase 3 decision, once you know which queries repeat hundreds of times a month. When you get there, having a documented API is the difference between a project of weeks and one of quarters.
How do I stop my team from uploading information they shouldn't?
With three things, in this order: a green, amber and red classification that fits on one page; a kickoff session explained with real examples; and one identified person to ask, with an answer within 24 hours. On Team or Enterprise, also centralize connector authorization and turn on single sign-on with automatic provisioning.
What does Paraguay's Ley 7593/2025 mean if I use AI with customer data?
It was enacted on November 27, 2025 and has a 24-month compliance window, which leaves 2026 and 2027 as the preparation period. It requires a legal basis for each processing activity; prior, freely given, informed and unambiguous consent where that's the basis; written contracts with processors; and incident notification. Using AI doesn't create new obligations. This is general guidance, not legal advice.
Is there an official Anthropic certification for my team?
No. Claude Academy offers free courses with no credential and Anthropic doesn't announce any professional certification program. If a vendor offers to «certify your team in Claude», they're selling a label that doesn't exist. What does have demonstrable value is an internal library of use cases and prompts, written procedures and adoption metrics.
Do I need to hire someone new to lead this?
At companies under 300 people, almost never. It works better with someone from inside who knows the processes and the people, with protected hours and outside support for the training and plan design. A new hire with no knowledge of the business usually produces a technically correct project that the sales team doesn't adopt.
Related articles
- Claude Projects: how to build your company's knowledge base
- An eight-week AI training program for teams
- Internal AI usage policy and governance
- How to measure AI adoption and return at your company
- Claude use cases by department at SMBs
- AI agents on WhatsApp Business: the definitive guide
- Process automation with AI
- More from the blog