AI agents in Paraguay: which process first

Ready for a pilot?

Tick what you already have; heavier items stall a pilot more. Nothing is sent.

Readiness for a first agent

0 % · 0 of 6 controls that applyFirst missing: process measured and written rules

Not yet: the basics for testing with real data are missing.

Higher-risk controls weigh more in the score.

We know how many supplier invoices arrive per month and how long each takes.
We have written price and quantity tolerances and who approves each exception.
There is an ERP user with limited permissions for the agent, with no access to payments.
We have 100 real invoices with their order and receipt for testing.
Every agent action is logged with date, user and version.
A named manager is accountable for the pilot.

What it costs and when it pays back

Hours freed by AI minus what it costs to run. Examples are marked; public prices link to their source.

Supplier invoices per month1,500

Minutes of work per invoice today35 min

Minutes per invoice with the agent (example)7 min

AI model

Cost per hour (PYG)

Floor: the minimum-wage hourly value (MTESS, July 2026). Typical real cost (example): ×2.5, a salary of two minimum wages plus employer contributions and the 13th month. Source

Today875 h/month

Automated175 h/month

Value of hours freed per month700 h × PYG 12,683: PYG 8,878,100
AI model public price1,500 invoices × USD 0.0010 (example: 6,000 input + 800 output tokens; gpt-6-luna at USD 0.10 / 0.50 per million) · OpenAI · API pricing: −PYG 8,542
Platform, monitoring and support exampleTypical monthly range: −PYG 1,138,894 to −PYG 2,847,235
Net savings per month: PYG 6,022,323 to PYG 7,730,664
ERP connection, tolerance rules, testing and pilot typical range · exampleOne-off, up front: PYG 28,472,350 to PYG 68,333,640
Months to pay back the investment: 3.7 to 11 months

Tokens: the unit AI providers bill for the text the model reads (input) and writes (output). OpenAI puts 100 tokens at about 75 English words.

Range: the lower figure uses low costs, the higher one high costs. Public prices checked on 11 Oct 2026. USD prices converted at PYG 5,694.47 per dollar (reference rate of 9 Oct 2026): BCP

estimated

4-week pilot

A typical plan; real timing depends on your systems.

Pilot · weeks 1 to 4

Week 1Pick and measureOne process (supplier invoices) and its baseline of time and differences.KPI: Signed baseline
Week 2Rules and permissionsWritten tolerances, a scoped ERP user and every step logged.KPI: Permissions approved
Week 3Controlled testThe agent matches 100 real invoices; accounts corrects every error.KPI: Correct matches (example target: 90%)
Week 4Scoped real useOne group of suppliers goes through the agent and is compared to the baseline.KPI: Minutes per invoice vs. baseline

Before you start

Process chosen and monthly invoice volume measured
Price and quantity tolerances written and approved
ERP user with limited permissions and no access to payments
100 real invoices with purchase order and goods receipt
Every action logged with user, date and version
Pilot owner and success metric agreed

Frequently asked questions

What's the difference between an agent and a chatbot?

A chatbot replies to messages. An agent executes work: queries systems, makes bounded decisions, completes multi-step processes and logs everything it does for audit.

Do we need perfect data to start?

No. Part of the diagnostic is evaluating your data sources and deciding which use case is viable today and which needs data engineering work first.

How do you keep the agent from making serious mistakes?

With three layers: bounded action scope, automated evaluations before each release and human oversight on sensitive decisions. Nothing critical is delegated without a safety net.

Do you work with companies outside Asunción?

Yes. We work remotely with companies across Paraguay and the region, meeting in local business hours.