Activación·Web// Tech Agency · est. 2015Agents that do the work, not demos that impress.
If it does not save hours, cut errors or open new revenue, it is not worth building. We pick the right model for the case rather than whichever one is trending, and we measure the result against a baseline you set beforehand.
Where AI actually earns its keep
Most requests fall into one of these three. Which one fits depends on where your data can live and how much judgement the task requires.
Not the generic bot that confuses people. An assistant that reads your documentation, products, policies and processes to answer precisely and cite its source. WhatsApp, web, app or voice.
- Retrieval over your own documents, not a generic model guessing
- Cites the source so an answer can be checked
- Human handoff the moment confidence drops
- Deployed on WhatsApp Business, web, Slack or voice
The work that eats hours is rarely one big task — it is twenty small handoffs between systems that do not talk. We wire them together and put AI only at the steps that need judgement.
- Document pipelines: invoices, contracts, IDs, receipts
- Automatic classification of tickets, emails and calls
- Orchestration with n8n, Make and your existing APIs
- Human validation at the steps where a mistake is expensive
Open-weight models running on your own servers or in a private cloud region. Slower to set up and more expensive to run than an API, and the right answer when confidentiality or regulation demands it.
- Open models self-hosted on your infrastructure
- No data leaves the perimeter you control
- Fine-tuning on your own corpus where it earns its cost
- Full audit trail of every inference
Twelve shapes this takes in practice
Most real projects combine two or three of these rather than being a pure example of one.
Conversational agents
WhatsApp, web, voice, Slack, Telegram
Document processing
Invoices, contracts, IDs, deeds, receipts
Voice agents
Automated inbound and outbound calls
Predictive analytics
Anticipate churn, demand, fraud, conversion
Content generation
Text, images, video and voice at scale
Internal knowledge assistant
Your team asks, the AI answers from your docs
Computer-use agents
AI that operates a browser or your systems
Personalisation engines
Recommendations, content and dynamic pricing
Workflow orchestration
AI plus n8n plus Make plus your systems
Transcription and translation
Calls, meetings and audio, searchable
Computer vision
Visual inspection, counting, classification
Multi-agent systems
AI that decomposes tasks and coordinates itself
The eight numbers we hold ourselves to
Ask any AI vendor for these before you sign. If they cannot name their hallucination rate or their cost per response, they are not measuring the system — they are hoping.
Deflection rate
Share of enquiries the AI resolves without escalating to a human. The most direct measure of whether it is working.
Hallucination rate
Share of factually wrong answers, measured by human evaluation over a sample. Above 2% in production, we stop and fix before scaling.
Latency
Response time. Under 3 seconds at the 95th percentile for chat, under 800ms for autocomplete. Latency kills adoption faster than a wrong answer.
Cost per response
What each answer costs in tokens. Caching, correct RAG and model routing typically cut this by an order of magnitude.
Precision and recall
For classification and extraction. Precision: of the ones it flagged, how many were right. Recall: of the ones that mattered, how many it caught.
Hours freed per month
The business metric that decides everything. Measured against a pre-AI baseline, not estimated afterwards.
Data drift
In classical predictive models: how far the incoming data has moved from the training set. Silent drift is how a good model quietly becomes wrong.
User satisfaction
A short survey after the conversation: did this resolve your problem? The best available proxy for perceived quality.
Three ways to start
Almost everyone should start with the first one.
We pick the single highest-value process, build it end to end and measure it against the baseline. You leave with evidence, not a slide deck — including the case where the answer is that AI is not worth it here.
We build the solution and run it: monitoring, evaluations, prompt and model updates as the providers move. Monthly, with an SLA.
Open models on your infrastructure, fine-tuned on your corpus. You own the weights and the pipeline. Higher upfront cost, no per-token bill and no data leaving your perimeter.
What a technical buyer asks first
Including the one about when not to use AI at all.
Whichever fits the case: Claude from Anthropic, GPT from OpenAI, Gemini, and open-weight models such as Llama, Mistral, DeepSeek, Kimi or Qwen when the deployment must stay private. We route per task rather than betting the system on one provider.
Retrieval over your own documents so the model answers from source rather than memory, citations on every answer, confidence thresholds that trigger human handoff, and a measured hallucination rate we hold under 2% before scaling.
We tell you. Some processes are better fixed with a rule, a form or a database index than with a model. It is a shorter invoice for us and a working process for you.
You decide before we start. With API providers, data is not used for training under their business terms. If that is not enough for your case, we deploy open models on your own infrastructure and nothing leaves it.
Name the process that eats the most hours
That is usually where the first proof of concept belongs — and thirty minutes is enough to tell whether it is a good candidate.