Loading...
Loading...
Services / AI Automations
Two layers of automation in one engagement. Workflows for the deterministic plumbing — lead capture, document intake, cross-system reporting. Agents for the long tail where the right action depends on what the input actually says. POPIA-reviewed, monitored, and handed over with the code in your repository.
Book a CallWhat it actually does
Most SA SMBs have SaaS tools that should talk to each other and don’t. The website form drops a lead, but someone still copies it into the CRM. Invoices arrive by email but someone still types them into the accounting system. Reports get assembled manually from three systems on the first Monday of every month. The first layer of an AAA Automations build connects those tools properly — pragmatic, reliable, and matched to your scale.
The second layer handles the work the workflow layer can’t. When an email arrives with an attachment, what kind of document is it? Which client does it belong to? Should it go to the bookkeeper or get flagged for a partner? Those are judgment calls, not data-shape checks. AAA builds the judgment layer with production-grade agent frameworks paired with the leading reasoning model for your use case — small, focused, with explicit escalation triggers and a monitoring layer that catches drift before users do.
At the end you own the code. Every build ships to a private GitHub repository under your organisation. Every credential lives in a secrets vault you control. There is no AAA platform, no licence we control, no “you need to stay with us to keep it running” clause. The 30-day support tail covers bugs; after that you either self-manage (the runbook is designed for it) or move to the Monthly Advisory Retainer.
When to use this
Not a fit if
What’s included
How it’s delivered
Each stage is fixed-price and self-contained. The Map stage is free; you commit to the Pilot before any work begins, and to Ship before production hardening starts. The Pilot exists to prove the approach before the bigger spend — if it doesn’t earn its place, we don’t move on.
A structured discovery session — on-site for Cape Town and Stellenbosch, remote for everyone else. We walk the current manual process with your team and identify where automation actually pays off. You leave with a one-page scope note covering proposed approach, integration points, human-in-the-loop triggers, and the fixed Pilot price. No work begins without your sign-off.
A working prototype running against your real data. You see the system handle actual inputs from your business and decide whether to commit to a full production build. Pilot deliverable: a demo session and a Production scope recommendation. If the Pilot doesn’t earn its place, we don’t move on.
Production-grade build. Monitoring, alerting, security hardening, prompt-injection guards, and a POPIA compliance review with documented data flows. Final acceptance against your operations team using real production inputs before go-live. The handover includes a training session and a 30-day support tail.
After the 30-day support tail you either self-manage (full documentation provided) or roll onto the Monthly Advisory Retainer. The Run option covers ongoing monitoring, bug fixes, prompt tuning as input patterns evolve, model migrations when providers update, and a monthly cost report.
Why this approach
Three findings shape the AAA Automations methodology — the production-gap problem, the citation-density lever for any agent that generates content, and the AI-search transformation that is now load-bearing for South African businesses.
of generative-AI projects forecast to be abandoned after proof-of-concept — the failure mode AAA’s Pilot gate is designed to catch early. Gartner, July 2024.
visibility lift for content with cited statistics. Named quotations add +31%. Agents that produce client-facing outputs benefit from the same rules that govern human writers. Aggarwal et al., Princeton, KDD 2024.
lift for lower-ranked pages that add outbound citations — the single largest effect-size in the Princeton GEO study. Same paper, different rank cohort.
The most expensive AI project I’ve seen was a custom agent built for a process that could have been handled by a Zapier workflow with a GPT step.
Matt Owen, in From ChatGPT to Custom AI Agents: What’s Actually Worth Building.
The cheapest tool that solves the problem usually wins. AAA’s Discovery session sorts workflows from agents from I-could-write-this-down before any code is written. The bias is against over-engineering — the “standalone-extractable” principle from Aleyda Solis’s AI Search Optimization Checklist applies as much to the system architecture as it does to the content that comes out of it.
Common questions
A workflow runs a fixed sequence of steps when a trigger fires — if X happens, do Y, then Z. Workflows are appropriate when every input fits a predictable structure and the right action is decidable from data shape alone. An agent reads the content of the input, decides what to do based on that content, and adjusts its path. Agents are appropriate when inputs are unstructured, when the right action depends on what the message says not just what fields it has, or when different inputs should produce meaningfully different outputs. Most AAA Automations builds blend both — workflows for the deterministic plumbing, agents at the specific points where judgment is required.
Every AAA Automations build defaults to human-in-the-loop for any outbound communication or write-back to a system of record until the agent has demonstrated sustained accuracy on your real data. Lower-confidence outputs route to a review queue. Higher-confidence routine outputs auto-process. The “never send without review” rule is made explicit in the SOW on the channels where the cost of a wrong action is high.
Every Production engagement includes a POPIA compliance review and data-flow documentation handed to your information officer. We use enterprise-tier accounts with the LLM providers, which publish their Data Processing Agreements explicitly for business customers. Where feasible we strip or pseudonymise personal information before it leaves SA jurisdiction. Hosting defaults to SA or EU regions — both acceptable for POPIA Section 72 cross-border transfer.
You can. The handover transfers the GitHub repository to your organisation, gives you owner access to the secrets vault, and removes AAA access at the end of the 30-day support tail. The runbook is detailed enough that any developer can pick up the system. We earn the Monthly Run engagement on quality, not lock-in.
Tell me what your team is doing manually that feels like it should be automatic. I’ll tell you whether it’s a workflow problem, an agent problem, or something simpler.
Book a Call