060 279 5587 info@sitect.co.za 139 Davies Street, Doornfontein, Johannesburg, 2001 Gauteng, SA
Services / AI Consultation & Integration / Custom AI Agents & Copilots
Tool-using · Memory-keeping · Human-supervised

Not chatbots. AI teammates that actually get work done.

Custom AI agents — Atlas (sales), Drafter (legal), Ledger (finance) — that work alongside your team, take real actions in your tools, and check in with humans when it matters. The flagship of the Sitect AI stack, built for businesses ready to put AI to work.

Live in 6–10 weeks Multi-step reasoning Approval gates + audit From R24,000

Chatbot, copilot, agent — three different things.

Most teams confuse them. The difference matters because they cost different amounts, take different times to build, and solve different problems. Here's how we draw the line.

Level 1

Chatbot

Reactive · answers questions

Waits for a question. Answers it from a knowledge base. Sometimes books a meeting if you ask nicely. Doesn't initiate anything.

Initiates workNo
Multi-step reasoningRarely
Uses your toolsSome
Build costR 10k+
Build time4 weeks
Level 2

Copilot

Assistive · drafts for a human

Sits next to a person doing a job. Drafts the email, summarises the meeting, suggests the next action — but a human reviews and decides.

Initiates workNo
Multi-step reasoningYes
Uses your toolsYes (read)
Build costR 24k+
Build time6–8 weeks
Level 3 · flagship

Agent

Autonomous · gets work done

Given a goal, breaks it into steps, picks tools, executes, checks the result, retries, and asks for human help at high-stakes decisions. Works while you sleep.

Initiates workYes
Multi-step reasoningAlways
Uses your toolsRead + write
Build costR 32k+
Build time6–10 weeks

The agents that run Sitect itself.

We dog-food every agent before we sell it. The eight below run our own sales, legal, finance and ops. Each can be re-themed and re-trained for your business in 6–10 weeks.

A
Atlas
Sales SDR

Qualifies inbound leads, enriches profiles, scores fit, drafts outreach. Books discovery calls into a real human's calendar.

clearbit crm.write cal.book
D
Drafter
Legal · proposals

Drafts proposals, MSAs, NDAs and contracts. SA-legal-aware. Always human-approved before send. Tracks redlines.

docs.draft redline e-sign.send
L
Ledger
Finance · AP/AR

Reads invoices, matches POs, drafts dunning emails, reconciles Yoco/EFT. Stops at R10k+ for human approval.

xero match.po dunning.send
R
Aria
Support · after-hours

Handles WhatsApp + IG support 24/7. Escalates anything spicy to humans within 4 hours with a context summary.

whatsapp kb.search escalate
K
Archive
Internal Q&A

"What did we agree with this client in March?" Answers internal staff questions from your knowledge base, with citations.

rag.search cite.source slack.reply
S
Scout
Research · prospect

Researches accounts before sales calls. News, hires, funding, competitors. Produces a 1-pager briefing 30 min before each meeting.

web.search news.scrape linkedin
O
Onboarder
Ops · client setup

New client signed? Onboarder provisions accounts, sends welcome pack, schedules kick-off, sets up Slack channel, files contract.

provision slack.create cal.invite
V
Verifier
Compliance · FICA

Runs FICA checks on new clients. Verifies SA ID, proof of residence, CIPC. Flags anything weird for human eyes.

id.verify cipc.lookup flag.review

Six capabilities. All baked in.

Every Sitect agent ships with these six superpowers. They're the difference between "an LLM call" and "an autonomous worker you can trust with real work."

Multi-step reasoning

Given a goal, the agent plans a sequence of steps, executes them, and adjusts the plan as new information comes in. Not a single LLM call — a thought process.

Tool use

Reads from and writes to your real tools — CRM, accounting, calendar, payment, email, Slack, your own APIs. Every tool call is logged and replayable.

Memory

Short-term working memory across the task. Long-term episodic memory across tasks — the agent remembers what worked, what failed, what each client prefers.

Self-correction

When a tool fails or a step doesn't work, the agent notices, re-plans, and tries a different approach. Up to 3 retry attempts before escalating to a human.

Approval gates

Configurable per-action: agents act autonomously up to a threshold (e.g. send R<10k invoices), then require a Slack thumbs-up for anything bigger.

Sacred audit log

Every thought, tool call, decision and action stored forever, replayable. POPIA-compliant retention. When something goes wrong, you can replay the exact run.

Under the hood

The agent loop, visualised.

Every Sitect agent runs the same five-stage loop. Plan → choose tool → call → reflect → repeat or finish. The loop is what separates an agent from a single LLM prompt.

01 · GOAL

Receive goal

From human, schedule or trigger.
"Qualify L-3417"
02 · PLAN

Plan steps

Decompose into actions.
[enrich, score, decide]
03 · ACT

Call tool

Real API, real action.
clearbit.enrich(...)
04 · REFLECT

Reflect

Did it work? Next step?
score = 87, proceed
05 · DECIDE

Finish or escalate

Done · retry · human gate.
done(opp_created)
The loop iterates until done or escalated · capped at 15 steps · runaway-protection built in. Every iteration logged · failures retried with backoff · human-in-the-loop on demand.

One trigger. An overnight outcome.

This is a real run of Onboarder, the agent that handles new-client setup at Sitect. Triggered when a contract is signed in DocuSeal. Total elapsed: 47 seconds. Zero human involvement until the welcome call.

  • Triggered automatically by a webhook from DocuSeal when contract signed.
  • Multi-tool execution — Slack, Calendar, CRM, S3, GitHub. Five tools touched.
  • Self-correcting — Slack channel creation failed (name collision), agent picked an alternative.
  • Audited — every action stored in the action log, replayable from any timestamp.
  • Hand-off — finished with a Slack message to the project manager, ready for kick-off.

Six artefacts. A teammate that ships with you.

Every Sitect agent comes with the same scaffolding. You own all of it — code, configs, prompts, audit trail. Run it on your infra forever.

01 · Agent persona

Personality + system prompt

Name, voice, role, constraints, escalation rules. Written in plain markdown so your team can refine it without a developer.

Markdown · Git-tracked
02 · Tool registry

Typed catalogue of capabilities

JSON-Schema for every tool the agent can call. Auth tokens stored in your vault. Each tool tagged with cost + criticality.

tools.json + secrets
03 · Memory store

Episodic + semantic memory

Working memory in Redis · long-term in your Postgres/vector DB. Cleanable per-user for POPIA right-to-be-forgotten.

Redis + pgvector / Qdrant
04 · Approval inbox

Slack-native human gate

Pending approvals land in a dedicated Slack channel — thumbs-up to approve, thumbs-down to reject, reply with text to send back to the agent.

Slack · email · web dashboard
05 · Action log

Sacred audit trail

Every thought, tool call, result, decision. Searchable by run-id, agent, user, time. Replay any past run end-to-end. POPIA-compliant retention.

JSON log · replay UI
06 · Runbook + training

Your team owns it from week 11

Markdown runbook for the top 12 incident types. 3-hour live training session. 90 days of Sitect on-call hand-off, then your team takes over.

Notion + PagerDuty

6–10 weeks. 5 phases. One agent you trust.

Most of the work is in the front (defining the job clearly) and the back (training your team to supervise). The middle is the actual build, which is the smallest piece.

01
Week 1–2 · 6 hrs

Job description

What's the agent's role? Boundaries? Tools needed? KPIs? You write its "job description" with us. Signed off before any code.

02
Week 2–4 · 3 hrs

Tools + memory

Wire up the tools the agent will use. Provision memory store. Define POPIA boundaries. Auth tokens in your vault.

03
Week 4–6 · 2 hrs

Train + approve gates

System prompt drafted, approval rules configured, eval suite built. Agent runs in shadow mode against real data.

04
Week 7–8 · 3 hrs

Pilot

Agent goes live to a closed pilot. Heavy human supervision. We tune from real failures daily.

05
Week 9–10 · 2 hrs

Go live + train

Full traffic. 90 days of Sitect on-call hand-off. Your team trained. Approval gates progressively relaxed.

The numbers behind a working agent.

Aggregated across our deployed agents. Your numbers will be yours — these are the order-of-magnitude bracket to plan against.

11.4×
Task throughput
Tasks completed per week vs the manual baseline.
96%
Auto-completion rate
Tasks finished without human intervention after 90 days of tuning.
R 84k
Monthly savings
Median FTE-equivalent displaced per month.
0
Unauthorised actions
Approval gates have a perfect record across our agents to date.

Pick the scope. We deliver in 6–10 weeks.

Build fee is fixed in writing. Variable cost = LLM tokens + cloud (typically R 1.5k–R 8k/month) billed direct. No per-agent platform fee. You own everything.

Copilot

Single-agent build

R24,000
/project · ex 15% VAT
  • 1 copilot (human-supervised)
  • Up to 4 tool integrations
  • Working + episodic memory
  • Slack approval inbox
  • Action log + audit replay
  • 60 days post-launch tuning
Choose Copilot →
Enterprise

Agent fleet

R100,000+
scoped per engagement · ex VAT
  • 3+ coordinated agents (a fleet)
  • Multi-agent orchestration
  • Shared memory + tool catalogue
  • SSO · RBAC · audit · DSAR
  • Custom evals + ops dashboard
  • SLA + dedicated PM
  • 90 days hyper-care + QBR
Talk to us →

Common questions from teams hiring an AI.

If yours isn't here, send us the job description. We'll come back with a build plan in 48h. No sales pitch.

How is this different from a chatbot or a Zapier workflow?
A chatbot answers when spoken to. A Zapier workflow runs a fixed sequence (if X then Y). An agent figures out the sequence on the fly — it plans, picks tools, retries when things break, asks for help when stuck. The difference matters when the work isn't predictable: lead qualification, content review, exception handling, research. Predictable work — stick with Zapier (we'll tell you so).
What if the agent does something we don't want?
Three layers: (1) System-prompt boundaries — explicit rules about what the agent must and must not do. (2) Approval gates — configurable per action; anything above a threshold (e.g. spend > R10k, or "send email to anyone outside the company") requires a Slack thumbs-up. (3) Audit log — every action stored forever, replayable. Plus a kill-switch in the dashboard pauses all agent runs in one click.
Will the agent replace someone on our team?
Honest answer: an agent typically does the repeatable portion of someone's job — maybe 30–60% of it. The remaining work (judgement calls, relationship building, exception handling) becomes the human's focus. Most teams don't reduce headcount; they reallocate. The teams that do reduce typically use it to avoid a hire they were about to make.
POPIA — where does customer data go?
Customer data stays in your tenant. The agent's memory store runs on your infrastructure (Postgres + Redis on AWS / Azure / on-prem). LLM provider calls strip PII before the prompt leaves your network. We sign a POPIA-compliant DPA before kick-off, and your Information Officer reviews the agent's "boundary doc" before go-live.
What model do you use under the hood?
Provider-agnostic. We default to claude-3.5-sonnet for reasoning and gpt-4o-mini for cheap routing calls. Can swap to Gemini, Llama 3 on-prem, or any OpenAI-compatible endpoint with a config change. For sensitive industries (finance, legal, health) we deploy on-prem options.
How much does it cost to run per month?
Depends entirely on volume. A copilot handling 50 tasks/day costs typically R 1.5k–R 3k/month in tokens + compute. A high-volume sales agent (1,000+ tasks/day) might land at R 6k–R 8k/month. We give you a usage dashboard and a monthly cap — never get a budget surprise.
What happens if the LLM is down?
Multi-provider fallback by default — if OpenAI is down, the agent falls back to Anthropic, then Google. If all three are down (rare), the agent pauses with a Slack alert. In-flight runs are checkpointed every step, so they resume from where they left off when service returns.
Can the agent learn over time?
Yes, in two ways: (1) Episodic memory — past successful runs become reference examples for similar future tasks. (2) Eval suite — every prompt change runs against a golden set of past tasks; we catch regressions before they ship. We don't fine-tune the underlying LLM (cost / privacy / drift risk) — improvements come from better prompts, better tools, and better memory.

Describe the job.
We'll design the agent.

Write down a job your team does repeatedly — qualifying leads, drafting proposals, handling claims, onboarding clients. Send it to us. Within 5 business days we'll come back with an agent design document: tools needed, build cost, indicative timeline, and three reasons we might be wrong about it being worth automating.

You'll get back in 5 days

  • Agent design document
  • Recommended tools + memory
  • Indicative price + timeline
  • Honest "should we even build this" verdict

More AI Consultation services

AI Strategy & Roadmap

Where does AI fit in your business?

AI Chatbot Integration

Smart chatbots that capture, qualify and convert.

AI Voice & Phone Agents

24/7 AI agents that handle calls.