iqsoft

AI agents & automation

AI that does real work, safely

We build AI into the systems you already run, to take routine work off your team. Your data stays in your own cloud account and is never used to train AI models.AI built into your systems. Your data stays in your own cloud account.

Start a projectWhat we build
How an AI feature runsFig. 1
  1. Your datastays here
  2. Finds the facts
  3. AI modelnever trained on it
  4. Checkstested for accuracy
  5. A person signs off
  6. The work gets done
Your datastays here
Finds the facts
AI modelnever trained on it
Checkstested for accuracy
A person signs off
The work gets done
Inside the system

How an AI system we build actually works

One request, start to finish.

System mapIllustration of a typical agentic system
YOUR CLOUD ACCOUNT · your data never trains the modelRequestReconcile Marchinvoices for40 clinicsOrchestratorLLMRetrieverAnalystWriterCheckerDocumentsDatabaseERP and CRMyour APIsMemoryshort-termKnowledgelong-termHuman approvala person signs offEvaluationsand feedbackretryapprovals, correctionslearning loop · improves prompts, retrieval, checksACCURACY · LAST 12 RUNS93.3%from feedback, not retraining
Agent tracesample run
  1. request: "Reconcile March invoices for 40 clinics"
  2. plan: 4 sub-tasks
  3. retriever: vector search · 12 matches
  4. analyst: sql · 40 clinics, 318 invoices
  5. analyst: 3 mismatches found
  6. writer: drafted 40 entries · 3 flagged
  7. checker: confidence 0.82 < 0.9 · retry
  8. writer: revised
  9. checker: confidence 0.94 · pass
  10. gate: waiting for approval
  11. human: approved
  12. erp: 40 records updated
  13. eval: +1 labelled example
  14. learn: prompts, retrieval, checks improved

A sample run

The full trace of one run, from request to learning. The retry is shown too.

Hover, focus or tap any part to see what it does.

A working example

Seven self-learning agents that trade gold, inside limits people set

A desk we built and run ourselves.

Read the case study
How a decision runs · illustrationNot investment advice
  1. Zeusindicators
  2. Veratrend
  3. Marcusrisk
  4. Irissetups
  5. Danedecision
  6. Noracritic
  7. Theotrades
  1. Herculesorchestrator decides
  2. Risk controlsthe final say
  3. Orderwith a stop-loss

Zeus reads the indicators: votes sell.

What we build

Five kinds of AI, built into your systems

  • Assistants

    questions → sourced answers

  • Document processing

    paperwork → data

  • Routing and triage

    inbox → right desk

  • Drafting

    routine → reviewed

  • Agents with limits

    task → approved steps

Where AI helps in your sector
How we keep it safe

Useful AI is careful AI

  • Your data stays yours

  • A person signs off

  • Tested on your work

  • You can see why

See it work

One email, start to finish

  1. 01Email arrives
  2. 02AI reads it
  3. 03AI drafts a reply
  4. 04A person approves
  5. 05Filed
Support inboxIllustration

01 / Email arrives

From Operations, Riverside Clinic

Invoice for March is wrong

The March invoice lists 14 licences, but only 10 of us use it.

02 / AI reads it

  • CategoryBilling
  • UrgencyHigh
  • AccountMatched

03 / AI drafts a reply

We checked it against your 10 users. A corrected copy follows today.

Source: Billing policy, section 4

04 / A person approves

ApprovedNothing is sent until a person says so.

05 / Filed

Ticket #4821 updated · reply sent

Scripted example, no live AI
When we say no

Where AI is not worth it

We will tell you if it is one of these.

  • A simple rule or form already does the job
  • The decision needs a person’s judgement from start to finish
  • The data the AI would need does not exist yet
How we build

AI drafts. Engineers decide.

Our own engineeringFig. 2
AI draftsroutine code, tests
Engineer reviewsevery change
Shipsdemo every 2 weeks
Ways to start

Start small, with a clear end

Each has a clear scope and a clear result.

Ask which fits you

  • The fast way in

    100 hoursof engineering

    100-hour pilot

    A first working version.

    Start a 100-hour pilotHow it works

  • Not sure where to start

    2 weeks

    AI opportunity review

    A ranked list and one pilot plan.

    Start with a review

  • Then

    4–6 weeks

    AI pilot

    Live with real users, and measured.

    Start a pilot
FAQ

Questions about AI

Leading models from OpenAI, Anthropic and Google, chosen per project.

A person approves anything risky, and answers show their source.

Usually, yes, through its API.

Tell us the routine work that takes the most time.

Often paired

Goes well with

Which routine work takes your team the longest?

Tell us. We will say plainly whether AI is worth it, and what a first version would look like.

Start a project