AI implementation
Let’s put AI on the work you still do by hand.
KZN has shipped software for nearly two decades and now runs itself on AI: the briefings, the invoicing, the inbox triage, the meeting record. We build the same agents for research centers, government agencies and companies, and train your team to run them after we leave.
Thirty minutes, no deck, and a straight answer either way. Two weeks and a fixed fee if you want the answer in writing.
A typical build
8 weeks
- Week 0–2
-
Scope
We map the workflow, the data and the sign-off. A written brief, a fixed price. We map the workflow, the data, and the review process that will decide whether this can go live. You get a written brief with the scope and the fixed price.
- Week 2–6
-
Build
Built on your real data, in your environment, with security and records people from week one. The system gets built against your real data, in your environment, with security and records people in the room from the first week rather than the last.
- Week 6–8
-
Handover
Your team is trained, the runbook is written, and it runs without us. No retainer. Your team is trained, the runbook is written, and the system runs without us. No retainer required to keep it alive.
In production, owned by you
01
We run on it
We rebuilt this firm on AI before we sold it.
There is no spreadsheet left in this company. One agent runs the operations on thirteen scheduled jobs, and each job loads its own playbook when it fires.
- A brief on what needs attention, and nothing else
- Weekdays, 9am and 1pm
- Inbox triage, where only the urgent gets through
- Every fifteen minutes
- Meeting recordings turned into commitments on the record
- Hourly
- Contracts, insurance and certificates checked for expiry
- Daily, 7am
- New federal awards read and qualified against our pipeline
- Daily, 8am
- Purchase orders and budgets checked against burn
- Daily, 9am
- Contractor invoices reconciled against logged time
- Weekly
- Utilization, time logging, and the financial summary
- Weekly
It did not start this way
The first version was thirteen separate agents, one per specialty, and every one of them decided on its own whether something was worth interrupting for. The result was noise. One agent decides now, and the other twelve became scheduled jobs. Finding that out cost us the time it took to build thirteen agents, and it is the kind of thing you are paying us to already know.
Your security people will ask about permissions
Ours did. Contractors here get their own agent, with their own credentials, scoped to exactly what they are allowed to see and blocked from writing anything operational. That question arrives on every engagement, and we answered it for ourselves first.
Public and inspectable. The toolkit behind our India tax agent is open source under MIT, with the full cost architecture written up: what the build cost, what the first returns cost, and the design decisions that got it there. Read the code and the decisions rather than take our word for it. The write-up is here.
02
What an agent is for
Work taken off people, not renamed.
The pattern we look for is expert people spending their week on data entry, chasing and reformatting. The agent takes that part of the job, shows its work, and asks before anything that matters. What is left is the judgment they were hired for.
- An advisor that knows the solicitation
- Proposal help grounded in the solicitation, the program guidance, and the office’s own funded proposals, so the advice is specific enough to act on at four in the afternoon. Common structured questions get an expert answer verbatim; the rest go to the model.
- University research office · In production
- A capability finder that cites every answer
- Ask which facility can characterize a given material and get an answer built from the program’s own pages, every claim linked to its source, ending at the named manager who owns that instrument. It declines anything outside the corpus.
- Federally funded multi-site research program · In production
- The scheduled jobs that run this firm
- The morning brief, the meeting record, the expiry checks, the invoice reconciliation: the jobs from the section above, running on our own operations. This is where we learned what breaks, before we sold any of it.
- This firm · Running since 2024
03
Where to start
Two ways most people start, both fixed fee.
Each is a fixed price agreed before it starts, two weeks on the clock, and a written document at the end. Neither obliges you to do anything afterwards, and neither fee is credited against a build. If neither is the right shape, start with a call and we will work out what is.
You have work being done by hand
AI Workflow Review
- You bring the workflows
- Two or three pieces of work you already suspect: the month-end close, the intake queue, the annual report, the thing your best person spends Thursdays on. We do not audit your organization. We look at the work you brought.
- What you get
- Each workflow costed: what it would take to build, what it would save, what would block it, and which one goes first. Plus the ones we tell you not to automate, and why. Whoever owns the budget can read it without a translator.
- Duration
- Two weeks
- Fee
- $12,000 fixed
- Covers
- Up to three workflows
- Your time
- About six hours
- Obligation
- None
You have a pilot that has not shipped
Production Readiness Review
- What we do
- We read the code, talk to the people who built it and the people who have to approve it, and put your pilot through the review it will eventually face.
- What you get
- A written verdict of ship, fix, or stop. The specific things that will fail review, named, with what each one takes to close and what finishing it would cost.
- The fee is not credited against a build
- A firm that credits the audit against the follow-on project has a reason to recommend the project. Ship, fix and stop are all real verdicts, and stop is the one that saves you the most.
- Duration
- Two weeks
- Fee
- $12,000 fixed
- Covers
- One pilot or system
- Your time
- About four hours
- Obligation
- None
95%
of enterprise AI pilots produce no measurable impact on the bottom line. The model is almost never the reason: the demo usually works. Review readiness, answerability and handover are what kill a pilot after the demo.
MIT, The GenAI Divide, 2025
We reply within two business days.
Bring us the workflow you already suspect.
Thirty minutes, no deck. You will leave knowing whether it is worth building, roughly what it would cost, and how long it would take.