Blog · 12 November 2026
AI pays off in a small business when it is pointed at one repetitive job that lives in documents and tools: answering the same questions, reading invoices and forms, drafting replies, keeping records straight. It wastes money when a company buys a chatbot instead of choosing a job, or automates a decision that needs judgement before the data is clean.
Start with the job, not the technology
Every failed AI project we have seen started with a tool: a chat assistant, a copilot licence, a platform someone saw at a conference. Every one that worked started with a sentence a person in the business could say out loud. Answer these refund questions. Read these invoices. Chase these overdue payments. Fill this form from that email.
The job has to be frequent, rule shaped and annoying. If it happens twice a month, no one cares how well it is automated. If it happens forty times a week and eats a person's morning, everyone cares.
The three jobs AI is genuinely good at now
- Finding and explaining. Answering a question from a set of documents, with the source shown, is a solved problem when the documents are decent and the answer is in them.
- Reading and filing. Extracting fields from invoices, receipts, forms, delivery notes and emails, then putting them where they belong, is reliable enough to save real hours.
- Drafting and routing. Writing the first reply, the summary, the chase email or the internal ticket, and sending it to the right queue, is where most teams get their first real win.
The three jobs it is bad at, for now
- Decisions with money or reputation at stake, without a person approving. The failure is rare, expensive and memorable, and one of them can end the project politically.
- Arithmetic over numbers that change. If the answer depends on today's stock or today's balance, the agent must query the database, not read a document about it.
- Anything with no data behind it. If the knowledge lives in one person's head, no model will find it, and the first step is writing it down.
Why companies start in the wrong place
The usual sequence is: buy a licence, run a pilot on clean sample data, impress a meeting, then discover that the real data sits in four systems, the process has exceptions nobody documented, and no one owns the result after the demo. The pilot dies quietly and the budget goes to something else.
The fix is boring. Pick one job. Use the real data, including the mess. Decide who approves what. Measure one number before and after. Do it in days, not quarters, so the cost of being wrong stays small.
What a working deployment needs
Ten things, and none of them is the model: a clear job, access to real data, answers grounded in sources, scoped tool access, written guardrails, human approval for risky actions, tests on real examples, logging of every run, cost tracking and a named owner.
The AI agent production checklist. The ten items, and where teams usually cut corners.
How Infrenic does it
You send us one problem and a sample of the data behind it. Within 48 hours you get a working version on that data, free, so you are judging results instead of slides. If it is worth building, we take it to production in your own cloud in two to six weeks, then run and improve it every month.
All of it runs on Keel, the platform we built after seeing the same plumbing rebuilt on every project. Keel already holds the parts that kill pilots:
- Agents that do a job end to end, with tools and memory.
- Knowledge: answers from your documents, with the source every time.
- Connectors to email, help desk, CRM, drives, databases, your own APIs and MCP servers.
- Approvals and identity: single sign-on for people, scoped access for agents, a person approves what matters.
- Hardware: cameras, sensors, scanners and machines through an edge gateway in your network.
- Monitoring: every run logged, costed and tested before each change.
The point of the platform is not the feature list. It is that the second app takes days instead of months, because it reuses the plumbing of the first.
Send us a problem. A 30 minute call, then a working version on your data in 48 hours.
A realistic first year
- Months one to three: one job automated properly, in production, with one number moving. The team learns what good data looks like.
- Months three to six: two more jobs on the same platform, usually in a different department, reusing the connectors and approvals.
- Months six to twelve: the boring part that decides whether it lasts. An owner reviews failures weekly, the test set grows, and the assistant gets more trusted, not less.
Companies that skip the last step end up with three impressive demos and no system anyone relies on.
FAQ
Is AI worth it for a company of 50 people?
Yes, if the first job is frequent and document shaped. A single support or finance workflow that saves ten hours a week pays for a small build within months. It is not worth it as a licence bought without a job attached.
Do we need a data team first?
No. You need one person who owns the problem and can give access to the documents and systems involved. The data work is part of the build, not a prerequisite.
What if our data is messy?
It usually is, and finding out is part of the value. A 48 hour proof of concept on real data tells you what is usable, what needs cleaning and what should stay out of scope, before you commit a budget.
What does it cost?
The 48 hour proof of concept is free. Production work is a fixed quote after that, sized to the one job. Model and cloud costs are billed to your own accounts, so nothing is hidden in a subscription.
Infrenic builds AI agents on your data and tools, ships them to production and runs them, all on Keel. Send us a problem and see it working in 48 hours.
The Infrenic AI Dispatch
Practical architecture notes on AI agents, RAG, and production systems. No buzzwords, no spam. Sent when we have something real to share.
Read earlier editions at infrenic.beehiiv.com