AI software a company actually needs

A long AI tool list feels thorough. In practice it creates work. Every extra tool needs an owner, a login policy, a data review, and a renewal date.

The teams that get value pick a small number of tools and use them properly. They ask a simple question for each one: which job does this do, and who is accountable for the output?

That question kills most of the list. A summariser that nobody owns produces nothing. A tool that sits inside work people already do produces a lot.

It also makes the buying decision honest. If two tools do the same job, one of them goes. If a job has no tool, that is the gap worth spending on.

Engineering is the easiest place to start. Cursor sits in the editor. The model suggests, the engineer accepts or rejects, and the pull request still runs through review and tests. Nothing changes about who is accountable.

Knowledge work is next. Claude handles the long document, the messy research question, and the first draft that nobody wants to write. The value is time saved before a human edit, not a finished artefact.

The office suite is where most companies land by default, because the licences are already there. Microsoft 365 Copilot drafts an email, summarises a meeting, and pulls a table together. It is broad and shallow by design.

Finance is the odd one out. The other three tools help a person do a task faster. Finance work is not one task. It is a chain: an invoice arrives, it gets read, coded, approved, paid, matched to a bank line, and then closed into a period.

A copilot bolted onto a batch ledger can help with a step in that chain. It drafts a bill or answers a question about a report. Someone still posts, and someone still chases the missing receipt.

Light works on the chain rather than the step. AI reads the bill or receipt, including line items, tax, and terms. Approvals go to Slack or Teams. Agents handle AP, AR, reconciliation, and the close on one live ledger, and finance reviews the result.

Because the ledger is live and multi-entity, the group view is not a monthly rebuild. Intercompany, multi-book, and revenue recognition are natively included rather than assembled later.

That is why finance is a separate pick on this page. It is a different shape of problem, so it needs a different shape of tool.

Cursor will not decide what to build. It speeds up the writing, not the product judgement.

Claude will not act on your systems. It has no ledger, no approval chain, and no audit trail of your transactions.

Microsoft 365 Copilot will not run a finance process. It drafts and summarises inside documents and mail.

Light will not run inventory or a shop floor, and it is not a full operational ERP. If the business needs manufacturing, warehousing, or one suite for operations, look at Oracle NetSuite or Microsoft Dynamics 365 Business Central. Light is cloud-only, hosted in AWS EU, pricing starts from $35,000 per year, and the usual fit is 30 to 5,000 employees with two or more legal entities. Named customers include Tillo, KeyShot, and Alva Labs.

Pick the two jobs with the most repeated manual work. For most companies that is engineering and finance.

Run each tool against real work, not a demo dataset. For finance, that means a real month of invoices, real approvers, and a real bank feed.

Write down what a person still has to do afterwards. That number, not the demo, is the value.

Then decide the office layer last. It is the cheapest to add and the easiest to change.