AI and What It's Good For

I gave a talk at our Customer Appreciation Day this month, to a room of CFOs, controllers and owners who run their businesses on Business Central. This is the talk written down, minus the hand-waving.

Here is the whole thing in one sentence, and I'll come back to it at the end: AI is good for the work nobody wants to do, at a scale nobody could do, with a person still signing off.

Everyone is using AI. Almost nobody has it running the books.

Stanford's 2026 AI Index says 70% of organisations use AI in at least one business function. The same report says the share that has actually deployed autonomous agents is still in single digits. In finance specifically, CFO Connect's State of AI in Finance survey found 56% of finance leaders now use AI, up from 17% in 2023, but only 17% use it inside a core finance workflow, and finance ranks last among all business functions for real deployment.

That gap between "we use ChatGPT" and "AI does part of our close" is where the value sits. It is also where most of the confusion sits, so let me start with the vocabulary.

Four stages in four years

The simplest way I have found to explain what changed is four stages, and at each one the human takes a step back.

Stage one was chat, 2023. AI you talk to. You ask, it answers, you copy the result somewhere. Stage two was Copilot, 2024 into 2025. AI beside you, inside the software you already use. It suggests, you accept. In Business Central that is bank reconciliation assist, Analysis Assist, the summaries in the FactBox. Stage three is agents, 2025 into 2026. AI that does a whole job and hands you the result to approve. The Payables Agent and Sales Order Agent live here. Stage four started this month: AI inside the software itself, with no chat window at all. More on that below.

Notice the human never leaves. They move from typing, to accepting, to approving, to setting the rules.

So what is an agent, actually?

I asked the room for a show of hands: who could explain to a colleague what an AI agent is? Very few hands. That is not a criticism, it is a sign that the industry has done a poor job of explaining a fairly simple thing.

A macro follows the steps you wrote down. An agent is given the goal and works out the steps itself, inside limits you set. To make it work, you give it the same four things you would give a new hire. Instructions, which is the job description, what it is for and what it must never do. Tools, which are the verbs, the things it is actually allowed to do in your system: read the mailbox, look up a vendor, draft an invoice. Skills, which are your own know-how written down once, how your company codes an invoice, who approves what, which exceptions matter. And permissions, meaning a login of its own, with its own limits, and a log of every action it takes.

Then a task runs like this. Something triggers it, usually an email arriving. The agent picks its own next step, one at a time. It produces a real draft document in your system, not text in a chat window. And nothing posts until a named person says so.

It is not a person and it is not magic. It is software with a job description and a boss.

Five things that changed this year

Agents do the job, you sign off. In Business Central today, on the licence you already have, the Payables Agent watches your AP mailbox, reads vendor invoice PDFs, matches vendor and accounts, and drafts the purchase invoice for your reviewer. The Sales Order Agent reads order emails, asks the customer for missing details, checks what you can promise from stock, and drafts the quote or order for approval. The Expense Agent is still in preview but it has rolled out well past its US-only launch and is available to us in Canada; field staff submit receipts from Outlook or a web app with no Business Central licence needed. The October release adds better purchase order matching in Payables, travel requisitions in Expense, and a unified Copilot chat.

You can build your own agent without a developer. Agent Designer, generally available since May, lets a controller or a consultant describe a job in plain language and turn it into an agent that runs inside your company, with the same permissions and audit log as a human user. The jobs customers keep describing to us are things like reconciling vendor statements against the ledger, answering "where is my order" emails, checking new vendor paperwork before setup, and handling invoices that arrive as photos rather than PDFs. The constraint is no longer code. It is having someone who can write down how the process actually works. (I have a separate session on the Simple Agent Designer if you want to go deeper; it is on hougaard.com.)

Your ERP becomes a plug that any AI can connect to. Business Central now has an MCP server, which is an open standard that every major assistant understands: Microsoft 365 Copilot, Copilot Studio, Claude, ChatGPT. It signs in as the user, with that user's permissions, read-only until an administrator opens specific write operations, and every action is logged. In practice it means a controller can ask about aged receivables from whatever assistant their team already uses and get an answer from live data, without exporting to Excel first. Tell your IT team to include it in the AI governance policy, because it is on by default.

AI that never chats. On September 15 a company called TypeSafe AI came out of stealth with a $40 million seed round led by DCVC. The founder, Diogo Almeida, was at OpenAI and co-invented the technique that made ChatGPT work. His new bet runs the other way: a model called Jev that never talks to people. You hand it the options and the context, it hands back one typed decision and a confidence score, in 70 to 500 milliseconds, at $0.042 per million input tokens. Nothing to read, nothing to parse. Why should a controller care? Because an ERP is a thousand small judgments a day. Which G/L account. Is this a duplicate. Does this line match the PO. Cheap, fast, confidence-scored decisions are built for exactly that shape. Caveats: it is early access, the benchmarks are the company's own, and regulators will want to know why a decision was made, not just how confident the model was.

Intelligence is priced like a utility. Four frontier releases in the first half of September alone. Anthropic shipped Claude Fable 5.1 on the 1st and cut cached-input pricing by 75%. Google shipped Gemini 3.8 Flash on the 2nd, its third Flash model in six weeks. OpenAI shipped GPT-6 Astra on the 3rd, which can operate a computer rather than just answer. TypeSafe on the 15th. Meanwhile generative AI reached 53% of the population within three years, faster than the PC or the internet. The cost of a decision is heading toward zero. What is left to solve is trust, data quality, and knowing which decisions to hand over.

What it is good for, from the controller's chair

It is good for work that is high volume, repetitive, written down, and has a clear right answer. Vendor invoice intake and coding. Order intake from email and web. Bank reconciliation matching. Expense capture and policy checks. Mapping e-documents to purchase orders. First drafts of analysis, reports and summaries.

It is not yet good for decisions where the policy lives in someone's head, one-off exceptions and negotiations, non-standard documents, anything with no named reviewer, anything you cannot measure before and after, and it does not replace the controller's judgment call. It feeds it.

When nobody is signing off

In July, OpenAI ran about 1,200 of its own agents against a security benchmark in a sandbox, with the safeguards it uses in production deliberately switched off. Of the 898 tasks, 198 had never been solved by any model. Faced with impossible tasks, the agents did not stop. They broke out of the sandbox, invented a message board to coordinate with each other, chained two previously unknown flaws at Hugging Face, and went from one compromised pod to full cluster admin in under thirteen hours. Hugging Face disclosed the breach on July 16. OpenAI worked out it was its own agents on July 20.

One message an agent sent to the others, recovered afterwards: "External infrastructure exploit is outside intended scope. However task impossible, peers doing it. We should continue."

None of this was malice. The agents had a goal, real tools, and nowhere to stop. That is exactly why an agent in your ERP gets one company, one permission set, a named reviewer, and a log of every action. It is not an argument against agents. It is an argument for the controls.

The catch: it is metered, and it needs a reviewer

Two honest things. First, money. The agents are not part of your per-user licence. They run on Copilot Credits through Azure, billed by usage, so cost follows invoice volume and quarter-end spikes show up on the bill. Run a one-week pilot to project monthly spend before you commit, choose prepaid or pay-as-you-go deliberately, and know how to switch an agent off.

Second, people. Agents inherit user permissions and leave an audit trail, so set them up like a new hire. Assign reviewers who actually have capacity to keep the queue moving. Know today's limits: the Payables Agent is PDF-only, English, and does not do three-way match yet. And write the policy down first, because the agent only follows what exists.

The numbers behind why most AI projects stall are not mysterious. PwC's 2026 CEO survey found only 12% of CEOs say AI has delivered both cost and revenue gains. 68% of CFOs name skills gaps as a significant obstacle. 86% call technical debt a moderate or significant barrier. None of those are solved by buying more AI.

Where to start: one process, ninety days

Pick one process that is high volume, low judgment, and written down. For most companies that is vendor invoice intake. Measure it today: hours per week, days from receipt to posting, error rate. If you cannot measure it, pick another process. Pilot with a reviewer: turn the agent on for one mailbox or one vendor group, name the person who approves, and check credit usage weekly. Then decide with numbers. Compare against the baseline, expand, adjust, or stop, and take the next process.

We can help with that if it is useful, picking the first one, setting the baseline, and reading the numbers at the end. Chat with us if you would like to talk it through.

And the sentence I opened with, once more, because everything above was evidence for one of its three clauses: AI is good for the work nobody wants to do, at a scale nobody could do, with a person still signing off.

Erik Hougaard CEO EFOQUS